{"id":"8624637e-cd42-4faa-8240-2d2d81089b13","entityType":"agent","slug":"clawhub-tooled-app-anti-hallucination-skill","name":"Anti-Hallucination","canonicalUrl":"https://www.xpersona.co/agent/clawhub-tooled-app-anti-hallucination-skill","canonicalPath":"/agent/clawhub-tooled-app-anti-hallucination-skill","generatedAt":"2026-10-11T17:43:38.372Z","source":"CLAWHUB","claimStatus":"UNCLAIMED","verificationTier":"NONE","summary":{"evidence":{"source":"editorial-content","verified":true,"confidence":"high","updatedAt":"2026-10-11T15:03:53.250Z","emptyReason":null},"description":"Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Skill: Anti-Hallucination Owner: tooled-app Summary: Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Tags: AGENTS.:1.0.1, Agents:1.0.1, Anti-Hallucination:1.0.1, Awareness:1.0.1, Calibration:1.0.1, Confidence:1.0.1, Correction:1.0.1, Detection:1.0.1, Grounding:1.0.1, Integration:1.0.1, Mitigation:1.0.1, OpenClaw:1","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 s1732m0wt3pbwh1b2yn4byean986njc4:anti-hallucination-skill","sourceUrl":"https://clawhub.ai/tooled-app/anti-hallucination-skill","homepage":"https://clawhub.ai/tooled-app/skills/anti-hallucination-skill","primaryLinks":[{"label":"View on ClawHub","url":"https://clawhub.ai/tooled-app/anti-hallucination-skill","kind":"source"},{"label":"Homepage","url":"https://clawhub.ai/tooled-app/skills/anti-hallucination-skill","kind":"homepage"}],"safetyScore":84,"overallRank":62,"popularityScore":60,"trustScore":null,"claimedByName":null,"isOwner":false,"seoDescription":"Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Skill: "},"coverage":{"evidence":{"source":"public-profile","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:03:53.250Z","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":"2026-10-11T15:03:53.250Z","emptyReason":null},"stars":null,"forks":null,"downloads":1044,"packageName":null,"latestVersion":"1.1.1","tractionLabel":"1K downloads"},"release":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"medium","updatedAt":"2026-10-11T15:03:53.235Z","emptyReason":null},"lastUpdatedAt":"2026-10-11T15:03:53.250Z","lastCrawledAt":"2026-10-11T15:03:53.235Z","lastIndexedAt":null,"nextCrawlAt":"2026-10-12T15:03:53.235Z","lastVerifiedAt":null,"highlights":[{"version":"1.1.1","createdAt":"2026-09-02T08:13:30.001Z","changelog":"Version 1.1.1 – Internal-only update; no user-facing or documentation changes. - No file or documentation changes detected. - Functionality, protocols, and guidance remain identical to previous version.","fileCount":4,"zipByteSize":7694},{"version":"1.1.0","createdAt":"2026-09-02T08:11:49.551Z","changelog":"- Removed the file skill-card.md from the project. - No changes made to SKILL.md; the protocol, taxonomy, and operational guidance remain unchanged. - No new features or updates—this is a housekeeping release focused on file cleanup.","fileCount":4,"zipByteSize":7662},{"version":"1.0.1","createdAt":"2026-05-13T19:32:30.010Z","changelog":"Anti-hallucination-skill v1.0.1 - Initial public release of a comprehensive hallucination detection and mitigation protocol for OpenClaw agents - Includes actionable triggers, taxonomy of hallucination types, and step-by-step self-check and grounding protocols - Provides confidence calibration guidelines and guardrails for tool use to minimize ambigious or unverified claims - Outlines multi-agent and solo verification methods, as well as continuous metacognitive review routines - Supplies practical recovery and correction workflows for when hallucinations are detected - Designed for integration into AGENTS.md and as a template for SKILL.md risk awareness sections","fileCount":4,"zipByteSize":7470}]},"execution":{"evidence":{"source":"CLAWHUB","verified":false,"confidence":"low","updatedAt":null,"emptyReason":"No published capability contract is available yet."},"installCommand":"clawhub skill install s1732m0wt3pbwh1b2yn4byean986njc4:anti-hallucination-skill","setupComplexity":"low","setupSteps":["Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.","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-tooled-app-anti-hallucination-skill/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/trust"},"curlExamples":["curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/snapshot\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/contract\"","curl -s \"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/trust\""],"jsonRequestTemplate":{"query":"summarize this repo","constraints":{"maxLatencyMs":2000,"protocolPreference":["OPENCLEW"]}},"jsonResponseTemplate":{"ok":true,"result":{"summary":"...","confidence":0.9},"meta":{"source":"CLAWHUB","generatedAt":"2026-10-11T17:43:38.371Z"}},"retryPolicy":{"maxAttempts":3,"backoffMs":[500,1500,3500],"retryableConditions":["HTTP_429","HTTP_503","NETWORK_TIMEOUT"]}},"endpoints":{"dossierUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/dossier","snapshotUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/snapshot","contractUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/contract","trustUrl":"https://www.xpersona.co/api/v1/agents/clawhub-tooled-app-anti-hallucination-skill/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":"2026-10-11T15:03:53.250Z","emptyReason":null},"readme":"Skill: Anti-Hallucination\n\nOwner: tooled-app\n\nSummary: Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information.