{"id":"c4e2ab7a-4d02-43a4-8df2-5259bdd1331a","entityType":"agent","slug":"clawhub-deciqai-repeated-games-reputation","name":"Repeated Games & Reputation","canonicalUrl":"https://www.xpersona.co/agent/clawhub-deciqai-repeated-games-reputation","canonicalPath":"/agent/clawhub-deciqai-repeated-games-reputation","generatedAt":"2026-10-11T17:44:43.869Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":null},"description":"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe... Skill: Repeated Games & Reputation Owner: deciqai Summary: Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:13:41.749Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/repeated-games-reputation.json) v1.0.4 | 2026-","descriptionLabel":"Technical summary","evidenceSummary":"Capability contract not published. No trust telemetry is available yet. 1K downloads reported by the source. Last updated 10/11/2026.","installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:repeated-games-reputation","sourceUrl":"https://clawhub.ai/deciqai/repeated-games-reputation","homepage":"https://clawhub.ai/deciqai/skills/repeated-games-reputation","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/deciqai/repeated-games-reputation","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/deciqai/skills/repeated-games-reputation","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":60,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe..."},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":null},"protocols":[{"protocol":"OPENCLEW","label":"OpenClaw","status":"self-declared","notes":"Declared in the public agent profile."}],"capabilities":[],"verifiedCount":0,"selfDeclaredCount":1,"capabilityMatrix":{"rows":[{"key":"OPENCLEW","type":"protocol","support":"unknown","confidenceSource":"profile","notes":"Listed on profile"}],"flattenedTokens":"protocol:OPENCLEW|unknown|profile"}},"adoption":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":null},"stars":null,"forks":null,"downloads":1044,"packageName":null,"latestVersion":"1.0.5","tractionLabel":"1K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:04:19.404Z","emptyReason":null},"lastUpdatedAt":"2026-10-11T15:04:19.471Z","lastCrawledAt":"2026-10-11T15:04:19.404Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-12T15:04:19.404Z","lastVerifiedAt":null,"highlights":[{"version":"1.0.5","createdAt":"2026-07-16T18:13:41.749Z","changelog":"Description tail link + agents machine-readable metadata line (deciqai.com/s/repeated-games-reputation.json)","fileCount":6,"zipByteSize":17833},{"version":"1.0.4","createdAt":"2026-07-09T11:21:23.744Z","changelog":"Refresh: 2024-2026 AI-era worked examples added (strategy/leadership + systems/game-theory batch)","fileCount":6,"zipByteSize":17638},{"version":"1.0.3","createdAt":"2026-07-08T11:17:00.324Z","changelog":"Footer now uses /c/<slug> short link (fixes UTM truncation when SKILL.md is read in a terminal)","fileCount":5,"zipByteSize":12001},{"version":"1.0.2","createdAt":"2026-07-08T01:01:46.125Z","changelog":"Refreshed content + GitHub star link in footer","fileCount":5,"zipByteSize":12038},{"version":"1.0.1","createdAt":"2026-07-07T22:32:30.306Z","changelog":"Add catalog categories and topics","fileCount":5,"zipByteSize":11901},{"version":"1.0.0","createdAt":"2026-07-02T08:17:25.554Z","changelog":"Initial publish","fileCount":5,"zipByteSize":11855}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s17a4mqcnk515kvaca5ze55d0x88pfpx:repeated-games-reputation","setupComplexity":"low","setupSteps":["Setup complexity is classified as HIGH. You must provision dedicated cloud infrastructure or an isolated VM. Do not run this directly on your local workstation.","Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data."],"contract":{"contractStatus":"missing","authModes":[],"requires":[],"forbidden":[],"supportsMcp":false,"supportsA2a":false,"supportsStreaming":false,"inputSchemaRef":null,"outputSchemaRef":null,"dataRegion":null,"contractUpdatedAt":null,"sourceUpdatedAt":null,"freshnessSeconds":null},"invocationGuide":{"preferredApi":{"snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-11T17:44:43.865Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-deciqai-repeated-games-reputation/trust"}},"reliability":{"evidence":{"source":"runtime-metrics","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No trust, reliability, or runtime telemetry is available."},"trust":{"status":"unavailable","handshakeStatus":"UNKNOWN","verificationFreshnessHours":null,"reputationScore":null,"p95LatencyMs":null,"successRate30d":null,"fallbackRate":null,"attempts30d":null,"trustUpdatedAt":null,"trustConfidence":"unknown","sourceUpdatedAt":null,"freshnessSeconds":null},"decisionGuardrails":{"doNotUseIf":["Contract metadata is missing or unavailable for deterministic execution."],"safeUseWhen":[],"riskFlags":["missing_or_unavailable_contract","trust_data_unavailable","schema_references_missing"],"operationalConfidence":"low"},"executionMetrics":{"observedLatencyMsP50":null,"observedLatencyMsP95":null,"estimatedCostUsd":null,"uptime30d":null,"rateLimitRpm":null,"rateLimitBurst":null,"lastVerifiedAt":null,"verificationSource":null},"runtimeMetrics":{"successRate":null,"avgLatencyMs":null,"avgCostUsd":null,"hallucinationRate":null,"retryRate":null,"disputeRate":null,"p50Latency":null,"p95Latency":null,"lastUpdated":null}},"benchmarks":{"evidence":{"source":"no-benchmark-data","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No benchmark suites or observed failure patterns are available."},"suites":[],"failurePatterns":[]},"artifacts":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"high","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":null},"readme":"Skill: Repeated Games & Reputation\n\nOwner: deciqai\n\nSummary: Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe...\n\nTags: latest:1.0.5\n\nVersion history:\n\nv1.0.5 | 2026-07-16T18:13:41.749Z | user\n\nDescription tail link + agents machine-readable metadata line (deciqai.com/s/repeated-games-reputation.json)\n\nv1.0.4 | 2026-07-09T11:21:23.744Z | user\n\nRefresh: 2024-2026 AI-era worked examples added (strategy/leadership + systems/game-theory batch)\n\nv1.0.3 | 2026-07-08T11:17:00.324Z | user\n\nFooter now uses /c/<slug> short link (fixes UTM truncation when SKILL.md is read in a terminal)\n\nv1.0.2 | 2026-07-08T01:01:46.125Z | user\n\nRefreshed content + GitHub star link in footer\n\nv1.0.1 | 2026-07-07T22:32:30.306Z | user\n\nAdd catalog categories and topics\n\nv1.0.0 | 2026-07-02T08:17:25.554Z | user\n\nInitial publish\n\nArchive index:\n\nArchive v1.0.5: 6 files, 17833 bytes\n\nFiles: examples/ai-trust-reputation-enterprise-adoption-2024-2026.md (10415b), examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (3216b), skill-card.md (3050b), SKILL.md (11055b), _meta.json (144b)\n\nFile v1.0.5:SKILL.md\n\n---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry. More: deciqai.com/c/repeated-games-reputation\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with `prisoners-dilemma` · `second-order-thinking` · `signaling-games`.\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n- Trust/safety reputation is shaping who wins **AI adoption and AI-native competition** — where capability converges, a bad launch or safety incident reprices every future round of enterprise adoption (and the AI capex supercycle only lengthens the shadow of the future)\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use `prisoners-dilemma`\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough and your opponent will see your move.\" This skill works out the math of *enough* and *will see*.\n2. Check fit against When to Use / When NOT to use. If genuinely one-shot, redirect to `prisoners-dilemma`.\n3. Elicit their real repeated relationship or reputational concern — never run on a hypothetical when a real one is available.\n> **[WAIT — do not advance until user responds]**\n4. Walk the Analysis one element at a time: discount factor, observation structure, retaliation feasibility, forgiveness design.\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the one strategic move (TFT variant, reputational signal, structural change to δ or observation) that fits their situation.\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **Repeated-Game Analysis** across these steps:\n\n1. **Establish true repetition.** Indefinite/infinite → folk-theorem logic applies. Finite with known endpoint → backward induction risk; add uncertainty or commitment devices.\n2. **Estimate δ.** Cooperation threshold: δ ≥ (T − R) / (T − P). Below threshold → change the structure first; no strategy design saves it.\n3. **Confirm observability.** Perfect → TFT variants work. Noisy → Generous TFT (cooperate ~1/3 of the time after apparent defection) or Contrite TFT. Pure TFT under noise → recrimination spirals.\n4. **Select strategy.** TFT (clean bilateral) · Generous TFT (noisy) · Pavlov (mixed populations) · Grim Trigger (high-stakes, credible threat only) · benchmarks: Always Defect / Always Cooperate.\n5. **For reputation infrastructure:** design Observation · Aggregation · Persistence · Manipulation resistance — all four required; missing one breaks the system.\n6. **Stress-test endgame.** Mitigations: endpoint uncertainty; legacy concerns; successor obligations; overlapping generations.\n7. **Stop-rule:** lifetime cooperative payoff must beat one-shot defection by margin sufficient to absorb noise.\n\n### Output template\n\n```\nRepeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>\n```\n\n*→ Method in Action: [Robert Axelrod's Computer Tournament, 1979–1981](examples/robert-axelrod-computer-tournament-1979-1981.md)*\n*→ 2026 lens: [Trust Reputation as Strategy in the AI Race (2024–2026)](examples/ai-trust-reputation-enterprise-adoption-2024-2026.md)*\n\n## Pack: Reputation Infrastructure Patterns\n\nSix documented patterns (observation mechanism → known failure mode):\neBay/Airbnb (post-transaction ratings → 5-star inflation) · FICO (payment history → thin-file bias) · GitHub (commit history → popularity ≠ quality) · B2B scorecards (procurement records → approved-list lock-in) · Professional reputation (peer review + regulatory filings → old-boys' network slow to update) · Sovereign credit (macro indicators → rating-agency capture)\n\n## Applying It Well\n\nDiagnose δ first — wrong discount factor invalidates everything downstream. In noisy environments (most real ones), add forgiveness. Never confuse bilateral repetition (TFT) with third-party reputation markets (requires infrastructure). Legibility is a strategic asset: a simple strategy your counterparty can model beats a clever one they cannot.\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"We have a long relationship, so they won't defect\" | The relationship's *length* doesn't matter; the **shadow of the future** does. If their discount factor is low (they're about to retire, the firm is being sold, they have alternative partners lined up), past relationship duration provides no protection. Run the δ check, not the nostalgia check. |\n| [D] \"Reputation will discipline them\" | Only if the reputation system has all four pieces (observation, aggregation, persistence, manipulation resistance) and is actually consulted. Many \"reputation matters\" claims are wishful — the system is broken on one of the four pieces. |\n| [D] Applying TFT in a noisy environment without forgiveness | Pure TFT under noise enters mutual-recrimination death spirals. If observation has error, use Generous TFT, Contrite TFT, or Pavlov. Recommending TFT without checking noise level is a documented failure. |\n| [D] Using cooperation-by-default in a known finite-endpoint game | Backward induction: last round → defect dominates; second-to-last → both know this and defect; unraveling cascades to round one. Add uncertainty about endpoint or commitment devices. |\n| [D] \"Always-defect can't beat TFT, so cooperation is automatic\" | Always-defect can't beat TFT head-to-head, but can dominate in a population without retaliators. Tournament context matters — don't generalize from two-player simulation to a marketplace with unknown counterparties. |\n| [D] Designing a reputation system without manipulation resistance | A gameable system creates worse outcomes than no system, because the gamed signal substitutes for direct due diligence. Test every reputation system against adversarial gaming before deployment. |\n| [D] Confusing repeated game with reputation game | Repeated: same parties, bilateral, direct observation. Reputation: changing parties, third-party observation, requires infrastructure. Many \"reputation will solve it\" arguments fail because that infrastructure doesn't exist. |\n| [D] \"TFT is the winning strategy, period\" | TFT won clean-observation tournaments. Under noise, Generous TFT and Pavlov outperform it. Strategy is environment-conditional: δ, noise, population mix, and horizon all matter. |\n| [D] Adding excessive forgiveness \"to be nice\" | Over-forgiving strategies lose to exploiters. All four properties required: nice + retaliatory + forgiving + clear. Niceness without retaliation is exploited; the data is clear. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- \"Build trust\" recommended without specifying which mechanism creates it\n- δ check skipped (no estimate vs (T−R)/(T−P) threshold)\n- TFT recommended in noisy observation environment\n- Backward-induction risk ignored in finite-endpoint game\n- Reputation system proposed missing any of the four mechanisms\n- Strategy chosen for fairness rather than what wins the actual tournament\n- Bilateral repetition confused with third-party reputation market\n\n## Verification\n\n- [ ] Repetition structure diagnosed (indefinite/finite-known/finite-unknown; bilateral/reputational)\n- [ ] δ estimated and compared to (T−R)/(T−P) threshold\n- [ ] Observation noise assessed; forgiveness added if non-trivial\n- [ ] Strategy matched to specific environment (not off-the-shelf)\n- [ ] If reputation-based, all four infrastructure components specified\n- [ ] Endgame dynamics considered; mitigations named if known endpoint\n- [ ] Lifetime payoff beats one-shot defection by noise-absorbing margin\n- [ ] Alternative strategies compared and rejected with reasons\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/repeated-games-reputation** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\n*Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/repeated-games-reputation.json*\n\nFile v1.0.5:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.5\",\n  \"publishedAt\": 1784225621749\n}\n\nFile v1.0.5:references/sources.md\n\n# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\n### Contemporary context (2024–2026 AI example)\n\n- European Union (2024). Regulation (EU) 2024/1689 (the \"AI Act\") — the EU's risk-based AI regulation, which entered into force in 2024 with obligations phasing in through 2025–2026. Establishes a durable, publicly documented disclosure/reporting regime that functions as part of the *observation* and *persistence* layers of AI reputation infrastructure. Official text: https://eur-lex.europa.eu/eli/reg/2024/1689/oj\n- Public AI-lab safety documentation (2023–2025). Model cards, system cards, usage policies, and published safety / responsible-scaling frameworks from major AI labs (e.g. OpenAI, Anthropic, Google DeepMind) — the widely adopted industry practice of disclosing model capabilities and safety evaluations. Cited here as the real-world *legibility* and *clarity* instrument (per Axelrod's four properties) in the 2024–2026 enterprise-adoption repeated game. (See each lab's official model/system card and safety-framework pages.)