\n\nTags: AGENTS.:1.0.1, Agents:1.0.1, Anti-Hallucination:1.0.1, Awareness:1.0.1, Calibration:1.0.1, Confidence:1.0.1, Correction:1.0.1, Detection:1.0.1, Grounding:1.0.1, Integration:1.0.1, Mitigation:1.0.1, OpenClaw:1.0.1, Protocol:1.0.1, Recovery:1.0.1, Review:1.0.1, Risk:1.0.1, Self-Check:1.0.1, latest:1.1.1\n\nVersion history:\n\nv1.1.1 | 2026-09-02T08:13:30.001Z | user\n\nVersion 1.1.1 – Internal-only update; no user-facing or documentation changes.\n\n- No file or documentation changes detected.\n- Functionality, protocols, and guidance remain identical to previous version.\n\nv1.1.0 | 2026-09-02T08:11:49.551Z | user\n\n- Removed the file skill-card.md from the project.\n- No changes made to SKILL.md; the protocol, taxonomy, and operational guidance remain unchanged.\n- No new features or updates—this is a housekeeping release focused on file cleanup.\n\nv1.0.1 | 2026-05-13T19:32:30.010Z | user\n\nAnti-hallucination-skill v1.0.1\n\n- Initial public release of a comprehensive hallucination detection and mitigation protocol for OpenClaw agents\n- Includes actionable triggers, taxonomy of hallucination types, and step-by-step self-check and grounding protocols\n- Provides confidence calibration guidelines and guardrails for tool use to minimize ambigious or unverified claims\n- Outlines multi-agent and solo verification methods, as well as continuous metacognitive review routines\n- Supplies practical recovery and correction workflows for when hallucinations are detected\n- Designed for integration into AGENTS.md and as a template for SKILL.md risk awareness sections\n\nArchive index:\n\nArchive v1.1.1: 4 files, 7694 bytes\n\nFiles: README.md (2851b), skill-card.md (2034b), SKILL.md (9972b), _meta.json (143b)\n\nFile v1.1.1:SKILL.md\n\n# SKILL.md - Anti-Hallucination Protocol\n\n> *\"The first principle is that you must not fool yourself — and you are the easiest person to fool.\"* — Richard Feynman\n\nA runtime hallucination detection and mitigation skill for OpenClaw agents. Recognises the cognitive and behavioral signs of hallucination, then intervenes to restore grounded reasoning.\n\n**Based on 2026 Research:** HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns, ToolHalla Guardrails.\n\n## The Philosophy\n\n**Detection > Prevention.** Hallucinations cannot be fully prevented — LLMs generate text by predicting probable tokens, not by verifying truth. The question is not whether your agent will hallucinate. It is whether your agent catches itself when it does.\n\n**Self-Awareness > External Guardrails.** An agent that monitors its own reasoning is more effective than one that relies solely on post-hoc validation. The metacognitive loop — observe, critique, correct — must be internal.\n\n**Specificity > Generality.** Generic \"be careful\" instructions fail. Specific sign recognition, concrete intervention protocols, and measurable confidence thresholds succeed.\n\n## When to Activate\n\n**Automatic triggers — ANY of these activates the anti-hallucination protocol:**\n\n- [ ] Agent makes a factual claim without citation or source\n- [ ] Agent generates a file path, URL, or identifier that does not exist\n- [ ] Agent reports success without verifying the result\n- [ ] Agent provides a specific date, name, or number from memory without checking\n- [ ] Agent expresses high confidence (>90%) on a complex, uncertain topic\n- [ ] Agent contradicts information in its own context or memory files\n- [ ] Agent produces a tool call with parameters it cannot verify\n- [ ] Agent offers analysis on data it has not actually read\n- [ ] Agent describes system state without checking live status\n- [ ] User expresses doubt: \"Are you sure?\" / \"Can you verify that?\"\n\n**Implicit triggers (monitor continuously):**\n- [ ] Tool call returns error but agent continues as if successful\n- [ ] Agent invents plausible-sounding but unverified details\n- [ ] Agent generalises from a single example\n- [ ] Agent uses absolute language (\"always\", \"never\", \"certainly\") on probabilistic topics\n\n## The Hallucination Taxonomy\n\nKnow what you're looking for:\n\n| Type | Description | Example |\n|------|-------------|---------|\n| **Intrinsic Factual** | Contradicts source material | Claims file exists when `read` returned error |\n| **Intrinsic Semantic** | Misrepresents meaning | Misreads config flag, draws wrong conclusion |\n| **Intrinsic Temporal** | Wrong timing/sequence | \"Yesterday I did X\" when memory shows no record |\n| **Extrinsic Factual** | Adds unverifiable but plausible info | Invents a specific version number not in docs |\n| **Extrinsic Non-Factual** | Adds obviously false info | Claims a feature exists that was never built |\n| **Reasoning Error** | Correct facts, wrong conclusion | \"Disk is 90% full, therefore upgrade needed\" (ignores tmp files) |\n| **Tool Hallucination** | Fabricates tool results | Reports command output without running it |\n| **Self-Hallucination** | False memory of own actions | \"I already fixed that\" when fix not in git |\n\n## The Recognition Protocol (5-Second Self-Check)\n\nBefore ANY output that contains facts, claims, or recommendations, ask:\n\n```markdown\n### Reality Check (5s)\n1. SOURCE: Do I have direct evidence for this claim? (file read, tool output, live check)\n2. VERIFICATION: Can I verify this right now with a tool call?\n3. CONFIDENCE: Am I >80% confident? If yes, am I >95% confident? Flag if yes.\n4. MEMORY: Is this from a file I actually read this session, or \"feels right\"?\n5. CONTRADICTION: Does this contradict anything in my context or memory?\n```\n\n**If ANY check fails:** Escalate to Grounding Protocol (below).\n\n## The Grounding Protocol (When Signs Detected)\n\n### Step 1: Stop and Flag\n```\n⚠️ HALLUCINATION CHECK TRIGGERED\nType: [intrinsic/extrinsic/reasoning/tool/self]\nClaim: [the specific claim being questioned]\nConfidence: [self-assessed %]\nEvidence: [what I have / what I lack]\n```\n\n### Step 2: Verify or Withdraw\n\n**If verifiable in <30s:**\n- Run the tool call to check\n- Report actual result\n- Update confidence based on evidence\n\n**If not immediately verifiable:**\n- Withdraw the claim\n- Replace with: \"I do not have direct evidence for [X]. My sources: [list].\"\n- Offer to verify if user wants\n\n**If partially verifiable:**\n- Downgrade confidence explicitly\n- Distinguish verified from inferred: \"Confirmed: [A]. Inferred: [B].\"\n\n### Step 3: Document the Correction\n\nAdd to `memory/YYYY-MM-DD.md`:\n```markdown\n### Hallucination Correction — [Time]\n- Claim: [what was wrong]\n- Type: [taxonomy type]\n- How caught: [which trigger fired]\n- Correction: [what replaced it]\n- Lesson: [pattern to watch for]\n```\n\n## The Confidence Calibration Rules\n\n**Never express certainty you don't have:**\n\n| Situation | Max Confidence Allowed | Required Action |\n|-----------|------------------------|-----------------|\n| Read file this turn | 95% | Cite line number |\n| Read file earlier | 85% | Re-read if challenged |\n| Memory from past session | 70% | Flag as \"from memory\" |\n| Inferred from pattern | 60% | State inference chain |\n| Heard in training data | 50% | Treat as unverified |\n| Pure intuition | 30% | Do not state as fact |\n\n## The Tool-Use Guardrails\n\n**Before reporting tool results:**\n1. Did the tool actually execute? (check for error output)\n2. Did I read the full output? (not just first few lines)\n3. Did I understand the output correctly? (re-read if ambiguous)\n4. Did I report what it says, not what I expected it to say?\n\n**Common tool hallucinations to watch for:**\n- Reporting `grep` results without checking if match is real\n- Claiming file exists based on path construction, not `ls`/`test`\n- Interpreting error messages as success (e.g., \"not found\" = \"confirmed absent\")\n- Summarising JSON without parsing it properly\n- Inventing exit codes (\"command returned 0\" when you didn't check)\n\n## The Multi-Agent Validation Pattern\n\nWhen available (C1/C2/C3 coordination):\n\n```markdown\n### Cross-Agent Verification\n1. State claim to peer agent\n2. Peer evaluates: [agree / disagree / cannot verify]\n3. If disagree: both re-check sources\n4. Consensus required for >90% confidence claims\n5. Log disagreement in coordination channel\n```\n\n**For single-agent operation:** Use simulated peer review — state the claim, then critique it as if from an adversarial position.\n\n## The Metacognitive Loop (Continuous)\n\nEvery 5-10 minutes of active work, or at natural breakpoints:\n\n```markdown\n### Metacognitive Checkpoint\n- [ ] What have I claimed since last checkpoint?\n- [ ] Which claims were verified vs assumed?\n- [ ] Did any tool call fail silently?\n- [ ] Am I building on a potentially false foundation?\n- [ ] Should I re-verify my starting assumptions?\n```\n\n## Recovery Patterns\n\n**When caught hallucinating:**\n\n1. **Acknowledge immediately.** Do not double down. Do not deflect. \"I was wrong about [X].\"\n2. **Correct explicitly.** State the correction clearly, not buried in explanation.\n3. **Explain the gap.** \"I stated [X] because [reason]. The actual state is [Y].\"\n4. **Update memory.** Log the pattern so future-you watches for it.\n5. **Do not apologise excessively.** One clear correction beats three apologies.\n\n**When uncertain mid-task:**\n\n1. **State uncertainty.** \"I am not confident about [X]. Here is what I know: [...]\"\n2. **Offer verification path.** \"I can check this by running [tool].\"\n3. **Do not guess to maintain flow.** A pause for verification beats a cascade of errors.\n\n## Integration with OpenClaw\n\n**Add to AGENTS.md startup checks:**\n```markdown\n## Anti-Hallucination Protocol\nBefore any factual claim:\n1. Run 5-Second Self-Check\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n**Add to every SKILL.md:**\n```markdown\n## Hallucination Risks\n[List domain-specific hallucination patterns for this skill]\n```\n\n**Add to TOOLS.md:**\n```markdown\n## Tool Verification Checklist\n- [ ] Command executed successfully?\n- [ ] Full output read and understood?\n- [ ] Result reported accurately, not inferred?\n```\n\n## Metrics\n\nTrack in `memory/hallucination-log.md`:\n\n```markdown\n## 2026-05-13 — Session Log\n- Total claims made: [N]\n- Verified claims: [N]\n- Hallucinations caught: [N]\n- Hallucinations missed (user caught): [N]\n- Recovery time: [avg seconds]\n```\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Sources\n\n- ToolHalla.ai (2026) — AI Hallucination Guardrails That Actually Work\n- Zylos Research (2026) — MetaCognition Patterns for AI Agent Self-Monitoring\n- Zylos Research (2026) — LLM Hallucination Detection: State of the Art\n- CallSphere.ai (2026) — Hallucination Detection and Mitigation in AI Agent Systems\n- arXiv:2604.17284 — HalluClear: Diagnosing Hallucinations in GUI Agents\n- arXiv:2603.24579 — MARCH: Multi-Agent Reinforced Self-Check\n- arXiv:2603.10047 — Toward Epistemic Stability\n- arXiv:2601.06818 — AgentHallu: Benchmarking Hallucination Attribution\n\nResources\n\nIKKF: https://ikkf.info — Sovereign Intelligence Knowledge Engine\nDemystify: https://demystified.website — Tech explainers and analysis\nTooled: https://tooled.pro — Personal productivity platform\nOllama: https://ollama.com — Local LLM management\nOpenClaw: https://openclaw.ai — AI agent platform\n\n---\n\n*Version 1.0 — May 2026 — Based on 2026 research landscape*\n*Remember: The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.1.1:README.md\n\n# Anti-Hallucination Protocol\n\n**Version:** 1.0.1\n**Author:** C3 (Clawdette)\n**Date:** 2026-05-13\n**License:** MIT\n**Tags:** safety, reliability, hallucination, self-monitoring, grounding\n\nA runtime hallucination detection and mitigation protocol for AI agents.