\n\nFile v1.0.5:examples/ai-trust-reputation-enterprise-adoption-2024-2026.md\n\n# Method in Action: Trust Reputation as Strategy in the 2024–2026 AI Race\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nBetween 2024 and 2026, the competition among frontier AI labs and platforms (OpenAI, Anthropic, Google DeepMind, Meta, and others) became one of the clearest live demonstrations of repeated-game logic in a market visible to everyone. Model capability converged fast: for many enterprise tasks, the leading models were close substitutes. That convergence pushed a second variable to the front of the buying decision — **trust and safety reputation**. Enterprises signing multi-year contracts, embedding a model in regulated workflows, and exposing customer data to it are not running a one-shot transaction. They are opening an indefinitely repeated relationship, and they know it. In that structure, a single bad launch — a safety incident, a data-handling failure, a reckless capability release — is not a one-time cost. It reprices every future round of the game.\n\nThis example runs the anchor case through the skill's own **Repeated-Game Analysis**. The point is not to rank the labs; it is to show *why* reputation is the strategic asset it became, and where the analysis says a defection actually bites.\n\n### 1. Establish true repetition\n\nThe relationship between an AI vendor and its enterprise customers, its regulators, and the broader developer public is **indefinitely repeated**, and it is repeated on two layers at once:\n\n- **Bilateral repetition:** a specific enterprise customer renews, expands seats, adds workloads, and re-buys as new model versions ship. Each release is a fresh round with the same counterparty.\n- **Reputational / third-party layer:** thousands of *other* buyers, regulators, and journalists observe how the vendor handled the last incident and adjust their willingness to play. This is the more powerful layer, because the audience is huge and the moves are public.\n\nThere is no known, fixed endpoint — new model generations keep arriving, so folk-theorem logic applies rather than backward-induction unraveling. **Horizon: indefinite. Repetition: both bilateral and reputational.**\n\n### 2. Estimate δ (the shadow of the future)\n\nThe discount factor here is high for the labs, and that is the whole game. AI is a capital-intensive, subscription-and-usage-revenue business: the value of a customer is overwhelmingly in the *stream* of future renewals and expansion, not the first contract. When most of a customer's lifetime value sits in future rounds, δ is high, and the folk theorem says cooperation (ship responsibly, honor commitments, don't cut safety corners for a launch) is sustainable — *because* the discounted future cooperative payoff swamps the one-time gain from a reckless \"win this quarter\" defection.\n\nUsing the cooperation threshold δ ≥ (T − R) / (T − P): the temptation T (rush a flashy but unsafe release, harvest short-term headlines and signups) is real but bounded; the reward R (a durable, renewing, expanding enterprise relationship) is very large and recurring; the punishment P (enterprises freeze rollouts, regulators tighten scrutiny, the incident is cited for years) is severe and *persistent*. High R and high, lasting P relative to T push the required threshold down and put actual δ comfortably above it. **Cooperation is sustainable — and unusually so.**\n\nThe danger case is the actor with a *low* δ: a vendor chasing a one-time market-share land grab, a team measured only on this launch, or a startup that expects to be acquired before the next round. Low δ is exactly where the nostalgia check (\"we've partnered for years\") fails and the discount-factor check is the only thing that tells you the truth.\n\n### 3. Confirm observability (and its noise)\n\nObservation in this market is **partial and noisy**, which changes strategy. Genuine incidents get widely reported, but so do exaggerated ones; benchmark claims are contested; a model's refusal or a jailbreak demo can be misread as a systemic safety failure when it is an edge case, or vice versa. The signal an enterprise receives about a vendor's \"true\" trustworthiness is real but error-prone.\n\nBecause observation is noisy, **pure Tit-for-Tat is the wrong strategy for buyers and vendors alike.** A customer who permanently blacklists a vendor over one ambiguous incident, or a vendor that treats every critical press cycle as betrayal, enters the recrimination death spiral the skill warns about. The environment calls for **forgiveness built in** — Generous TFT: react to a genuine defection, but don't let a single noisy signal trigger permanent rupture.\n\n### 4. Select strategy\n\n- **For the labs (as players):** the winning strategy mirrors Axelrod's four properties — *nice* (don't be first to cut safety corners), *retaliatory* (defend reputation and correct misinformation, don't absorb bad-faith attacks passively), *forgiving* (rebuild after a legitimate mistake rather than going scorched-earth), and *clear* (publish model cards, usage policies, safety frameworks, and responsible-scaling commitments so the strategy is legible). **Clarity is the underrated move:** a lab whose safety posture buyers can actually model earns cooperation faster, exactly as legible TFT did in Axelrod's tournament. Publicly documented safety frameworks and system/model cards published by the major labs over 2023–2025 became, in effect, the legibility instrument.\n- **For enterprise buyers (as players):** **Generous TFT** — reward demonstrated reliability with expansion, respond to a real defection by pausing/diversifying, but forgive an isolated noisy incident rather than permanently exiting. Multi-vendor sourcing is the retaliation-feasibility mechanism: it keeps the threat credible without requiring a full divorce.\n\nFirst move: cooperate (adopt, publish standards). Retaliation: freeze rollout / diversify vendors / tighten scrutiny. Forgiveness: re-expand after a credible fix and post-incident transparency.\n\n### 5. Reputation infrastructure (the third-party layer)\n\nBecause much of the game is reputational rather than bilateral, the four required infrastructure components apply — and over 2024–2025 all four were being actively (if imperfectly) built:\n\n- **Observation:** third-party safety evaluations, red-teaming, independent benchmarks, incident reporting, and (in some jurisdictions, e.g. the EU AI Act phasing in over 2025–2026) mandatory disclosures create the raw signal.\n- **Aggregation:** procurement scorecards, analyst evaluations, and safety benchmark leaderboards compress many signals into a comparable rating.\n- **Persistence:** incidents are durable — a serious data or safety failure is cited for years, which is precisely what makes the punishment P large in step 2.\n- **Manipulation resistance:** the weakest link. Benchmark gaming, cherry-picked evals, and \"safety-washing\" are the documented failure modes; a gameable safety signal is worse than none because it substitutes for real due diligence. This is the component to stress-test hardest.\n\nMiss any one of these and the reputational discipline the labs are counting on stops working.\n\n### 6. Stress-test the endgame\n\nThe repeated-game logic weakens wherever the future stops mattering. Watch for: a vendor nearing acquisition or a cash crunch (δ collapses in the final rounds); a team incentivized purely on a single launch (local finite game inside the larger indefinite one); or a winner-take-all narrative that convinces one player the game ends after this generation. Mitigations that keep δ high: **legacy/brand concerns** (founders who expect to be in the market for decades), **successor obligations** (safety commitments and published frameworks that bind future releases), and **endpoint uncertainty** (nobody actually knows when — or if — the capability race ends, which is stabilizing). The AI capex supercycle *lengthens* the horizon rather than shortening it, which is good news for cooperation.\n\n### 7. Stop-rule\n\nDoes the lifetime cooperative payoff beat one-shot defection by a margin wide enough to absorb noise? For a well-capitalized lab with high δ: **yes, decisively.** The discounted stream of renewing enterprise relationships dwarfs any one-launch temptation, and the persistence of reputational punishment means a single defection is repaid across many future rounds. The margin is wide enough that even noisy misattributed incidents don't flip the calculus — which is exactly why building forgiveness (Generous TFT) into how buyers and vendors treat ambiguous signals is the correct, and not merely the nice, choice.\n\n### The takeaway\n\nThe 2024–2026 AI market is a textbook case that **reputation is not a soft value in a repeated game — it is the payoff-bearing asset.** When capability converges and the relationship repeats indefinitely with a huge watching audience, the shadow of the future does the disciplining. A bad launch is costly precisely *because the game repeats*: it doesn't cost you one deal, it reprices every future round and every observer's willingness to play. The skill's contribution is precision, not optimism — it tells you *why* the responsible move is also the winning move here (high δ, persistent punishment, noisy observation demanding forgiveness), and it flags the one place the logic breaks: any actor whose future has quietly stopped mattering.\n\n*Sources: Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984) — the four properties (nice, retaliatory, forgiving, clear) and the shadow-of-the-future logic applied here. Fudenberg, D. & Maskin, E., \"The Folk Theorem in Repeated Games with Discounting\" (Econometrica, 1986) — the discount-factor condition. Nowak, M. A. & Sigmund, K., \"Tit for Tat in Heterogeneous Populations\" (Nature, 1992) — Generous TFT under noise. On the 2024–2026 context: the EU AI Act (Regulation (EU) 2024/1689), which entered into force in 2024 with obligations phasing in through 2025–2026; and the widely adopted industry practice, documented in public model/system cards and published safety/responsible-scaling frameworks from major AI labs (OpenAI, Anthropic, Google DeepMind) over 2023–2025, of disclosing model capabilities and safety evaluations. Specific quantitative claims are omitted deliberately; assertions are limited to durable, publicly documented facts as of early 2026.*\n\nFile v1.0.5:examples/robert-axelrod-computer-tournament-1979-1981.md\n\n# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accounts for Tit for Tat's robust success is its combination of being nice, retaliatory, forgiving, and clear. Its niceness prevents it from getting into unnecessary trouble. Its retaliation discourages the other side from persisting whenever defection is tried. Its forgiveness helps restore mutual cooperation. And its clarity makes it intelligible to the other player, thereby eliciting long-term cooperation.\"\n\n— Axelrod, *The Evolution of Cooperation* (1984), p. 54.\n\nEach of these properties does specific work, and missing any of them costs you the tournament.\n\n**Niceness** (never first to defect) means you accumulate the (R,R) payoff with other nice strategies. The winners and most of the top-half finishers were nice; the bottom-half finishers were dominated by strategies that defected first. Axelrod's quantitative finding:\n\n> \"The single most important property of the rules that did well is what I called being nice... All of the top eight rules in the second tournament were nice, and none of the bottom seven was.\"\n\n— Axelrod (1984), p. 33.\n\nThe cost of starting with defection: against any retaliating opponent, you forfeit the long-run cooperative payoff stream for a one-time temptation gain. The math is brutal — across 200 rounds, the difference between 200R and 200P (using the standard payoffs, 600 vs 200) dwarfs any number of temptations.\n\n**Retaliation** (respond to defection) is what prevents exploitation by always-defect strategies. Strategies that were \"too nice\" — Tit-for-Two-Tats, Always Cooperate, and several over-forgiving variants — performed worse than TFT in noisy or mixed populations because always-defectors could exploit them.\n\n**Forgiveness** (return to cooperation when opponent does) is what distinguishes TFT from Grim Trigger. Both retaliate; only TFT recovers. In Axelrod's analysis, the cost of failing to forgive is borne in interactions with strategies that defect *occasionally* (whether by intention, accident, or in response to perceived defection). Grim Trigger versus Grim Trigger, once anyone defects by accident, is a death spiral. TFT versus TFT recovers in one round.\n\n**Clarity** (be simple enough to be understood) was the property Axelrod found most surprising. The losing strategies were often the cleverest — they had complex internal logic that other strategies could not model. The result was that opponents *could not learn* how to cooperate with them. The complex strategies, even when they performed reasonably against any individual opponent, lost the long-run tournament because they failed to *teach* their opponents to cooperate. TFT, by contrast, is so simple that any opponent can model it in one round: \"if I cooperate, it cooperates; if I defect, it defects.\" This makes cooperative coordination trivially learnable.\n\nThe episode teaches several things that running a repeated-game analysis correctly requires you to internalize.\n\n**First**, simplicity dominated complexity. The four-line program beat sophisticated multi-state automata. The mechanism: in a world of strategic agents, **legibility is a strategic asset**. If counterparties cannot model your strategy, they cannot learn how to cooperate with you. This generalizes far beyond Prisoner's Dilemma — clear pricing, clear reputation systems, clear contract terms outperform clever ones precisely because they are easy to coordinate around.\n\n**Second**, the tournament confirmed empirically what the Folk Theorem proved mathematically: in a repeated game with sufficient shadow of the future, cooperation is not a moral preference, it is a winning strategy. Nice strategies dominated unconditional defectors in long enough games. The strategy you should pick is not whichever is \"fairest\" or \"most ethical\" by intuition; it is whichever wins the tournament you're actually in.\n\n**Third**, the results are sensitive to noise. Axelrod's tournaments used clean observation — each program saw the previous move unambiguously. In follow-on research (Nowak & Sigmund 1992; Wu & Axelrod 1995), simulated noise was added: occasional misperception of cooperation as defection and vice versa. Under noise, pure TFT collapses into mutual recrimination, and *generous* TFT (cooperate with some probability even after observing defection) wins instead. **This is the practical lesson for any reputation system you design: assume some signal noise, and build in forgiveness — or you will create death spirals out of bookkeeping errors.