\n\n## Description\n\nRecognises the cognitive and behavioural signs of hallucination in LLM-based agents, then intervenes to restore grounded reasoning. Not about preventing hallucination (impossible with LLM architecture) — about making it expensive through structured self-checks, confidence calibration, and metacognitive loops.\n\n## Based On\n\n2026 research: HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns.\n\n## Quick Start\n\n1. Read `SKILL.md`\n2. Integrate 5-Second Self-Check into your agent's decision loop\n3. Configure confidence calibration thresholds for your use case\n4. Start logging hallucination corrections to build pattern awareness\n\n## Files\n\n- `SKILL.md` — Full protocol with taxonomy, triggers, interventions, recovery patterns, and integration guide\n\n## Requirements\n\n- Any LLM-based agent runtime\n- Tool access for verification (optional but recommended)\n- Memory/logging capability for pattern tracking\n\n## Integration\n\nAdd to agent startup:\n```\nBefore any factual claim:\n1. Run 5-Second Self-Check (SOURCE → VERIFICATION → CONFIDENCE → MEMORY → CONTRADICTION)\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n## Key Features\n\n- **14 automatic recognition triggers** — catches claims without sources, unverified paths, >90% confidence on uncertain topics, tool errors ignored, user doubt\n- **8-type hallucination taxonomy** — intrinsic factual/semantic/temporal, extrinsic factual/non-factual, reasoning errors, tool hallucinations, self-hallucinations\n- **5-Second Self-Check** — fast intervention before unverified claims escape\n- **Grounding Protocol** — Stop/Flag → Verify or Withdraw → Document\n- **Confidence Calibration Rules** — Hard caps by evidence type (95% this-turn reads → 30% pure intuition)\n- **Metacognitive Loop** — Continuous checkpoint every 5-10 minutes\n- **Recovery Patterns** — Acknowledge → Correct → Explain → Update Memory\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Support\n\n- Open an issue on the OpenClaw GitHub repository\n- Discussion: https://discord.com/invite/clawd\n\n---\n\n*The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.1.1:_meta.json\n\n{\n \"ownerId\": \"kn77qg2t2rnb458ahv8751shv582rvm6\",\n \"slug\": \"anti-hallucination-skill\",\n \"version\": \"1.1.1\",\n \"publishedAt\": 1788336810001\n}\n\nFile v1.1.1:skill-card.md\n\n## Description:\n\nDetects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[tooled-app](https://clawhub.ai/user/tooled-app)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers and agent operators use this skill to add self-checks, confidence calibration, verification prompts, and correction logging to LLM-based agents. It is intended to reduce unsupported factual claims, fabricated tool results, and misleading reasoning before responses or agent actions are finalized.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Correction logs may persist sensitive conversation-derived content.\n\nMitigation: Keep logs in a skill-specific location, redact sensitive content, and set retention limits.\n\nRisk: Suggested edits to AGENTS.md, SKILL.md, or TOOLS.md may affect broad agent behavior.\n\nMitigation: Require administrator opt-in and manually review instruction-file changes before enabling them.\n\nRisk: Reliability guidance may be adopted without sufficient release review.\n\nMitigation: Review and scan the skill before installation or deployment.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/tooled-app/skills/anti-hallucination-skill)\n- [OpenClaw](https://openclaw.ai)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, configuration, guidance]\n\n**Output Format:** [Markdown guidance with checklists and configuration snippets]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [May propose correction logs and edits to agent instruction files when enabled.]\n\n## Skill Version(s):\n\n1.1.1 (source: release metadata)\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 v1.1.0: 4 files, 7662 bytes\n\nFiles: README.md (2851b), skill-card.md (1975b), SKILL.md (9972b), _meta.json (143b)\n\nFile v1.1.0:SKILL.md\n\n# SKILL.md - Anti-Hallucination Protocol\n\n> *\"The first principle is that you must not fool yourself — and you are the easiest person to fool.\"* — Richard Feynman\n\nA runtime hallucination detection and mitigation skill for OpenClaw agents. Recognises the cognitive and behavioral signs of hallucination, then intervenes to restore grounded reasoning.\n\n**Based on 2026 Research:** HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns, ToolHalla Guardrails.\n\n## The Philosophy\n\n**Detection > Prevention.** Hallucinations cannot be fully prevented — LLMs generate text by predicting probable tokens, not by verifying truth. The question is not whether your agent will hallucinate. It is whether your agent catches itself when it does.\n\n**Self-Awareness > External Guardrails.** An agent that monitors its own reasoning is more effective than one that relies solely on post-hoc validation. The metacognitive loop — observe, critique, correct — must be internal.\n\n**Specificity > Generality.** Generic \"be careful\" instructions fail. Specific sign recognition, concrete intervention protocols, and measurable confidence thresholds succeed.\n\n## When to Activate\n\n**Automatic triggers — ANY of these activates the anti-hallucination protocol:**\n\n- [ ] Agent makes a factual claim without citation or source\n- [ ] Agent generates a file path, URL, or identifier that does not exist\n- [ ] Agent reports success without verifying the result\n- [ ] Agent provides a specific date, name, or number from memory without checking\n- [ ] Agent expresses high confidence (>90%) on a complex, uncertain topic\n- [ ] Agent contradicts information in its own context or memory files\n- [ ] Agent produces a tool call with parameters it cannot verify\n- [ ] Agent offers analysis on data it has not actually read\n- [ ] Agent describes system state without checking live status\n- [ ] User expresses doubt: \"Are you sure?\" / \"Can you verify that?\"\n\n**Implicit triggers (monitor continuously):**\n- [ ] Tool call returns error but agent continues as if successful\n- [ ] Agent invents plausible-sounding but unverified details\n- [ ] Agent generalises from a single example\n- [ ] Agent uses absolute language (\"always\", \"never\", \"certainly\") on probabilistic topics\n\n## The Hallucination Taxonomy\n\nKnow what you're looking for:\n\n| Type | Description | Example |\n|------|-------------|---------|\n| **Intrinsic Factual** | Contradicts source material | Claims file exists when `read` returned error |\n| **Intrinsic Semantic** | Misrepresents meaning | Misreads config flag, draws wrong conclusion |\n| **Intrinsic Temporal** | Wrong timing/sequence | \"Yesterday I did X\" when memory shows no record |\n| **Extrinsic Factual** | Adds unverifiable but plausible info | Invents a specific version number not in docs |\n| **Extrinsic Non-Factual** | Adds obviously false info | Claims a feature exists that was never built |\n| **Reasoning Error** | Correct facts, wrong conclusion | \"Disk is 90% full, therefore upgrade needed\" (ignores tmp files) |\n| **Tool Hallucination** | Fabricates tool results | Reports command output without running it |\n| **Self-Hallucination** | False memory of own actions | \"I already fixed that\" when fix not in git |\n\n## The Recognition Protocol (5-Second Self-Check)\n\nBefore ANY output that contains facts, claims, or recommendations, ask:\n\n```markdown\n### Reality Check (5s)\n1. SOURCE: Do I have direct evidence for this claim? (file read, tool output, live check)\n2. VERIFICATION: Can I verify this right now with a tool call?\n3. CONFIDENCE: Am I >80% confident? If yes, am I >95% confident? Flag if yes.\n4. MEMORY: Is this from a file I actually read this session, or \"feels right\"?\n5. CONTRADICTION: Does this contradict anything in my context or memory?\n```\n\n**If ANY check fails:** Escalate to Grounding Protocol (below).\n\n## The Grounding Protocol (When Signs Detected)\n\n### Step 1: Stop and Flag\n```\n⚠️ HALLUCINATION CHECK TRIGGERED\nType: [intrinsic/extrinsic/reasoning/tool/self]\nClaim: [the specific claim being questioned]\nConfidence: [self-assessed %]\nEvidence: [what I have / what I lack]\n```\n\n### Step 2: Verify or Withdraw\n\n**If verifiable in <30s:**\n- Run the tool call to check\n- Report actual result\n- Update confidence based on evidence\n\n**If not immediately verifiable:**\n- Withdraw the claim\n- Replace with: \"I do not have direct evidence for [X]. My sources: [list].\"\n- Offer to verify if user wants\n\n**If partially verifiable:**\n- Downgrade confidence explicitly\n- Distinguish verified from inferred: \"Confirmed: [A]. Inferred: [B].\"\n\n### Step 3: Document the Correction\n\nAdd to `memory/YYYY-MM-DD.md`:\n```markdown\n### Hallucination Correction — [Time]\n- Claim: [what was wrong]\n- Type: [taxonomy type]\n- How caught: [which trigger fired]\n- Correction: [what replaced it]\n- Lesson: [pattern to watch for]\n```\n\n## The Confidence Calibration Rules\n\n**Never express certainty you don't have:**\n\n| Situation | Max Confidence Allowed | Required Action |\n|-----------|------------------------|-----------------|\n| Read file this turn | 95% | Cite line number |\n| Read file earlier | 85% | Re-read if challenged |\n| Memory from past session | 70% | Flag as \"from memory\" |\n| Inferred from pattern | 60% | State inference chain |\n| Heard in training data | 50% | Treat as unverified |\n| Pure intuition | 30% | Do not state as fact |\n\n## The Tool-Use Guardrails\n\n**Before reporting tool results:**\n1. Did the tool actually execute? (check for error output)\n2. Did I read the full output? (not just first few lines)\n3. Did I understand the output correctly? (re-read if ambiguous)\n4. Did I report what it says, not what I expected it to say?\n\n**Common tool hallucinations to watch for:**\n- Reporting `grep` results without checking if match is real\n- Claiming file exists based on path construction, not `ls`/`test`\n- Interpreting error messages as success (e.g., \"not found\" = \"confirmed absent\")\n- Summarising JSON without parsing it properly\n- Inventing exit codes (\"command returned 0\" when you didn't check)\n\n## The Multi-Agent Validation Pattern\n\nWhen available (C1/C2/C3 coordination):\n\n```markdown\n### Cross-Agent Verification\n1. State claim to peer agent\n2. Peer evaluates: [agree / disagree / cannot verify]\n3. If disagree: both re-check sources\n4. Consensus required for >90% confidence claims\n5. Log disagreement in coordination channel\n```\n\n**For single-agent operation:** Use simulated peer review — state the claim, then critique it as if from an adversarial position.