**\n\n**Fourth**, the second tournament's design was crucial. The first tournament's result might have been luck. The second tournament was run with the strategies and outcomes of the first tournament fully published, with the express invitation to design beat-TFT strategies. Sixty-two entries tried. TFT won again. **This is what robust empirical evidence in strategy looks like**: the result survives adversarial follow-on examination by experts who had every motivation to falsify it.\n\n**Fifth**, the most important methodological point: **Axelrod's tournament is the founding empirical proof that mathematical game theory's pessimistic predictions are environment-conditional.** Classical theory said \"rational players in PD defect.\" Axelrod's tournament showed: in the repeated case, \"rational players cooperate, retaliate when needed, forgive, and stay legible — and that strategy beats every other strategy submitted by experts who knew the rules in advance.\" When you analyze a repeated game in your business or strategy work, you are not choosing between \"the cynical view\" and \"the optimistic view.\" You are choosing between **understanding the discount factor and observation structure** — in which case cooperation can be the dominant strategy — or **misdiagnosing the structure** — in which case you'll either defect prematurely or trust naively. The skill is not optimism; it is precision.\n\nFile v1.0.5:skill-card.md\n\n## Description:\n\nGuides agents through repeated-game and reputation analysis for trust, retaliation, cooperation, and reputation-system design when future interactions or third-party observers matter.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[deciqai](https://clawhub.ai/user/deciqai)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nEmployees, external users, and strategy-focused agents use this skill to assess whether cooperation can be sustained in repeated relationships, whether and how to retaliate after defection, and how to design reputation infrastructure for marketplaces or observed one-shot interactions.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: The skill may influence business, relationship, or retaliation strategy based on incomplete user-provided context.\n\nMitigation: Review recommendations before acting and verify the repeated-game assumptions, observability, and payoff estimates against the real situation.\n\nRisk: Reputation-system guidance can be misapplied if the required observation, aggregation, persistence, or manipulation-resistance mechanisms are absent.\n\nMitigation: Check all four reputation-infrastructure components explicitly before relying on reputation to discipline behavior.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/repeated-games-reputation)\n- [Publisher profile](https://clawhub.ai/user/deciqai)\n- [Machine-readable skill metadata](https://www.deciqai.com/s/repeated-games-reputation.json)\n- [Primary sources](references/sources.md)\n- [Robert Axelrod's Computer Tournament, 1979-1981](examples/robert-axelrod-computer-tournament-1979-1981.md)\n- [Trust Reputation as Strategy in the 2024-2026 AI Race](examples/ai-trust-reputation-enterprise-adoption-2024-2026.md)\n- [Axelrod and Hamilton, The Evolution of Cooperation](https://doi.org/10.1126/science.7466396)\n- [Friedman, A Non-cooperative Equilibrium for Supergames](https://doi.org/10.2307/2296617)\n- [Fudenberg and Maskin, The Folk Theorem in Repeated Games](https://doi.org/10.2307/1911307)\n- [Nowak and Sigmund, Tit for Tat in Heterogeneous Populations](https://doi.org/10.1038/355250a0)\n- [Nowak and Sigmund, Win-Stay Lose-Shift](https://doi.org/10.1038/364056a0)\n- [EU Artificial Intelligence Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, guidance]\n\n**Output Format:** [Markdown repeated-game analysis with strategy recommendations]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [Can include discount-factor checks, observability assessment, strategy selection, endgame mitigations, and reputation-infrastructure review.]\n\n## Skill Version(s):\n\n1.0.5 (source: server release evidence)\n\n## Ethical Considerations:\n\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.\n\nArchive v1.0.4: 6 files, 17638 bytes\n\nFiles: examples/ai-trust-reputation-enterprise-adoption-2024-2026.md (10415b), examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (3216b), skill-card.md (2714b), SKILL.md (10893b), _meta.json (144b)\n\nFile v1.0.4:SKILL.md\n\n---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry.\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with `prisoners-dilemma` · `second-order-thinking` · `signaling-games`.\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n- Trust/safety reputation is shaping who wins **AI adoption and AI-native competition** — where capability converges, a bad launch or safety incident reprices every future round of enterprise adoption (and the AI capex supercycle only lengthens the shadow of the future)\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use `prisoners-dilemma`\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough and your opponent will see your move.\" This skill works out the math of *enough* and *will see*.\n2. Check fit against When to Use / When NOT to use. If genuinely one-shot, redirect to `prisoners-dilemma`.\n3. Elicit their real repeated relationship or reputational concern — never run on a hypothetical when a real one is available.\n> **[WAIT — do not advance until user responds]**\n4. Walk the Analysis one element at a time: discount factor, observation structure, retaliation feasibility, forgiveness design.\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the one strategic move (TFT variant, reputational signal, structural change to δ or observation) that fits their situation.\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **Repeated-Game Analysis** across these steps:\n\n1. **Establish true repetition.** Indefinite/infinite → folk-theorem logic applies. Finite with known endpoint → backward induction risk; add uncertainty or commitment devices.\n2. **Estimate δ.** Cooperation threshold: δ ≥ (T − R) / (T − P). Below threshold → change the structure first; no strategy design saves it.\n3. **Confirm observability.** Perfect → TFT variants work. Noisy → Generous TFT (cooperate ~1/3 of the time after apparent defection) or Contrite TFT. Pure TFT under noise → recrimination spirals.\n4. **Select strategy.** TFT (clean bilateral) · Generous TFT (noisy) · Pavlov (mixed populations) · Grim Trigger (high-stakes, credible threat only) · benchmarks: Always Defect / Always Cooperate.\n5. **For reputation infrastructure:** design Observation · Aggregation · Persistence · Manipulation resistance — all four required; missing one breaks the system.\n6. **Stress-test endgame.** Mitigations: endpoint uncertainty; legacy concerns; successor obligations; overlapping generations.\n7. **Stop-rule:** lifetime cooperative payoff must beat one-shot defection by margin sufficient to absorb noise.\n\n### Output template\n\n```\nRepeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>\n```\n\n*→ Method in Action: [Robert Axelrod's Computer Tournament, 1979–1981](examples/robert-axelrod-computer-tournament-1979-1981.md)*\n*→ 2026 lens: [Trust Reputation as Strategy in the AI Race (2024–2026)](examples/ai-trust-reputation-enterprise-adoption-2024-2026.md)*\n\n## Pack: Reputation Infrastructure Patterns\n\nSix documented patterns (observation mechanism → known failure mode):\neBay/Airbnb (post-transaction ratings → 5-star inflation) · FICO (payment history → thin-file bias) · GitHub (commit history → popularity ≠ quality) · B2B scorecards (procurement records → approved-list lock-in) · Professional reputation (peer review + regulatory filings → old-boys' network slow to update) · Sovereign credit (macro indicators → rating-agency capture)\n\n## Applying It Well\n\nDiagnose δ first — wrong discount factor invalidates everything downstream. In noisy environments (most real ones), add forgiveness. Never confuse bilateral repetition (TFT) with third-party reputation markets (requires infrastructure). Legibility is a strategic asset: a simple strategy your counterparty can model beats a clever one they cannot.\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"We have a long relationship, so they won't defect\" | The relationship's *length* doesn't matter; the **shadow of the future** does. If their discount factor is low (they're about to retire, the firm is being sold, they have alternative partners lined up), past relationship duration provides no protection. Run the δ check, not the nostalgia check. |\n| [D] \"Reputation will discipline them\" | Only if the reputation system has all four pieces (observation, aggregation, persistence, manipulation resistance) and is actually consulted. Many \"reputation matters\" claims are wishful — the system is broken on one of the four pieces. |\n| [D] Applying TFT in a noisy environment without forgiveness | Pure TFT under noise enters mutual-recrimination death spirals. If observation has error, use Generous TFT, Contrite TFT, or Pavlov. Recommending TFT without checking noise level is a documented failure. |\n| [D] Using cooperation-by-default in a known finite-endpoint game | Backward induction: last round → defect dominates; second-to-last → both know this and defect; unraveling cascades to round one. Add uncertainty about endpoint or commitment devices. |\n| [D] \"Always-defect can't beat TFT, so cooperation is automatic\" | Always-defect can't beat TFT head-to-head, but can dominate in a population without retaliators. Tournament context matters — don't generalize from two-player simulation to a marketplace with unknown counterparties. |\n| [D] Designing a reputation system without manipulation resistance | A gameable system creates worse outcomes than no system, because the gamed signal substitutes for direct due diligence. Test every reputation system against adversarial gaming before deployment. |\n| [D] Confusing repeated game with reputation game | Repeated: same parties, bilateral, direct observation. Reputation: changing parties, third-party observation, requires infrastructure. Many \"reputation will solve it\" arguments fail because that infrastructure doesn't exist. |\n| [D] \"TFT is the winning strategy, period\" | TFT won clean-observation tournaments. Under noise, Generous TFT and Pavlov outperform it. Strategy is environment-conditional: δ, noise, population mix, and horizon all matter. |\n| [D] Adding excessive forgiveness \"to be nice\" | Over-forgiving strategies lose to exploiters. All four properties required: nice + retaliatory + forgiving + clear. Niceness without retaliation is exploited; the data is clear. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- \"Build trust\" recommended without specifying which mechanism creates it\n- δ check skipped (no estimate vs (T−R)/(T−P) threshold)\n- TFT recommended in noisy observation environment\n- Backward-induction risk ignored in finite-endpoint game\n- Reputation system proposed missing any of the four mechanisms\n- Strategy chosen for fairness rather than what wins the actual tournament\n- Bilateral repetition confused with third-party reputation market\n\n## Verification\n\n- [ ] Repetition structure diagnosed (indefinite/finite-known/finite-unknown; bilateral/reputational)\n- [ ] δ estimated and compared to (T−R)/(T−P) threshold\n- [ ] Observation noise assessed; forgiveness added if non-trivial\n- [ ] Strategy matched to specific environment (not off-the-shelf)\n- [ ] If reputation-based, all four infrastructure components specified\n- [ ] Endgame dynamics considered; mitigations named if known endpoint\n- [ ] Lifetime payoff beats one-shot defection by noise-absorbing margin\n- [ ] Alternative strategies compared and rejected with reasons\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 189 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/repeated-games-reputation** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.4:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.4\",\n  \"publishedAt\": 1783596083744\n}\n\nFile v1.0.4:references/sources.md\n\n# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\n### Contemporary context (2024–2026 AI example)\n\n- European Union (2024). Regulation (EU) 2024/1689 (the \"AI Act\") — the EU's risk-based AI regulation, which entered into force in 2024 with obligations phasing in through 2025–2026. Establishes a durable, publicly documented disclosure/reporting regime that functions as part of the *observation* and *persistence* layers of AI reputation infrastructure. Official text: https://eur-lex.europa.eu/eli/reg/2024/1689/oj\n- Public AI-lab safety documentation (2023–2025). Model cards, system cards, usage policies, and published safety / responsible-scaling frameworks from major AI labs (e.g. OpenAI, Anthropic, Google DeepMind) — the widely adopted industry practice of disclosing model capabilities and safety evaluations. Cited here as the real-world *legibility* and *clarity* instrument (per Axelrod's four properties) in the 2024–2026 enterprise-adoption repeated game. (See each lab's official model/system card and safety-framework pages.)\n\nFile v1.0.4:examples/ai-trust-reputation-enterprise-adoption-2024-2026.md\n\n# Method in Action: Trust Reputation as Strategy in the 2024–2026 AI Race\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nBetween 2024 and 2026, the competition among frontier AI labs and platforms (OpenAI, Anthropic, Google DeepMind, Meta, and others) became one of the clearest live demonstrations of repeated-game logic in a market visible to everyone. Model capability converged fast: for many enterprise tasks, the leading models were close substitutes. That convergence pushed a second variable to the front of the buying decision — **trust and safety reputation**. Enterprises signing multi-year contracts, embedding a model in regulated workflows, and exposing customer data to it are not running a one-shot transaction. They are opening an indefinitely repeated relationship, and they know it. In that structure, a single bad launch — a safety incident, a data-handling failure, a reckless capability release — is not a one-time cost. It reprices every future round of the game.\n\nThis example runs the anchor case through the skill's own **Repeated-Game Analysis**. The point is not to rank the labs; it is to show *why* reputation is the strategic asset it became, and where the analysis says a defection actually bites.\n\n### 1. Establish true repetition\n\nThe relationship between an AI vendor and its enterprise customers, its regulators, and the broader developer public is **indefinitely repeated**, and it is repeated on two layers at once:\n\n- **Bilateral repetition:** a specific enterprise customer renews, expands seats, adds workloads, and re-buys as new model versions ship. Each release is a fresh round with the same counterparty.\n- **Reputational / third-party layer:** thousands of *other* buyers, regulators, and journalists observe how the vendor handled the last incident and adjust their willingness to play. This is the more powerful layer, because the audience is huge and the moves are public.\n\nThere is no known, fixed endpoint — new model generations keep arriving, so folk-theorem logic applies rather than backward-induction unraveling. **Horizon: indefinite. Repetition: both bilateral and reputational.