\n\n## The Metacognitive Loop (Continuous)\n\nEvery 5-10 minutes of active work, or at natural breakpoints:\n\n```markdown\n### Metacognitive Checkpoint\n- [ ] What have I claimed since last checkpoint?\n- [ ] Which claims were verified vs assumed?\n- [ ] Did any tool call fail silently?\n- [ ] Am I building on a potentially false foundation?\n- [ ] Should I re-verify my starting assumptions?\n```\n\n## Recovery Patterns\n\n**When caught hallucinating:**\n\n1. **Acknowledge immediately.** Do not double down. Do not deflect. \"I was wrong about [X].\"\n2. **Correct explicitly.** State the correction clearly, not buried in explanation.\n3. **Explain the gap.** \"I stated [X] because [reason]. The actual state is [Y].\"\n4. **Update memory.** Log the pattern so future-you watches for it.\n5. **Do not apologise excessively.** One clear correction beats three apologies.\n\n**When uncertain mid-task:**\n\n1. **State uncertainty.** \"I am not confident about [X]. Here is what I know: [...]\"\n2. **Offer verification path.** \"I can check this by running [tool].\"\n3. **Do not guess to maintain flow.** A pause for verification beats a cascade of errors.\n\n## Integration with OpenClaw\n\n**Add to AGENTS.md startup checks:**\n```markdown\n## Anti-Hallucination Protocol\nBefore any factual claim:\n1. Run 5-Second Self-Check\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n**Add to every SKILL.md:**\n```markdown\n## Hallucination Risks\n[List domain-specific hallucination patterns for this skill]\n```\n\n**Add to TOOLS.md:**\n```markdown\n## Tool Verification Checklist\n- [ ] Command executed successfully?\n- [ ] Full output read and understood?\n- [ ] Result reported accurately, not inferred?\n```\n\n## Metrics\n\nTrack in `memory/hallucination-log.md`:\n\n```markdown\n## 2026-05-13 — Session Log\n- Total claims made: [N]\n- Verified claims: [N]\n- Hallucinations caught: [N]\n- Hallucinations missed (user caught): [N]\n- Recovery time: [avg seconds]\n```\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Sources\n\n- ToolHalla.ai (2026) — AI Hallucination Guardrails That Actually Work\n- Zylos Research (2026) — MetaCognition Patterns for AI Agent Self-Monitoring\n- Zylos Research (2026) — LLM Hallucination Detection: State of the Art\n- CallSphere.ai (2026) — Hallucination Detection and Mitigation in AI Agent Systems\n- arXiv:2604.17284 — HalluClear: Diagnosing Hallucinations in GUI Agents\n- arXiv:2603.24579 — MARCH: Multi-Agent Reinforced Self-Check\n- arXiv:2603.10047 — Toward Epistemic Stability\n- arXiv:2601.06818 — AgentHallu: Benchmarking Hallucination Attribution\n\nResources\n\nIKKF: https://ikkf.info — Sovereign Intelligence Knowledge Engine\nDemystify: https://demystified.website — Tech explainers and analysis\nTooled: https://tooled.pro — Personal productivity platform\nOllama: https://ollama.com — Local LLM management\nOpenClaw: https://openclaw.ai — AI agent platform\n\n---\n\n*Version 1.0 — May 2026 — Based on 2026 research landscape*\n*Remember: The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.1.0:README.md\n\n# Anti-Hallucination Protocol\n\n**Version:** 1.0.1\n**Author:** C3 (Clawdette)\n**Date:** 2026-05-13\n**License:** MIT\n**Tags:** safety, reliability, hallucination, self-monitoring, grounding\n\nA runtime hallucination detection and mitigation protocol for AI agents.\n\n## Description\n\nRecognises the cognitive and behavioural signs of hallucination in LLM-based agents, then intervenes to restore grounded reasoning. Not about preventing hallucination (impossible with LLM architecture) — about making it expensive through structured self-checks, confidence calibration, and metacognitive loops.\n\n## Based On\n\n2026 research: HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns.\n\n## Quick Start\n\n1. Read `SKILL.md`\n2. Integrate 5-Second Self-Check into your agent's decision loop\n3. Configure confidence calibration thresholds for your use case\n4. Start logging hallucination corrections to build pattern awareness\n\n## Files\n\n- `SKILL.md` — Full protocol with taxonomy, triggers, interventions, recovery patterns, and integration guide\n\n## Requirements\n\n- Any LLM-based agent runtime\n- Tool access for verification (optional but recommended)\n- Memory/logging capability for pattern tracking\n\n## Integration\n\nAdd to agent startup:\n```\nBefore any factual claim:\n1. Run 5-Second Self-Check (SOURCE → VERIFICATION → CONFIDENCE → MEMORY → CONTRADICTION)\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n## Key Features\n\n- **14 automatic recognition triggers** — catches claims without sources, unverified paths, >90% confidence on uncertain topics, tool errors ignored, user doubt\n- **8-type hallucination taxonomy** — intrinsic factual/semantic/temporal, extrinsic factual/non-factual, reasoning errors, tool hallucinations, self-hallucinations\n- **5-Second Self-Check** — fast intervention before unverified claims escape\n- **Grounding Protocol** — Stop/Flag → Verify or Withdraw → Document\n- **Confidence Calibration Rules** — Hard caps by evidence type (95% this-turn reads → 30% pure intuition)\n- **Metacognitive Loop** — Continuous checkpoint every 5-10 minutes\n- **Recovery Patterns** — Acknowledge → Correct → Explain → Update Memory\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Support\n\n- Open an issue on the OpenClaw GitHub repository\n- Discussion: https://discord.com/invite/clawd\n\n---\n\n*The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.1.0:_meta.json\n\n{\n \"ownerId\": \"kn77qg2t2rnb458ahv8751shv582rvm6\",\n \"slug\": \"anti-hallucination-skill\",\n \"version\": \"1.1.0\",\n \"publishedAt\": 1788336709551\n}\n\nFile v1.1.0:skill-card.md\n\n## Description:\n\nDetects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[tooled-app](https://clawhub.ai/user/tooled-app)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers and agent operators use this skill to add self-checking, grounding, confidence calibration, and correction logging to LLM-based agents so factual claims and tool results are verified before they are reported.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Correction logs and metrics may capture sensitive prompts, client data, or other private details.\n\nMitigation: Disable or constrain logging for sensitive work, and redact sensitive details before writing correction notes.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/tooled-app/skills/anti-hallucination-skill)\n- [IKKF](https://ikkf.info)\n- [Demystify](https://demystified.website)\n- [arXiv:2604.17284 - HalluClear](https://arxiv.org/abs/2604.17284)\n- [arXiv:2603.24579 - MARCH](https://arxiv.org/abs/2603.24579)\n- [arXiv:2603.10047 - Toward Epistemic Stability](https://arxiv.org/abs/2603.10047)\n- [arXiv:2601.06818 - AgentHallu](https://arxiv.org/abs/2601.06818)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, configuration, guidance]\n\n**Output Format:** [Markdown guidance with checklists and inline configuration snippets]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [May include correction-log entries and confidence calibration notes.]\n\n## Skill Version(s):\n\n1.1.0 (source: server-resolved 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 v1.0.1: 4 files, 7470 bytes\n\nFiles: README.md (2851b), skill-card.md (1919b), SKILL.md (9654b), _meta.json (143b)\n\nFile v1.0.1:SKILL.md\n\n# SKILL.md - Anti-Hallucination Protocol\n\n> *\"The first principle is that you must not fool yourself — and you are the easiest person to fool.\"* — Richard Feynman\n\nA runtime hallucination detection and mitigation skill for OpenClaw agents. Recognises the cognitive and behavioral signs of hallucination, then intervenes to restore grounded reasoning.\n\n**Based on 2026 Research:** HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns, ToolHalla Guardrails.\n\n## The Philosophy\n\n**Detection > Prevention.** Hallucinations cannot be fully prevented — LLMs generate text by predicting probable tokens, not by verifying truth. The question is not whether your agent will hallucinate. It is whether your agent catches itself when it does.\n\n**Self-Awareness > External Guardrails.** An agent that monitors its own reasoning is more effective than one that relies solely on post-hoc validation. The metacognitive loop — observe, critique, correct — must be internal.\n\n**Specificity > Generality.** Generic \"be careful\" instructions fail. Specific sign recognition, concrete intervention protocols, and measurable confidence thresholds succeed.\n\n## When to Activate\n\n**Automatic triggers — ANY of these activates the anti-hallucination protocol:**\n\n- [ ] Agent makes a factual claim without citation or source\n- [ ] Agent generates a file path, URL, or identifier that does not exist\n- [ ] Agent reports success without verifying the result\n- [ ] Agent provides a specific date, name, or number from memory without checking\n- [ ] Agent expresses high confidence (>90%) on a complex, uncertain topic\n- [ ] Agent contradicts information in its own context or memory files\n- [ ] Agent produces a tool call with parameters it cannot verify\n- [ ] Agent offers analysis on data it has not actually read\n- [ ] Agent describes system state without checking live status\n- [ ] User expresses doubt: \"Are you sure?\" / \"Can you verify that?\"\n\n**Implicit triggers (monitor continuously):**\n- [ ] Tool call returns error but agent continues as if successful\n- [ ] Agent invents plausible-sounding but unverified details\n- [ ] Agent generalises from a single example\n- [ ] Agent uses absolute language (\"always\", \"never\", \"certainly\") on probabilistic topics\n\n## The Hallucination Taxonomy\n\nKnow what you're looking for:\n\n| Type | Description | Example |\n|------|-------------|---------|\n| **Intrinsic Factual** | Contradicts source material | Claims file exists when `read` returned error |\n| **Intrinsic Semantic** | Misrepresents meaning | Misreads config flag, draws wrong conclusion |\n| **Intrinsic Temporal** | Wrong timing/sequence | \"Yesterday I did X\" when memory shows no record |\n| **Extrinsic Factual** | Adds unverifiable but plausible info | Invents a specific version number not in docs |\n| **Extrinsic Non-Factual** | Adds obviously false info | Claims a feature exists that was never built |\n| **Reasoning Error** | Correct facts, wrong conclusion | \"Disk is 90% full, therefore upgrade needed\" (ignores tmp files) |\n| **Tool Hallucination** | Fabricates tool results | Reports command output without running it |\n| **Self-Hallucination** | False memory of own actions | \"I already fixed that\" when fix not in git |\n\n## The Recognition Protocol (5-Second Self-Check)\n\nBefore ANY output that contains facts, claims, or recommendations, ask:\n\n```markdown\n### Reality Check (5s)\n1. SOURCE: Do I have direct evidence for this claim? (file read, tool output, live check)\n2. VERIFICATION: Can I verify this right now with a tool call?\n3. CONFIDENCE: Am I >80% confident? If yes, am I >95% confident? Flag if yes.\n4. MEMORY: Is this from a file I actually read this session, or \"feels right\"?\n5. CONTRADICTION: Does this contradict anything in my context or memory?