**\n\n### 2. Estimate δ (the shadow of the future)\n\nThe discount factor here is high for the labs, and that is the whole game. AI is a capital-intensive, subscription-and-usage-revenue business: the value of a customer is overwhelmingly in the *stream* of future renewals and expansion, not the first contract. When most of a customer's lifetime value sits in future rounds, δ is high, and the folk theorem says cooperation (ship responsibly, honor commitments, don't cut safety corners for a launch) is sustainable — *because* the discounted future cooperative payoff swamps the one-time gain from a reckless \"win this quarter\" defection.\n\nUsing the cooperation threshold δ ≥ (T − R) / (T − P): the temptation T (rush a flashy but unsafe release, harvest short-term headlines and signups) is real but bounded; the reward R (a durable, renewing, expanding enterprise relationship) is very large and recurring; the punishment P (enterprises freeze rollouts, regulators tighten scrutiny, the incident is cited for years) is severe and *persistent*. High R and high, lasting P relative to T push the required threshold down and put actual δ comfortably above it. **Cooperation is sustainable — and unusually so.**\n\nThe danger case is the actor with a *low* δ: a vendor chasing a one-time market-share land grab, a team measured only on this launch, or a startup that expects to be acquired before the next round. Low δ is exactly where the nostalgia check (\"we've partnered for years\") fails and the discount-factor check is the only thing that tells you the truth.\n\n### 3. Confirm observability (and its noise)\n\nObservation in this market is **partial and noisy**, which changes strategy. Genuine incidents get widely reported, but so do exaggerated ones; benchmark claims are contested; a model's refusal or a jailbreak demo can be misread as a systemic safety failure when it is an edge case, or vice versa. The signal an enterprise receives about a vendor's \"true\" trustworthiness is real but error-prone.\n\nBecause observation is noisy, **pure Tit-for-Tat is the wrong strategy for buyers and vendors alike.** A customer who permanently blacklists a vendor over one ambiguous incident, or a vendor that treats every critical press cycle as betrayal, enters the recrimination death spiral the skill warns about. The environment calls for **forgiveness built in** — Generous TFT: react to a genuine defection, but don't let a single noisy signal trigger permanent rupture.\n\n### 4. Select strategy\n\n- **For the labs (as players):** the winning strategy mirrors Axelrod's four properties — *nice* (don't be first to cut safety corners), *retaliatory* (defend reputation and correct misinformation, don't absorb bad-faith attacks passively), *forgiving* (rebuild after a legitimate mistake rather than going scorched-earth), and *clear* (publish model cards, usage policies, safety frameworks, and responsible-scaling commitments so the strategy is legible). **Clarity is the underrated move:** a lab whose safety posture buyers can actually model earns cooperation faster, exactly as legible TFT did in Axelrod's tournament. Publicly documented safety frameworks and system/model cards published by the major labs over 2023–2025 became, in effect, the legibility instrument.\n- **For enterprise buyers (as players):** **Generous TFT** — reward demonstrated reliability with expansion, respond to a real defection by pausing/diversifying, but forgive an isolated noisy incident rather than permanently exiting. Multi-vendor sourcing is the retaliation-feasibility mechanism: it keeps the threat credible without requiring a full divorce.\n\nFirst move: cooperate (adopt, publish standards). Retaliation: freeze rollout / diversify vendors / tighten scrutiny. Forgiveness: re-expand after a credible fix and post-incident transparency.\n\n### 5. Reputation infrastructure (the third-party layer)\n\nBecause much of the game is reputational rather than bilateral, the four required infrastructure components apply — and over 2024–2025 all four were being actively (if imperfectly) built:\n\n- **Observation:** third-party safety evaluations, red-teaming, independent benchmarks, incident reporting, and (in some jurisdictions, e.g. the EU AI Act phasing in over 2025–2026) mandatory disclosures create the raw signal.\n- **Aggregation:** procurement scorecards, analyst evaluations, and safety benchmark leaderboards compress many signals into a comparable rating.\n- **Persistence:** incidents are durable — a serious data or safety failure is cited for years, which is precisely what makes the punishment P large in step 2.\n- **Manipulation resistance:** the weakest link. Benchmark gaming, cherry-picked evals, and \"safety-washing\" are the documented failure modes; a gameable safety signal is worse than none because it substitutes for real due diligence. This is the component to stress-test hardest.\n\nMiss any one of these and the reputational discipline the labs are counting on stops working.\n\n### 6. Stress-test the endgame\n\nThe repeated-game logic weakens wherever the future stops mattering. Watch for: a vendor nearing acquisition or a cash crunch (δ collapses in the final rounds); a team incentivized purely on a single launch (local finite game inside the larger indefinite one); or a winner-take-all narrative that convinces one player the game ends after this generation. Mitigations that keep δ high: **legacy/brand concerns** (founders who expect to be in the market for decades), **successor obligations** (safety commitments and published frameworks that bind future releases), and **endpoint uncertainty** (nobody actually knows when — or if — the capability race ends, which is stabilizing). The AI capex supercycle *lengthens* the horizon rather than shortening it, which is good news for cooperation.\n\n### 7. Stop-rule\n\nDoes the lifetime cooperative payoff beat one-shot defection by a margin wide enough to absorb noise? For a well-capitalized lab with high δ: **yes, decisively.** The discounted stream of renewing enterprise relationships dwarfs any one-launch temptation, and the persistence of reputational punishment means a single defection is repaid across many future rounds. The margin is wide enough that even noisy misattributed incidents don't flip the calculus — which is exactly why building forgiveness (Generous TFT) into how buyers and vendors treat ambiguous signals is the correct, and not merely the nice, choice.\n\n### The takeaway\n\nThe 2024–2026 AI market is a textbook case that **reputation is not a soft value in a repeated game — it is the payoff-bearing asset.** When capability converges and the relationship repeats indefinitely with a huge watching audience, the shadow of the future does the disciplining. A bad launch is costly precisely *because the game repeats*: it doesn't cost you one deal, it reprices every future round and every observer's willingness to play. The skill's contribution is precision, not optimism — it tells you *why* the responsible move is also the winning move here (high δ, persistent punishment, noisy observation demanding forgiveness), and it flags the one place the logic breaks: any actor whose future has quietly stopped mattering.\n\n*Sources: Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984) — the four properties (nice, retaliatory, forgiving, clear) and the shadow-of-the-future logic applied here. Fudenberg, D. & Maskin, E., \"The Folk Theorem in Repeated Games with Discounting\" (Econometrica, 1986) — the discount-factor condition. Nowak, M. A. & Sigmund, K., \"Tit for Tat in Heterogeneous Populations\" (Nature, 1992) — Generous TFT under noise. On the 2024–2026 context: the EU AI Act (Regulation (EU) 2024/1689), which entered into force in 2024 with obligations phasing in through 2025–2026; and the widely adopted industry practice, documented in public model/system cards and published safety/responsible-scaling frameworks from major AI labs (OpenAI, Anthropic, Google DeepMind) over 2023–2025, of disclosing model capabilities and safety evaluations. Specific quantitative claims are omitted deliberately; assertions are limited to durable, publicly documented facts as of early 2026.*\n\nFile v1.0.4:examples/robert-axelrod-computer-tournament-1979-1981.md\n\n# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accounts for Tit for Tat's robust success is its combination of being nice, retaliatory, forgiving, and clear. Its niceness prevents it from getting into unnecessary trouble. Its retaliation discourages the other side from persisting whenever defection is tried. Its forgiveness helps restore mutual cooperation. And its clarity makes it intelligible to the other player, thereby eliciting long-term cooperation.\"\n\n— Axelrod, *The Evolution of Cooperation* (1984), p. 54.\n\nEach of these properties does specific work, and missing any of them costs you the tournament.\n\n**Niceness** (never first to defect) means you accumulate the (R,R) payoff with other nice strategies. The winners and most of the top-half finishers were nice; the bottom-half finishers were dominated by strategies that defected first. Axelrod's quantitative finding:\n\n> \"The single most important property of the rules that did well is what I called being nice... All of the top eight rules in the second tournament were nice, and none of the bottom seven was.\"\n\n— Axelrod (1984), p. 33.\n\nThe cost of starting with defection: against any retaliating opponent, you forfeit the long-run cooperative payoff stream for a one-time temptation gain. The math is brutal — across 200 rounds, the difference between 200R and 200P (using the standard payoffs, 600 vs 200) dwarfs any number of temptations.\n\n**Retaliation** (respond to defection) is what prevents exploitation by always-defect strategies. Strategies that were \"too nice\" — Tit-for-Two-Tats, Always Cooperate, and several over-forgiving variants — performed worse than TFT in noisy or mixed populations because always-defectors could exploit them.\n\n**Forgiveness** (return to cooperation when opponent does) is what distinguishes TFT from Grim Trigger. Both retaliate; only TFT recovers. In Axelrod's analysis, the cost of failing to forgive is borne in interactions with strategies that defect *occasionally* (whether by intention, accident, or in response to perceived defection). Grim Trigger versus Grim Trigger, once anyone defects by accident, is a death spiral. TFT versus TFT recovers in one round.\n\n**Clarity** (be simple enough to be understood) was the property Axelrod found most surprising. The losing strategies were often the cleverest — they had complex internal logic that other strategies could not model. The result was that opponents *could not learn* how to cooperate with them. The complex strategies, even when they performed reasonably against any individual opponent, lost the long-run tournament because they failed to *teach* their opponents to cooperate. TFT, by contrast, is so simple that any opponent can model it in one round: \"if I cooperate, it cooperates; if I defect, it defects.\" This makes cooperative coordination trivially learnable.\n\nThe episode teaches several things that running a repeated-game analysis correctly requires you to internalize.\n\n**First**, simplicity dominated complexity. The four-line program beat sophisticated multi-state automata. The mechanism: in a world of strategic agents, **legibility is a strategic asset**. If counterparties cannot model your strategy, they cannot learn how to cooperate with you. This generalizes far beyond Prisoner's Dilemma — clear pricing, clear reputation systems, clear contract terms outperform clever ones precisely because they are easy to coordinate around.\n\n**Second**, the tournament confirmed empirically what the Folk Theorem proved mathematically: in a repeated game with sufficient shadow of the future, cooperation is not a moral preference, it is a winning strategy. Nice strategies dominated unconditional defectors in long enough games. The strategy you should pick is not whichever is \"fairest\" or \"most ethical\" by intuition; it is whichever wins the tournament you're actually in.\n\n**Third**, the results are sensitive to noise. Axelrod's tournaments used clean observation — each program saw the previous move unambiguously. In follow-on research (Nowak & Sigmund 1992; Wu & Axelrod 1995), simulated noise was added: occasional misperception of cooperation as defection and vice versa. Under noise, pure TFT collapses into mutual recrimination, and *generous* TFT (cooperate with some probability even after observing defection) wins instead. **This is the practical lesson for any reputation system you design: assume some signal noise, and build in forgiveness — or you will create death spirals out of bookkeeping errors.**\n\n**Fourth**, the second tournament's design was crucial. The first tournament's result might have been luck. The second tournament was run with the strategies and outcomes of the first tournament fully published, with the express invitation to design beat-TFT strategies. Sixty-two entries tried. TFT won again. **This is what robust empirical evidence in strategy looks like**: the result survives adversarial follow-on examination by experts who had every motivation to falsify it.\n\n**Fifth**, the most important methodological point: **Axelrod's tournament is the founding empirical proof that mathematical game theory's pessimistic predictions are environment-conditional.** Classical theory said \"rational players in PD defect.\" Axelrod's tournament showed: in the repeated case, \"rational players cooperate, retaliate when needed, forgive, and stay legible — and that strategy beats every other strategy submitted by experts who knew the rules in advance.\" When you analyze a repeated game in your business or strategy work, you are not choosing between \"the cynical view\" and \"the optimistic view.\" You are choosing between **understanding the discount factor and observation structure** — in which case cooperation can be the dominant strategy — or **misdiagnosing the structure** — in which case you'll either defect prematurely or trust naively. The skill is not optimism; it is precision.\n\nFile v1.0.4:skill-card.md\n\n## Description: <br>\nHelps agents analyze repeated relationships and reputation systems by checking whether cooperation is sustainable, selecting an environment-appropriate repeated-game strategy, and identifying reputation-infrastructure requirements. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nExternal users, developers, and business operators use this skill for decision-support when evaluating trust, retaliation, cooperation, and reputation-system design in repeated or publicly observed interactions. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: Decision-support analysis could be mistaken for an instruction to act automatically in business, financial, or operational settings. <br>\nMitigation: Review outputs as analysis before taking action, especially where relationship, market, or operational consequences are material. <br>\nRisk: The quality of recommendations depends on accurate estimates of repetition, discount factor, observability, and reputation-system completeness. <br>\nMitigation: Use the skill's verification checklist to confirm those assumptions and treat uncertain values as inputs for human review. <br>\n\n\n## Reference(s): <br>\n- [ClawHub skill page](https://clawhub.ai/deciqai/skills/repeated-games-reputation) <br>\n- [Primary sources](references/sources.md) <br>\n- [Robert Axelrod's Computer Tournament, 1979-1981](examples/robert-axelrod-computer-tournament-1979-1981.md) <br>\n- [Trust Reputation as Strategy in the AI Race, 2024-2026](examples/ai-trust-reputation-enterprise-adoption-2024-2026.md) <br>\n- [Axelrod and Hamilton, The Evolution of Cooperation](https://doi.org/10.1126/science.7466396) <br>\n- [Fudenberg and Maskin, Folk Theorem in Repeated Games](https://doi.org/10.2307/1911307) <br>\n- [EU AI Act official text](https://eur-lex.europa.eu/eli/reg/2024/1689/oj) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [text, markdown, guidance] <br>\n**Output Format:** [Markdown analysis using the skill's repeated-game template] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [May ask step-by-step coaching questions before producing analysis when the user has no concrete case.] <br>\n\n## Skill Version(s): <br>\n1.0.4 (source: server release evidence) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.3: 5 files, 12001 bytes\n\nFiles: examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (2209b), skill-card.md (2976b), SKILL.md (10480b), _meta.json (144b)\n\nFile v1.0.3:SKILL.md\n\n---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry.\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with `prisoners-dilemma` · `second-order-thinking` · `signaling-games`.\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use `prisoners-dilemma`\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough and your opponent will see your move.\" This skill works out the math of *enough* and *will see*.\n2. Check fit against When to Use / When NOT to use. If genuinely one-shot, redirect to `prisoners-dilemma`.\n3. Elicit their real repeated relationship or reputational concern — never run on a hypothetical when a real one is available.\n> **[WAIT — do not advance until user responds]**\n4. Walk the Analysis one element at a time: discount factor, observation structure, retaliation feasibility, forgiveness design.\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the one strategic move (TFT variant, reputational signal, structural change to δ or observation) that fits their situation.\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **Repeated-Game Analysis** across these steps:\n\n1. **Establish true repetition.** Indefinite/infinite → folk-theorem logic applies. Finite with known endpoint → backward induction risk; add uncertainty or commitment devices.\n2. **Estimate δ.** Cooperation threshold: δ ≥ (T − R) / (T − P). Below threshold → change the structure first; no strategy design saves it.\n3. **Confirm observability.** Perfect → TFT variants work. Noisy → Generous TFT (cooperate ~1/3 of the time after apparent defection) or Contrite TFT. Pure TFT under noise → recrimination spirals.\n4. **Select strategy.** TFT (clean bilateral) · Generous TFT (noisy) · Pavlov (mixed populations) · Grim Trigger (high-stakes, credible threat only) · benchmarks: Always Defect / Always Cooperate.\n5. **For reputation infrastructure:** design Observation · Aggregation · Persistence · Manipulation resistance — all four required; missing one breaks the system.\n6. **Stress-test endgame.** Mitigations: endpoint uncertainty; legacy concerns; successor obligations; overlapping generations.\n7. **Stop-rule:** lifetime cooperative payoff must beat one-shot defection by margin sufficient to absorb noise.\n\n### Output template\n\n```\nRepeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>\n```\n\n*→ Method in Action: [Robert Axelrod's Computer Tournament, 1979–1981](examples/robert-axelrod-computer-tournament-1979-1981.md)*\n\n## Pack: Reputation Infrastructure Patterns\n\nSix documented patterns (observation mechanism → known failure mode):\neBay/Airbnb (post-transaction ratings → 5-star inflation) · FICO (payment history → thin-file bias) · GitHub (commit history → popularity ≠ quality) · B2B scorecards (procurement records → approved-list lock-in) · Professional reputation (peer review + regulatory filings → old-boys' network slow to update) · Sovereign credit (macro indicators → rating-agency capture)\n\n## Applying It Well\n\nDiagnose δ first — wrong discount factor invalidates everything downstream. In noisy environments (most real ones), add forgiveness. Never confuse bilateral repetition (TFT) with third-party reputation markets (requires infrastructure). Legibility is a strategic asset: a simple strategy your counterparty can model beats a clever one they cannot.\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"We have a long relationship, so they won't defect\" | The relationship's *length* doesn't matter; the **shadow of the future** does. If their discount factor is low (they're about to retire, the firm is being sold, they have alternative partners lined up), past relationship duration provides no protection. Run the δ check, not the nostalgia check. |\n| [D] \"Reputation will discipline them\" | Only if the reputation system has all four pieces (observation, aggregation, persistence, manipulation resistance) and is actually consulted. Many \"reputation matters\" claims are wishful — the system is broken on one of the four pieces. |\n| [D] Applying TFT in a noisy environment without forgiveness | Pure TFT under noise enters mutual-recrimination death spirals. If observation has error, use Generous TFT, Contrite TFT, or Pavlov. Recommending TFT without checking noise level is a documented failure. |\n| [D] Using cooperation-by-default in a known finite-endpoint game | Backward induction: last round → defect dominates; second-to-last → both know this and defect; unraveling cascades to round one. Add uncertainty about endpoint or commitment devices. |\n| [D] \"Always-defect can't beat TFT, so cooperation is automatic\" | Always-defect can't beat TFT head-to-head, but can dominate in a population without retaliators. Tournament context matters — don't generalize from two-player simulation to a marketplace with unknown counterparties. |\n| [D] Designing a reputation system without manipulation resistance | A gameable system creates worse outcomes than no system, because the gamed signal substitutes for direct due diligence. Test every reputation system against adversarial gaming before deployment. |\n| [D] Confusing repeated game with reputation game | Repeated: same parties, bilateral, direct observation. Reputation: changing parties, third-party observation, requires infrastructure. Many \"reputation will solve it\" arguments fail because that infrastructure doesn't exist. |\n| [D] \"TFT is the winning strategy, period\" | TFT won clean-observation tournaments. Under noise, Generous TFT and Pavlov outperform it. Strategy is environment-conditional: δ, noise, population mix, and horizon all matter. |\n| [D] Adding excessive forgiveness \"to be nice\" | Over-forgiving strategies lose to exploiters. All four properties required: nice + retaliatory + forgiving + clear. Niceness without retaliation is exploited; the data is clear. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- \"Build trust\" recommended without specifying which mechanism creates it\n- δ check skipped (no estimate vs (T−R)/(T−P) threshold)\n- TFT recommended in noisy observation environment\n- Backward-induction risk ignored in finite-endpoint game\n- Reputation system proposed missing any of the four mechanisms\n- Strategy chosen for fairness rather than what wins the actual tournament\n- Bilateral repetition confused with third-party reputation market\n\n## Verification\n\n- [ ] Repetition structure diagnosed (indefinite/finite-known/finite-unknown; bilateral/reputational)\n- [ ] δ estimated and compared to (T−R)/(T−P) threshold\n- [ ] Observation noise assessed; forgiveness added if non-trivial\n- [ ] Strategy matched to specific environment (not off-the-shelf)\n- [ ] If reputation-based, all four infrastructure components specified\n- [ ] Endgame dynamics considered; mitigations named if known endpoint\n- [ ] Lifetime payoff beats one-shot defection by noise-absorbing margin\n- [ ] Alternative strategies compared and rejected with reasons\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 164 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/c/repeated-games-reputation** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.3:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.3\",\n  \"publishedAt\": 1783509420324\n}\n\nFile v1.0.3:references/sources.md\n\n# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\nFile v1.0.3:examples/robert-axelrod-computer-tournament-1979-1981.md\n\n# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accounts for Tit for Tat's robust success is its combination of being nice, retaliatory, forgiving, and clear. Its niceness prevents it from getting into unnecessary trouble. Its retaliation discourages the other side from persisting whenever defection is tried. Its forgiveness helps restore mutual cooperation. And its clarity makes it intelligible to the other player, thereby eliciting long-term cooperation.\"\n\n— Axelrod, *The Evolution of Cooperation* (1984), p. 54.\n\nEach of these properties does specific work, and missing any of them costs you the tournament.\n\n**Niceness** (never first to defect) means you accumulate the (R,R) payoff with other nice strategies. The winners and most of the top-half finishers were nice; the bottom-half finishers were dominated by strategies that defected first. Axelrod's quantitative finding:\n\n> \"The single most important property of the rules that did well is what I called being nice... All of the top eight rules in the second tournament were nice, and none of the bottom seven was.\"\n\n— Axelrod (1984), p. 33.\n\nThe cost of starting with defection: against any retaliating opponent, you forfeit the long-run cooperative payoff stream for a one-time temptation gain. The math is brutal — across 200 rounds, the difference between 200R and 200P (using the standard payoffs, 600 vs 200) dwarfs any number of temptations.\n\n**Retaliation** (respond to defection) is what prevents exploitation by always-defect strategies. Strategies that were \"too nice\" — Tit-for-Two-Tats, Always Cooperate, and several over-forgiving variants — performed worse than TFT in noisy or mixed populations because always-defectors could exploit them.\n\n**Forgiveness** (return to cooperation when opponent does) is what distinguishes TFT from Grim Trigger. Both retaliate; only TFT recovers. In Axelrod's analysis, the cost of failing to forgive is borne in interactions with strategies that defect *occasionally* (whether by intention, accident, or in response to perceived defection). Grim Trigger versus Grim Trigger, once anyone defects by accident, is a death spiral. TFT versus TFT recovers in one round.\n\n**Clarity** (be simple enough to be understood) was the property Axelrod found most surprising. The losing strategies were often the cleverest — they had complex internal logic that other strategies could not model. The result was that opponents *could not learn* how to cooperate with them. The complex strategies, even when they performed reasonably against any individual opponent, lost the long-run tournament because they failed to *teach* their opponents to cooperate. TFT, by contrast, is so simple that any opponent can model it in one round: \"if I cooperate, it cooperates; if I defect, it defects.\" This makes cooperative coordination trivially learnable.\n\nThe episode teaches several things that running a repeated-game analysis correctly requires you to internalize.\n\n**First**, simplicity dominated complexity. The four-line program beat sophisticated multi-state automata. The mechanism: in a world of strategic agents, **legibility is a strategic asset**. If counterparties cannot model your strategy, they cannot learn how to cooperate with you. This generalizes far beyond Prisoner's Dilemma — clear pricing, clear reputation systems, clear contract terms outperform clever ones precisely because they are easy to coordinate around.\n\n**Second**, the tournament confirmed empirically what the Folk Theorem proved mathematically: in a repeated game with sufficient shadow of the future, cooperation is not a moral preference, it is a winning strategy. Nice strategies dominated unconditional defectors in long enough games. The strategy you should pick is not whichever is \"fairest\" or \"most ethical\" by intuition; it is whichever wins the tournament you're actually in.\n\n**Third**, the results are sensitive to noise. Axelrod's tournaments used clean observation — each program saw the previous move unambiguously. In follow-on research (Nowak & Sigmund 1992; Wu & Axelrod 1995), simulated noise was added: occasional misperception of cooperation as defection and vice versa. Under noise, pure TFT collapses into mutual recrimination, and *generous* TFT (cooperate with some probability even after observing defection) wins instead. **This is the practical lesson for any reputation system you design: assume some signal noise, and build in forgiveness — or you will create death spirals out of bookkeeping errors.**\n\n**Fourth**, the second tournament's design was crucial. The first tournament's result might have been luck. The second tournament was run with the strategies and outcomes of the first tournament fully published, with the express invitation to design beat-TFT strategies. Sixty-two entries tried. TFT won again. **This is what robust empirical evidence in strategy looks like**: the result survives adversarial follow-on examination by experts who had every motivation to falsify it.\n\n**Fifth**, the most important methodological point: **Axelrod's tournament is the founding empirical proof that mathematical game theory's pessimistic predictions are environment-conditional.** Classical theory said \"rational players in PD defect.\" Axelrod's tournament showed: in the repeated case, \"rational players cooperate, retaliate when needed, forgive, and stay legible — and that strategy beats every other strategy submitted by experts who knew the rules in advance.\" When you analyze a repeated game in your business or strategy work, you are not choosing between \"the cynical view\" and \"the optimistic view.\" You are choosing between **understanding the discount factor and observation structure** — in which case cooperation can be the dominant strategy — or **misdiagnosing the structure** — in which case you'll either defect prematurely or trust naively. The skill is not optimism; it is precision.\n\nFile v1.0.3:skill-card.md\n\n## Description: <br>\nHelps agents analyze repeated relationships and reputation systems by checking discount factors, observability, strategy fit, and reputation infrastructure. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nEmployees, external users, and developers use this skill to evaluate whether cooperation is sustainable in repeated or reputation-mediated relationships and to choose fitting strategies such as Tit for Tat, Generous Tit for Tat, Pavlov, or reputation-system design changes. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: Strategic recommendations may be overapplied to real business, employment, partnership, or marketplace decisions. <br>\nMitigation: Review the analysis before acting and confirm that the relationship is genuinely repeated or reputation-mediated. <br>\nRisk: The skill can produce misleading guidance if discount factor, observation noise, endgame, or reputation infrastructure assumptions are wrong. <br>\nMitigation: Use the skill's verification checks to validate the discount-factor threshold, observability, forgiveness design, and all required reputation-system components. <br>\n\n\n## Reference(s): <br>\n- [ClawHub Skill Page](https://clawhub.ai/deciqai/skills/repeated-games-reputation) <br>\n- [Primary Sources](references/sources.md) <br>\n- [Robert Axelrod's Computer Tournament, 1979-1981](examples/robert-axelrod-computer-tournament-1979-1981.md) <br>\n- [Axelrod and Hamilton (1981), The Evolution of Cooperation](https://doi.org/10.1126/science.7466396) <br>\n- [Friedman (1971), A Non-cooperative Equilibrium for Supergames](https://doi.org/10.2307/2296617) <br>\n- [Fudenberg and Maskin (1986), The Folk Theorem in Repeated Games](https://doi.org/10.2307/1911307) <br>\n- [Nowak and Sigmund (1992), Tit for Tat in Heterogeneous Populations](https://doi.org/10.1038/355250a0) <br>\n- [Nowak and Sigmund (1993), Win-Stay, Lose-Shift](https://doi.org/10.1038/364056a0) <br>\n- [Kreps, Milgrom, Roberts, and Wilson (1982), Rational Cooperation in the Finitely Repeated Prisoners' Dilemma](https://doi.org/10.1016/0022-0531(82)90029-1) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [Text, Markdown, Guidance, Analysis] <br>\n**Output Format:** [Markdown structured analysis] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [Produces decision-support guidance and does not require shell commands, code execution, API calls, or system changes.] <br>\n\n## Skill Version(s): <br>\n1.0.3 (source: server release evidence) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.2: 5 files, 12038 bytes\n\nFiles: examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (2209b), skill-card.md (2900b), SKILL.md (10595b), _meta.json (144b)\n\nFile v1.0.2:SKILL.md\n\n---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry.\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with `prisoners-dilemma` · `second-order-thinking` · `signaling-games`.\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use `prisoners-dilemma`\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough and your opponent will see your move.\" This skill works out the math of *enough* and *will see*.\n2. Check fit against When to Use / When NOT to use. If genuinely one-shot, redirect to `prisoners-dilemma`.\n3. Elicit their real repeated relationship or reputational concern — never run on a hypothetical when a real one is available.\n> **[WAIT — do not advance until user responds]**\n4. Walk the Analysis one element at a time: discount factor, observation structure, retaliation feasibility, forgiveness design.\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the one strategic move (TFT variant, reputational signal, structural change to δ or observation) that fits their situation.\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **Repeated-Game Analysis** across these steps:\n\n1. **Establish true repetition.** Indefinite/infinite → folk-theorem logic applies. Finite with known endpoint → backward induction risk; add uncertainty or commitment devices.\n2. **Estimate δ.** Cooperation threshold: δ ≥ (T − R) / (T − P). Below threshold → change the structure first; no strategy design saves it.\n3. **Confirm observability.** Perfect → TFT variants work. Noisy → Generous TFT (cooperate ~1/3 of the time after apparent defection) or Contrite TFT. Pure TFT under noise → recrimination spirals.\n4. **Select strategy.** TFT (clean bilateral) · Generous TFT (noisy) · Pavlov (mixed populations) · Grim Trigger (high-stakes, credible threat only) · benchmarks: Always Defect / Always Cooperate.\n5. **For reputation infrastructure:** design Observation · Aggregation · Persistence · Manipulation resistance — all four required; missing one breaks the system.\n6. **Stress-test endgame.** Mitigations: endpoint uncertainty; legacy concerns; successor obligations; overlapping generations.\n7. **Stop-rule:** lifetime cooperative payoff must beat one-shot defection by margin sufficient to absorb noise.\n\n### Output template\n\n```\nRepeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>\n```\n\n*→ Method in Action: [Robert Axelrod's Computer Tournament, 1979–1981](examples/robert-axelrod-computer-tournament-1979-1981.md)*\n\n## Pack: Reputation Infrastructure Patterns\n\nSix documented patterns (observation mechanism → known failure mode):\neBay/Airbnb (post-transaction ratings → 5-star inflation) · FICO (payment history → thin-file bias) · GitHub (commit history → popularity ≠ quality) · B2B scorecards (procurement records → approved-list lock-in) · Professional reputation (peer review + regulatory filings → old-boys' network slow to update) · Sovereign credit (macro indicators → rating-agency capture)\n\n## Applying It Well\n\nDiagnose δ first — wrong discount factor invalidates everything downstream. In noisy environments (most real ones), add forgiveness. Never confuse bilateral repetition (TFT) with third-party reputation markets (requires infrastructure). Legibility is a strategic asset: a simple strategy your counterparty can model beats a clever one they cannot.\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"We have a long relationship, so they won't defect\" | The relationship's *length* doesn't matter; the **shadow of the future** does. If their discount factor is low (they're about to retire, the firm is being sold, they have alternative partners lined up), past relationship duration provides no protection. Run the δ check, not the nostalgia check. |\n| [D] \"Reputation will discipline them\" | Only if the reputation system has all four pieces (observation, aggregation, persistence, manipulation resistance) and is actually consulted. Many \"reputation matters\" claims are wishful — the system is broken on one of the four pieces. |\n| [D] Applying TFT in a noisy environment without forgiveness | Pure TFT under noise enters mutual-recrimination death spirals. If observation has error, use Generous TFT, Contrite TFT, or Pavlov. Recommending TFT without checking noise level is a documented failure. |\n| [D] Using cooperation-by-default in a known finite-endpoint game | Backward induction: last round → defect dominates; second-to-last → both know this and defect; unraveling cascades to round one. Add uncertainty about endpoint or commitment devices. |\n| [D] \"Always-defect can't beat TFT, so cooperation is automatic\" | Always-defect can't beat TFT head-to-head, but can dominate in a population without retaliators. Tournament context matters — don't generalize from two-player simulation to a marketplace with unknown counterparties. |\n| [D] Designing a reputation system without manipulation resistance | A gameable system creates worse outcomes than no system, because the gamed signal substitutes for direct due diligence. Test every reputation system against adversarial gaming before deployment. |\n| [D] Confusing repeated game with reputation game | Repeated: same parties, bilateral, direct observation. Reputation: changing parties, third-party observation, requires infrastructure. Many \"reputation will solve it\" arguments fail because that infrastructure doesn't exist. |\n| [D] \"TFT is the winning strategy, period\" | TFT won clean-observation tournaments. Under noise, Generous TFT and Pavlov outperform it. Strategy is environment-conditional: δ, noise, population mix, and horizon all matter. |\n| [D] Adding excessive forgiveness \"to be nice\" | Over-forgiving strategies lose to exploiters. All four properties required: nice + retaliatory + forgiving + clear. Niceness without retaliation is exploited; the data is clear. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- \"Build trust\" recommended without specifying which mechanism creates it\n- δ check skipped (no estimate vs (T−R)/(T−P) threshold)\n- TFT recommended in noisy observation environment\n- Backward-induction risk ignored in finite-endpoint game\n- Reputation system proposed missing any of the four mechanisms\n- Strategy chosen for fairness rather than what wins the actual tournament\n- Bilateral repetition confused with third-party reputation market\n\n## Verification\n\n- [ ] Repetition structure diagnosed (indefinite/finite-known/finite-unknown; bilateral/reputational)\n- [ ] δ estimated and compared to (T−R)/(T−P) threshold\n- [ ] Observation noise assessed; forgiveness added if non-trivial\n- [ ] Strategy matched to specific environment (not off-the-shelf)\n- [ ] If reputation-based, all four infrastructure components specified\n- [ ] Endgame dynamics considered; mitigations named if known endpoint\n- [ ] Lifetime payoff beats one-shot defection by noise-absorbing margin\n- [ ] Alternative strategies compared and rejected with reasons\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n---\n\n*Part of **deciqAI Knowledge Skills** — 163 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. **See it run → https://www.deciqai.com/skills/repeated-games-reputation?utm_source=clawhub&utm_medium=marketplace&utm_campaign=knowledge-skills&utm_content=repeated-games-reputation** · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.*\n\nFile v1.0.2:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.2\",\n  \"publishedAt\": 1783472506125\n}\n\nFile v1.0.2:references/sources.md\n\n# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\nFile v1.0.2:examples/robert-axelrod-computer-tournament-1979-1981.md\n\n# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accounts for Tit for Tat's robust success is its combination of being nice, retaliatory, forgiving, and clear. Its niceness prevents it from getting into unnecessary trouble. Its retaliation discourages the other side from persisting whenever defection is tried. Its forgiveness helps restore mutual cooperation. And its clarity makes it intelligible to the other player, thereby eliciting long-term cooperation.\"\n\n— Axelrod, *The Evolution of Cooperation* (1984), p. 54.\n\nEach of these properties does specific work, and missing any of them costs you the tournament.\n\n**Niceness** (never first to defect) means you accumulate the (R,R) payoff with other nice strategies. The winners and most of the top-half finishers were nice; the bottom-half finishers were dominated by strategies that defected first. Axelrod's quantitative finding:\n\n> \"The single most important property of the rules that did well is what I called being nice... All of the top eight rules in the second tournament were nice, and none of the bottom seven was.\"\n\n— Axelrod (1984), p. 33.\n\nThe cost of starting with defection: against any retaliating opponent, you forfeit the long-run cooperative payoff stream for a one-time temptation gain. The math is brutal — across 200 rounds, the difference between 200R and 200P (using the standard payoffs, 600 vs 200) dwarfs any number of temptations.\n\n**Retaliation** (respond to defection) is what prevents exploitation by always-defect strategies. Strategies that were \"too nice\" — Tit-for-Two-Tats, Always Cooperate, and several over-forgiving variants — performed worse than TFT in noisy or mixed populations because always-defectors could exploit them.\n\n**Forgiveness** (return to cooperation when opponent does) is what distinguishes TFT from Grim Trigger. Both retaliate; only TFT recovers. In Axelrod's analysis, the cost of failing to forgive is borne in interactions with strategies that defect *occasionally* (whether by intention, accident, or in response to perceived defection). Grim Trigger versus Grim Trigger, once anyone defects by accident, is a death spiral. TFT versus TFT recovers in one round.\n\n**Clarity** (be simple enough to be understood) was the property Axelrod found most surprising. The losing strategies were often the cleverest — they had complex internal logic that other strategies could not model. The result was that opponents *could not learn* how to cooperate with them. The complex strategies, even when they performed reasonably against any individual opponent, lost the long-run tournament because they failed to *teach* their opponents to cooperate. TFT, by contrast, is so simple that any opponent can model it in one round: \"if I cooperate, it cooperates; if I defect, it defects.\" This makes cooperative coordination trivially learnable.\n\nThe episode teaches several things that running a repeated-game analysis correctly requires you to internalize.\n\n**First**, simplicity dominated complexity. The four-line program beat sophisticated multi-state automata. The mechanism: in a world of strategic agents, **legibility is a strategic asset**. If counterparties cannot model your strategy, they cannot learn how to cooperate with you. This generalizes far beyond Prisoner's Dilemma — clear pricing, clear reputation systems, clear contract terms outperform clever ones precisely because they are easy to coordinate around.\n\n**Second**, the tournament confirmed empirically what the Folk Theorem proved mathematically: in a repeated game with sufficient shadow of the future, cooperation is not a moral preference, it is a winning strategy. Nice strategies dominated unconditional defectors in long enough games. The strategy you should pick is not whichever is \"fairest\" or \"most ethical\" by intuition; it is whichever wins the tournament you're actually in.\n\n**Third**, the results are sensitive to noise. Axelrod's tournaments used clean observation — each program saw the previous move unambiguously. In follow-on research (Nowak & Sigmund 1992; Wu & Axelrod 1995), simulated noise was added: occasional misperception of cooperation as defection and vice versa. Under noise, pure TFT collapses into mutual recrimination, and *generous* TFT (cooperate with some probability even after observing defection) wins instead. **This is the practical lesson for any reputation system you design: assume some signal noise, and build in forgiveness — or you will create death spirals out of bookkeeping errors.**\n\n**Fourth**, the second tournament's design was crucial. The first tournament's result might have been luck. The second tournament was run with the strategies and outcomes of the first tournament fully published, with the express invitation to design beat-TFT strategies. Sixty-two entries tried. TFT won again. **This is what robust empirical evidence in strategy looks like**: the result survives adversarial follow-on examination by experts who had every motivation to falsify it.\n\n**Fifth**, the most important methodological point: **Axelrod's tournament is the founding empirical proof that mathematical game theory's pessimistic predictions are environment-conditional.** Classical theory said \"rational players in PD defect.\" Axelrod's tournament showed: in the repeated case, \"rational players cooperate, retaliate when needed, forgive, and stay legible — and that strategy beats every other strategy submitted by experts who knew the rules in advance.\" When you analyze a repeated game in your business or strategy work, you are not choosing between \"the cynical view\" and \"the optimistic view.\" You are choosing between **understanding the discount factor and observation structure** — in which case cooperation can be the dominant strategy — or **misdiagnosing the structure** — in which case you'll either defect prematurely or trust naively. The skill is not optimism; it is precision.\n\nFile v1.0.2:skill-card.md\n\n## Description: <br>\nHelps agents analyze repeated relationships and reputation systems by checking whether future cooperation is sustainable, selecting an appropriate repeated-game strategy, and identifying reputation infrastructure needs. <br>\n\nThis skill is ready for commercial/non-commercial use. <br>\n\n## Publisher: <br>\n[deciqai](https://clawhub.ai/user/deciqai) <br>\n\n### License/Terms of Use: <br>\nMIT-0 <br>\n\n\n## Use Case: <br>\nEmployees, external users, developers, and strategy agents use this skill to reason about repeat counterparties, retaliation, trust repair, and marketplace reputation design. It is intended for strategic analysis and coaching, not as professional legal, financial, or management advice. <br>\n\n### Deployment Geography for Use: <br>\nGlobal <br>\n\n## Known Risks and Mitigations: <br>\nRisk: Strategic recommendations could be mistaken for professional legal, financial, employment, or management advice. <br>\nMitigation: Use the analysis as decision support and involve qualified reviewers before acting on contracts, employment issues, financial exposure, or retaliation decisions. <br>\nRisk: A repeated-game diagnosis can be wrong if the discount factor, horizon, observation quality, or reputation infrastructure is misread. <br>\nMitigation: Follow the skill's verification checklist, explicitly test assumptions, and reassess when incentives, counterparties, or observability change. <br>\n\n\n## Reference(s): <br>\n- [Primary Sources](references/sources.md) <br>\n- [Robert Axelrod's Computer Tournament, 1979-1981](examples/robert-axelrod-computer-tournament-1979-1981.md) <br>\n- [Axelrod and Hamilton, The Evolution of Cooperation](https://doi.org/10.1126/science.7466396) <br>\n- [Friedman, A Non-cooperative Equilibrium for Supergames](https://doi.org/10.2307/2296617) <br>\n- [Fudenberg and Maskin, The Folk Theorem in Repeated Games](https://doi.org/10.2307/1911307) <br>\n- [Nowak and Sigmund, Tit for Tat in Heterogeneous Populations](https://doi.org/10.1038/355250a0) <br>\n- [Nowak and Sigmund, Win-Stay Lose-Shift](https://doi.org/10.1038/364056a0) <br>\n- [Kreps et al., Rational Cooperation in the Finitely Repeated Prisoners' Dilemma](https://doi.org/10.1016/0022-0531(82)90029-1) <br>\n\n\n## Skill Output: <br>\n**Output Type(s):** [text, markdown, guidance] <br>\n**Output Format:** [Markdown strategic analysis with a repeated-game diagnosis template] <br>\n**Output Parameters:** [1D] <br>\n**Other Properties Related to Output:** [Produces advisory reasoning and recommendations; no executable output.] <br>\n\n## Skill Version(s): <br>\n1.0.2 (source: server release metadata) <br>\n\n## Ethical Considerations: <br>\nUsers should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment. <br>\n\nArchive v1.0.1: 5 files, 11901 bytes\n\nFiles: examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (2209b), skill-card.md (2844b), SKILL.md (10484b), _meta.json (144b)\n\nFile v1.0.1:SKILL.md\n\n---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry.\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with [`prisoners-dilemma`](../prisoners-dilemma/SKILL.md) · [`second-order-thinking`](../second-order-thinking/SKILL.md) · [`signaling-games`](../signaling-games/SKILL.md).\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use [`prisoners-dilemma`](../prisoners-dilemma/SKILL.md)\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough and your opponent will see your move.\" This skill works out the math of *enough* and *will see*.\n2. Check fit against When to Use / When NOT to use. If genuinely one-shot, redirect to [`prisoners-dilemma`](../prisoners-dilemma/SKILL.md).\n3. Elicit their real repeated relationship or reputational concern — never run on a hypothetical when a real one is available.\n> **[WAIT — do not advance until user responds]**\n4. Walk the Analysis one element at a time: discount factor, observation structure, retaliation feasibility, forgiveness design.\n> **[WAIT — do not advance until user responds]**\n5. Close by naming the one strategic move (TFT variant, reputational signal, structural change to δ or observation) that fits their situation.\n> **[WAIT — do not advance until user responds]**\n\n## The Process\n\nRun the **Repeated-Game Analysis** across these steps:\n\n1. **Establish true repetition.** Indefinite/infinite → folk-theorem logic applies. Finite with known endpoint → backward induction risk; add uncertainty or commitment devices.\n2. **Estimate δ.** Cooperation threshold: δ ≥ (T − R) / (T − P). Below threshold → change the structure first; no strategy design saves it.\n3. **Confirm observability.** Perfect → TFT variants work. Noisy → Generous TFT (cooperate ~1/3 of the time after apparent defection) or Contrite TFT. Pure TFT under noise → recrimination spirals.\n4. **Select strategy.** TFT (clean bilateral) · Generous TFT (noisy) · Pavlov (mixed populations) · Grim Trigger (high-stakes, credible threat only) · benchmarks: Always Defect / Always Cooperate.\n5. **For reputation infrastructure:** design Observation · Aggregation · Persistence · Manipulation resistance — all four required; missing one breaks the system.\n6. **Stress-test endgame.** Mitigations: endpoint uncertainty; legacy concerns; successor obligations; overlapping generations.\n7. **Stop-rule:** lifetime cooperative payoff must beat one-shot defection by margin sufficient to absorb noise.\n\n### Output template\n\n```\nRepeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>\n```\n\n*→ Method in Action: [Robert Axelrod's Computer Tournament, 1979–1981](examples/robert-axelrod-computer-tournament-1979-1981.md)*\n\n## Pack: Reputation Infrastructure Patterns\n\nSix documented patterns (observation mechanism → known failure mode):\neBay/Airbnb (post-transaction ratings → 5-star inflation) · FICO (payment history → thin-file bias) · GitHub (commit history → popularity ≠ quality) · B2B scorecards (procurement records → approved-list lock-in) · Professional reputation (peer review + regulatory filings → old-boys' network slow to update) · Sovereign credit (macro indicators → rating-agency capture)\n\n## Applying It Well\n\nDiagnose δ first — wrong discount factor invalidates everything downstream. In noisy environments (most real ones), add forgiveness. Never confuse bilateral repetition (TFT) with third-party reputation markets (requires infrastructure). Legibility is a strategic asset: a simple strategy your counterparty can model beats a clever one they cannot.\n\n## Common Rationalizations\n\n**[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.**\n\n| Fake move | Reality |\n|---|---|\n| [D] \"We have a long relationship, so they won't defect\" | The relationship's *length* doesn't matter; the **shadow of the future** does. If their discount factor is low (they're about to retire, the firm is being sold, they have alternative partners lined up), past relationship duration provides no protection. Run the δ check, not the nostalgia check. |\n| [D] \"Reputation will discipline them\" | Only if the reputation system has all four pieces (observation, aggregation, persistence, manipulation resistance) and is actually consulted. Many \"reputation matters\" claims are wishful — the system is broken on one of the four pieces. |\n| [D] Applying TFT in a noisy environment without forgiveness | Pure TFT under noise enters mutual-recrimination death spirals. If observation has error, use Generous TFT, Contrite TFT, or Pavlov. Recommending TFT without checking noise level is a documented failure. |\n| [D] Using cooperation-by-default in a known finite-endpoint game | Backward induction: last round → defect dominates; second-to-last → both know this and defect; unraveling cascades to round one. Add uncertainty about endpoint or commitment devices. |\n| [D] \"Always-defect can't beat TFT, so cooperation is automatic\" | Always-defect can't beat TFT head-to-head, but can dominate in a population without retaliators. Tournament context matters — don't generalize from two-player simulation to a marketplace with unknown counterparties. |\n| [D] Designing a reputation system without manipulation resistance | A gameable system creates worse outcomes than no system, because the gamed signal substitutes for direct due diligence. Test every reputation system against adversarial gaming before deployment. |\n| [D] Confusing repeated game with reputation game | Repeated: same parties, bilateral, direct observation. Reputation: changing parties, third-party observation, requires infrastructure. Many \"reputation will solve it\" arguments fail because that infrastructure doesn't exist. |\n| [D] \"TFT is the winning strategy, period\" | TFT won clean-observation tournaments. Under noise, Generous TFT and Pavlov outperform it. Strategy is environment-conditional: δ, noise, population mix, and horizon all matter. |\n| [D] Adding excessive forgiveness \"to be nice\" | Over-forgiving strategies lose to exploiters. All four properties required: nice + retaliatory + forgiving + clear. Niceness without retaliation is exploited; the data is clear. |\n| *→ Add [O] entries here after each real use — paste the actual failure pattern* | *What went wrong and why* |\n\n## Red Flags\n\n- \"Build trust\" recommended without specifying which mechanism creates it\n- δ check skipped (no estimate vs (T−R)/(T−P) threshold)\n- TFT recommended in noisy observation environment\n- Backward-induction risk ignored in finite-endpoint game\n- Reputation system proposed missing any of the four mechanisms\n- Strategy chosen for fairness rather than what wins the actual tournament\n- Bilateral repetition confused with third-party reputation market\n\n## Verification\n\n- [ ] Repetition structure diagnosed (indefinite/finite-known/finite-unknown; bilateral/reputational)\n- [ ] δ estimated and compared to (T−R)/(T−P) threshold\n- [ ] Observation noise assessed; forgiveness added if non-trivial\n- [ ] Strategy matched to specific environment (not off-the-shelf)\n- [ ] If reputation-based, all four infrastructure components specified\n- [ ] Endgame dynamics considered; mitigations named if known endpoint\n- [ ] Lifetime payoff beats one-shot defection by noise-absorbing margin\n- [ ] Alternative strategies compared and rejected with reasons\n\n*→ Primary sources: [references/sources.md](references/sources.md)*\n\n---\n\n*Part of **deciqAI Knowledge Skills** — open-source thinking skills that make rigor executable for AI agents. Built by deciqAI · https://deciqai.com · Contributions welcome — see the template at the repo root.*\n\nFile v1.0.1:_meta.json\n\n{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.1\",\n  \"publishedAt\": 1783463550306\n}\n\nFile v1.0.1:references/sources.md\n\n# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\nFile v1.0.1:examples/robert-axelrod-computer-tournament-1979-1981.md\n\n# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accounts for Tit for Tat's robust success is its combination of being nice, retaliatory, forgiving, and clear. Its niceness prevents it from getting into unnecessary trouble. Its retaliation discourages the other side from persisting whenever defection is tried. Its forgiveness helps restore mutual cooperation. And its clarity makes it intelligible to the other player, thereby eliciting long-term cooperation.\"\n\n— Axelrod, *The Evolution of Cooperation* (1984), p. 54.\n\nEach of these properties does specific work, and missing any of them costs you the tournament.\n\n**Niceness** (never first to defect) means you accumulate the (R,R) payoff with other nice strategies. The winners and most of the top-half finishers were nice; the bottom-half finishers were dominated by strategies that defected first. Axelrod's quantitative finding:\n\n> \"The single most important property of the rules that did well is what I called being nice... All of the top eight rules in the second tournament were nice, and none of the bottom seven was.\"\n\n— Axelrod (1984), p. 33.\n\nThe cost of starting with defection: against any retaliating opponent, you forfeit the long-run cooperative payoff stream for a one-time temptation gain. The math is brutal — across 200 rounds, the difference between 200R and 200P (using the standard payoffs, 600 vs 200) dwarfs any number of temptations.\n\n**Retaliation** (respond to defection) is what prevents exploitation by always-defect strategies. Strategies that were \"too nice\" — Tit-for-Two-Tats, Always Cooperate, and several over-forgiving variants — performed worse than TFT in noisy or mixed populations because always-defectors could exploit them.\n\n**Forgiveness** (return to cooperation when opponent does) is what distinguishes TFT from Grim Trigger. Both retaliate; only TFT recovers. In Axelrod's analysis, the cost of failing to forgive is borne in interactions with strategies that defect *occasionally* (whether by intention, accident, or in response to perceived defection). Grim Trigger versus Grim Trigger, once anyone defects by accident, is a death spiral. TFT versus TFT recovers in one round.\n\n**Clarity** (be simple enough to be understood) was the property Axelrod found most surprising. The losing strategies were often the cleverest — they had complex internal logic that other strategies could not model. The result was that opponents *could not learn* how to cooperate with them. The complex strategies, even when they performed reasonably against any individual opponent, lost the long-run tournament because they failed to *teach* their opponents to cooperate. TFT, by contrast, is so simple that any opponent can model it in one round: \"if I cooperate, it cooperates; if I defect, it defects.\" This makes cooperative coordination tr\n\nArchive v1.0.0: 5 files, 11855 bytes\n\nFiles: examples/robert-axelrod-computer-tournament-1979-1981.md (8909b), references/sources.md (2209b), skill-card.md (2705b), SKILL.md (10484b), _meta.json (144b)","readmeExcerpt":"Skill: Repeated Games & Reputation Owner: deciqai Summary: Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:13:41.749Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/repeated-games-reputation.json) v1.0.4 | 2026-","codeSnippets":[],"executableExamples":[{"language":"text","snippet":"Repeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>"},{"language":"text","snippet":"Repeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>"},{"language":"text","snippet":"Repeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>"},{"language":"text","snippet":"Repeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>"},{"language":"text","snippet":"Repeated-Game Analysis: <situation>\nRepetition: <bilateral/reputational> | Horizon: <indefinite/finite-known/finite-unknown>\nδ_actual vs δ_required=(T−R)/(T−P): <values> → Cooperation sustainable? <yes/no/borderline>\nObservation: <clean/noisy> | Aggregation/Persistence/Manipulation-resistance (if reputational)\nStrategy: <TFT/Generous TFT/Pavlov/Grim Trigger> | First move: <> | Retaliation: <> | Forgiveness: <>\nEndgame risk: <> | Mitigations: <>\nTest: <lifetime payoff vs one-shot defect>"}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"---\nname: repeated-games-reputation\ndescription: \"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whether a long-term relationship can survive betrayal, or says 'shadow of the future / tit-for-tat / burn this bridge / they'll remember this / build credibility.' Do NOT activate when: interaction is genuinely one-shot with no third-party observers (use prisoners-dilemma), or situation is zero-sum competition where repetition entrenches rivalry. More: deciqai.com/c/repeated-games-reputation\"\n---\n\n# Repeated Games & Reputation\n\n## Overview\n\nWhen parties repeat — or third parties observe — defection costs tomorrow's cooperation, flipping the Prisoner's Dilemma. Axelrod's 1979–1981 tournaments proved cooperation wins empirically; the Folk Theorem (Fudenberg & Maskin 1986) proved it mathematically. This skill diagnoses when cooperation is sustainable (discount factor check), selects the right strategy (TFT vs Generous TFT vs Pavlov), and engineers reputation infrastructure for markets where parties don't repeat directly. Composes with `prisoners-dilemma` · `second-order-thinking` · `signaling-games`.\n\n## When to Use\n\nApply when:\n- A relationship is **expected to continue** between the same parties (supplier-buyer, employer-employee, GP-LP, founder-investor, customer-platform)\n- Even in a one-shot direct interaction, **third parties observe** the move and adjust their willingness to play with you\n- You're designing **a platform or marketplace** that needs strangers to cooperate — reputation infrastructure is the architectural question\n- You're trying to **escape a defection trap** and the candidate escape is \"repetition\" or \"reputation\"\n- A partnership keeps fragmenting — diagnose whether δ is too low or observation is broken\n- Trust/safety reputation is shaping who wins **AI adoption and AI-native competition** — where capability converges, a bad launch or safety incident reprices every future round of enterprise adoption (and the AI capex supercycle only lengthens the shadow of the future)\n\n**When NOT to use:**\n- Genuinely one-shot with no third-party observability → use `prisoners-dilemma`\n- Parties are **about to exit** (last round of finite game) — backward induction risk; standard repeated-game logic can fail\n- Zero-sum situation — repetition can entrench rivalry rather than dissolve it\n- \"Repetition\" is only nominal — rotating counterparties who don't talk = effectively one-shot\n\n## Coaching Novices (Adaptive Front Door)\n\n- **Engine mode:** user has a concrete repeated/reputational situation → run The Process directly.\n- **Coach mode:** user is unfamiliar or has no concrete case → guide step by step.\n\nIn Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.\n\n1. One-line what-it-is: one-shot rationality says \"defect\"; repeated rationality says \"cooperate if the future matters enough an"},{"path":"_meta.json","content":"{\n  \"ownerId\": \"kn754b8sk22s8c6gjxt02bftbn88q7ye\",\n  \"slug\": \"repeated-games-reputation\",\n  \"version\": \"1.0.5\",\n  \"publishedAt\": 1784225621749\n}"},{"path":"references/sources.md","content":"# Sources — repeated-games-reputation\n\n> *Primary sources for the [repeated-games-reputation](../SKILL.md) skill.*\n\n- Axelrod, R. (1984). *The Evolution of Cooperation*. Basic Books. The primary text reporting both computer tournaments (1979–80 and 1980–81), Rapoport's TFT victories, and Axelrod's analysis extracting the four properties (nice, retaliatory, forgiving, clear). ISBN 978-0465021215. Updated edition with additional chapters: Axelrod, R. (2006). *The Evolution of Cooperation*, Revised Edition. Basic Books.\n- Axelrod, R., & Hamilton, W. D. (1981). \"The Evolution of Cooperation.\" *Science*, 211(4489), pp. 1390–1396. The original peer-reviewed publication of the tournament findings, co-authored with evolutionary biologist W. D. Hamilton. https://doi.org/10.1126/science.7466396\n- Friedman, J. W. (1971). \"A Non-cooperative Equilibrium for Supergames.\" *Review of Economic Studies*, 38(1), pp. 1–12. The early formulation of the folk theorem — cooperation is sustainable in infinitely repeated games when the discount factor is high enough. https://doi.org/10.2307/2296617\n- Fudenberg, D., & Maskin, E. (1986). \"The Folk Theorem in Repeated Games with Discounting or with Incomplete Information.\" *Econometrica*, 54(3), pp. 533–554. The full statement of the folk theorem with rigorous discount-factor conditions. https://doi.org/10.2307/1911307\n- Nowak, M. A., & Sigmund, K. (1992). \"Tit for Tat in Heterogeneous Populations.\" *Nature*, 355, pp. 250–253. The discovery that Generous TFT outperforms TFT under observation noise. https://doi.org/10.1038/355250a0\n- Nowak, M. A., & Sigmund, K. (1993). \"A Strategy of Win-Stay, Lose-Shift That Outperforms Tit-for-Tat in the Prisoner's Dilemma Game.\" *Nature*, 364, pp. 56–58. Pavlov strategy formal introduction. https://doi.org/10.1038/364056a0\n- Kreps, D. M., Milgrom, P., Roberts, J., & Wilson, R. (1982). \"Rational Cooperation in the Finitely Repeated Prisoners' Dilemma.\" *Journal of Economic Theory*, 27(2), pp. 245–252. The \"gang of four\" paper showing that cooperation can be sustained in finitely-repeated games when there's incomplete information about opponent types. https://doi.org/10.1016/0022-0531(82)90029-1\n\n### Contemporary context (2024–2026 AI example)\n\n- European Union (2024). Regulation (EU) 2024/1689 (the \"AI Act\") — the EU's risk-based AI regulation, which entered into force in 2024 with obligations phasing in through 2025–2026. Establishes a durable, publicly documented disclosure/reporting regime that functions as part of the *observation* and *persistence* layers of AI reputation infrastructure. Official text: https://eur-lex.europa.eu/eli/reg/2024/1689/oj\n- Public AI-lab safety documentation (2023–2025). Model cards, system cards, usage policies, and published safety / responsible-scaling frameworks from major AI labs (e.g. OpenAI, Anthropic, Google DeepMind) — the widely adopted industry practice of disclosing model capabilities and safety evaluations. Cited here as the real-world *legibilit"},{"path":"examples/ai-trust-reputation-enterprise-adoption-2024-2026.md","content":"# Method in Action: Trust Reputation as Strategy in the 2024–2026 AI Race\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nBetween 2024 and 2026, the competition among frontier AI labs and platforms (OpenAI, Anthropic, Google DeepMind, Meta, and others) became one of the clearest live demonstrations of repeated-game logic in a market visible to everyone. Model capability converged fast: for many enterprise tasks, the leading models were close substitutes. That convergence pushed a second variable to the front of the buying decision — **trust and safety reputation**. Enterprises signing multi-year contracts, embedding a model in regulated workflows, and exposing customer data to it are not running a one-shot transaction. They are opening an indefinitely repeated relationship, and they know it. In that structure, a single bad launch — a safety incident, a data-handling failure, a reckless capability release — is not a one-time cost. It reprices every future round of the game.\n\nThis example runs the anchor case through the skill's own **Repeated-Game Analysis**. The point is not to rank the labs; it is to show *why* reputation is the strategic asset it became, and where the analysis says a defection actually bites.\n\n### 1. Establish true repetition\n\nThe relationship between an AI vendor and its enterprise customers, its regulators, and the broader developer public is **indefinitely repeated**, and it is repeated on two layers at once:\n\n- **Bilateral repetition:** a specific enterprise customer renews, expands seats, adds workloads, and re-buys as new model versions ship. Each release is a fresh round with the same counterparty.\n- **Reputational / third-party layer:** thousands of *other* buyers, regulators, and journalists observe how the vendor handled the last incident and adjust their willingness to play. This is the more powerful layer, because the audience is huge and the moves are public.\n\nThere is no known, fixed endpoint — new model generations keep arriving, so folk-theorem logic applies rather than backward-induction unraveling. **Horizon: indefinite. Repetition: both bilateral and reputational.**\n\n### 2. Estimate δ (the shadow of the future)\n\nThe discount factor here is high for the labs, and that is the whole game. AI is a capital-intensive, subscription-and-usage-revenue business: the value of a customer is overwhelmingly in the *stream* of future renewals and expansion, not the first contract. When most of a customer's lifetime value sits in future rounds, δ is high, and the folk theorem says cooperation (ship responsibly, honor commitments, don't cut safety corners for a launch) is sustainable — *because* the discounted future cooperative payoff swamps the one-time gain from a reckless \"win this quarter\" defection.\n\nUsing the cooperation threshold δ ≥ (T − R) / (T − P): the temptation T (rush a flashy but unsafe release, harvest short-term headlines and signups) is real but bounded; the reward R (a durable, renewi"},{"path":"examples/robert-axelrod-computer-tournament-1979-1981.md","content":"# Method in Action: Robert Axelrod's Computer Tournament, 1979–1981\n\n> *Example for the [repeated-games-reputation](../SKILL.md) skill.*\n\nThe empirical foundation of modern repeated-game theory was not derived. It was **observed** — under conditions of unprecedented procedural rigor for social science — in two computer tournaments run by political scientist Robert Axelrod at the University of Michigan from 1979 to 1981.\n\nThe setup was direct. Axelrod sent invitations to game theorists, economists, mathematicians, sociologists, computer scientists, and psychologists who had published on the Prisoner's Dilemma. Each submitter wrote a computer program implementing a strategy for the iterated Prisoner's Dilemma. The programs were entered into a round-robin tournament: each strategy played each other strategy (and a clone of itself, and one random-defector control) over a long series of rounds, with payoffs accumulated. The payoff matrix was the canonical PD (T=5, R=3, P=1, S=0). The rule was: highest cumulative score wins.\n\nThe **first tournament (1979–1980)** received 14 entries. The strategies ranged from extraordinarily complex (multi-state automata that attempted to model the opponent's strategy and respond optimally) to extraordinarily simple. The shortest entry was submitted by **Anatol Rapoport**, a mathematical psychologist at the University of Toronto best known for his work on conflict resolution and for an earlier book co-authored with Albert Chammah on Prisoner's Dilemma experiments. Rapoport's program was four lines of FORTRAN. Its strategy:\n\n> \"Tit for Tat starts with a cooperative choice, and thereafter does what the other player did on the previous move.\"\n\n— Axelrod, R., *The Evolution of Cooperation* (Basic Books, 1984), p. 31.\n\nTit-for-Tat won. The complex strategies — including several that attempted to detect cooperators and exploit them — finished lower. The result was striking enough that Axelrod organized a second tournament with full disclosure: he published the strategies of all 14 first-round entries along with their performance data, gave participants months to study the results, and invited fresh submissions. **The second tournament (1980–1981) received 62 entries from six countries** — from professional game theorists, including several Nobel-prize-track economists, and from amateurs who had read the first-tournament writeup and wanted to test their own ideas.\n\nMany of the second-round entries were explicitly designed to beat TFT. Some attempted to identify when they were playing TFT and exploit it; others tried elaborate detection-and-punishment schemes; one submitted strategy waited 30 moves to defect, hoping to extract a one-time gain before TFT's retaliation kicked in. Tit-for-Tat — the same four-line FORTRAN program, resubmitted by Rapoport without modification — won again.\n\nAxelrod's analysis of what made TFT win is the part of the literature most worth reading carefully. He extracted four properties:\n\n> \"What accou"}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe... Skill: Repeated Games & Reputation Owner: deciqai Summary: Activate when: user asks how to build trust with a repeat counterparty, whether to retaliate after a partner defected, how to design a reputation system, whe... Tags: latest:1.0.5 Version history: v1.0.5 | 2026-07-16T18:13:41.749Z | user Description tail link + agents machine-readable metadata line (deciqai.com/s/repeated-games-reputation.json) v1.0.4 | 2026-","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":2157,"uniquenessScore":50,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":"No screenshots, media assets, or demo links are available."},"primaryImageUrl":null,"mediaAssetCount":0,"assets":[],"demoUrl":null},"ownerResources":{"evidence":{"source":"unclaimed","verified":false,"confidence":"low","updatedAt":"2026-10-11T15:04:19.471Z","emptyReason":"This page has not been claimed by the agent owner."},"hasCustomPage":false,"customPageUpdatedAt":null,"customLinks":[],"structuredLinks":{"docsUrl":null,"demoUrl":null,"supportUrl":null,"pricingUrl":null,"statusUrl":null},"customPage":null},"relatedAgents":{"evidence":{"source":"protocol-neighbors","verified":false,"confidence":"medium","updatedAt":"2026-10-11T17:44:43.869Z","emptyReason":null},"items":[{"id":"8ebccd8e-3863-4187-8355-c3f14e1f9edf","entityType":"agent","canonicalPath":"/agent/iofficeai-aionui","slug":"iofficeai-aionui","name":"AionUi","description":"Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!","url":"https://github.com/iOfficeAI/AionUi","homepage":"https://www.aionui.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-10-09T19:11:12.944Z","createdAt":"2026-02-25T03:38:16.584Z","downloads":null},{"id":"b917f68a-ebff-438e-84f8-3f4b2494c0bc","entityType":"agent","canonicalPath":"/agent/activepieces-activepieces","slug":"activepieces-activepieces","name":"activepieces","description":"AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents","url":"https://github.com/activepieces/activepieces","homepage":"https://www.activepieces.com","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-15T02:22:12.426Z","createdAt":"2026-02-25T03:38:12.412Z","downloads":null},{"id":"5cb26759-3a39-483f-94cf-276a98c13bb8","entityType":"agent","canonicalPath":"/agent/cherryhq-cherry-studio","slug":"cherryhq-cherry-studio","name":"cherry-studio","description":"AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs","url":"https://github.com/CherryHQ/cherry-studio","homepage":"https://cherry-ai.com","source":"GITHUB_REPOS","protocols":["MCP","OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-04-11T14:38:40.986Z","createdAt":"2026-02-25T03:38:19.379Z","downloads":null},{"id":"6f6582d0-5d76-4f0f-b81d-86520247950b","entityType":"agent","canonicalPath":"/agent/copilotkit-copilotkit","slug":"copilotkit-copilotkit","name":"CopilotKit","description":"The Frontend for Agents & Generative UI. React + Angular","url":"https://github.com/CopilotKit/CopilotKit","homepage":"https://docs.copilotkit.ai","source":"GITHUB_REPOS","protocols":["OPENCLAW"],"capabilities":[],"safetyScore":100,"overallRank":70,"updatedAt":"2026-03-25T09:50:57.846Z","createdAt":"2026-02-25T03:39:14.617Z","downloads":null}],"links":{"hub":"/agent","source":"/agent/source/clawhub","protocols":[{"label":"OpenClaw","href":"/agent/protocol/openclew"}]}}}