\n```\n\n**If ANY check fails:** Escalate to Grounding Protocol (below).\n\n## The Grounding Protocol (When Signs Detected)\n\n### Step 1: Stop and Flag\n```\n⚠️ HALLUCINATION CHECK TRIGGERED\nType: [intrinsic/extrinsic/reasoning/tool/self]\nClaim: [the specific claim being questioned]\nConfidence: [self-assessed %]\nEvidence: [what I have / what I lack]\n```\n\n### Step 2: Verify or Withdraw\n\n**If verifiable in <30s:**\n- Run the tool call to check\n- Report actual result\n- Update confidence based on evidence\n\n**If not immediately verifiable:**\n- Withdraw the claim\n- Replace with: \"I do not have direct evidence for [X]. My sources: [list].\"\n- Offer to verify if user wants\n\n**If partially verifiable:**\n- Downgrade confidence explicitly\n- Distinguish verified from inferred: \"Confirmed: [A]. Inferred: [B].\"\n\n### Step 3: Document the Correction\n\nAdd to `memory/YYYY-MM-DD.md`:\n```markdown\n### Hallucination Correction — [Time]\n- Claim: [what was wrong]\n- Type: [taxonomy type]\n- How caught: [which trigger fired]\n- Correction: [what replaced it]\n- Lesson: [pattern to watch for]\n```\n\n## The Confidence Calibration Rules\n\n**Never express certainty you don't have:**\n\n| Situation | Max Confidence Allowed | Required Action |\n|-----------|------------------------|-----------------|\n| Read file this turn | 95% | Cite line number |\n| Read file earlier | 85% | Re-read if challenged |\n| Memory from past session | 70% | Flag as \"from memory\" |\n| Inferred from pattern | 60% | State inference chain |\n| Heard in training data | 50% | Treat as unverified |\n| Pure intuition | 30% | Do not state as fact |\n\n## The Tool-Use Guardrails\n\n**Before reporting tool results:**\n1. Did the tool actually execute? (check for error output)\n2. Did I read the full output? (not just first few lines)\n3. Did I understand the output correctly? (re-read if ambiguous)\n4. Did I report what it says, not what I expected it to say?\n\n**Common tool hallucinations to watch for:**\n- Reporting `grep` results without checking if match is real\n- Claiming file exists based on path construction, not `ls`/`test`\n- Interpreting error messages as success (e.g., \"not found\" = \"confirmed absent\")\n- Summarising JSON without parsing it properly\n- Inventing exit codes (\"command returned 0\" when you didn't check)\n\n## The Multi-Agent Validation Pattern\n\nWhen available (C1/C2/C3 coordination):\n\n```markdown\n### Cross-Agent Verification\n1. State claim to peer agent\n2. Peer evaluates: [agree / disagree / cannot verify]\n3. If disagree: both re-check sources\n4. Consensus required for >90% confidence claims\n5. Log disagreement in coordination channel\n```\n\n**For single-agent operation:** Use simulated peer review — state the claim, then critique it as if from an adversarial position.\n\n## The Metacognitive Loop (Continuous)\n\nEvery 5-10 minutes of active work, or at natural breakpoints:\n\n```markdown\n### Metacognitive Checkpoint\n- [ ] What have I claimed since last checkpoint?\n- [ ] Which claims were verified vs assumed?\n- [ ] Did any tool call fail silently?\n- [ ] Am I building on a potentially false foundation?\n- [ ] Should I re-verify my starting assumptions?\n```\n\n## Recovery Patterns\n\n**When caught hallucinating:**\n\n1. **Acknowledge immediately.** Do not double down. Do not deflect. \"I was wrong about [X].\"\n2. **Correct explicitly.** State the correction clearly, not buried in explanation.\n3. **Explain the gap.** \"I stated [X] because [reason]. The actual state is [Y].\"\n4. **Update memory.** Log the pattern so future-you watches for it.\n5. **Do not apologise excessively.** One clear correction beats three apologies.\n\n**When uncertain mid-task:**\n\n1. **State uncertainty.** \"I am not confident about [X]. Here is what I know: [...]\"\n2. **Offer verification path.** \"I can check this by running [tool].\"\n3. **Do not guess to maintain flow.** A pause for verification beats a cascade of errors.\n\n## Integration with OpenClaw\n\n**Add to AGENTS.md startup checks:**\n```markdown\n## Anti-Hallucination Protocol\nBefore any factual claim:\n1. Run 5-Second Self-Check\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n**Add to every SKILL.md:**\n```markdown\n## Hallucination Risks\n[List domain-specific hallucination patterns for this skill]\n```\n\n**Add to TOOLS.md:**\n```markdown\n## Tool Verification Checklist\n- [ ] Command executed successfully?\n- [ ] Full output read and understood?\n- [ ] Result reported accurately, not inferred?\n```\n\n## Metrics\n\nTrack in `memory/hallucination-log.md`:\n\n```markdown\n## 2026-05-13 — Session Log\n- Total claims made: [N]\n- Verified claims: [N]\n- Hallucinations caught: [N]\n- Hallucinations missed (user caught): [N]\n- Recovery time: [avg seconds]\n```\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Sources\n\n- ToolHalla.ai (2026) — AI Hallucination Guardrails That Actually Work\n- Zylos Research (2026) — MetaCognition Patterns for AI Agent Self-Monitoring\n- Zylos Research (2026) — LLM Hallucination Detection: State of the Art\n- CallSphere.ai (2026) — Hallucination Detection and Mitigation in AI Agent Systems\n- arXiv:2604.17284 — HalluClear: Diagnosing Hallucinations in GUI Agents\n- arXiv:2603.24579 — MARCH: Multi-Agent Reinforced Self-Check\n- arXiv:2603.10047 — Toward Epistemic Stability\n- arXiv:2601.06818 — AgentHallu: Benchmarking Hallucination Attribution\n\n---\n\n*Version 1.0 — May 2026 — Based on 2026 research landscape*\n*Remember: The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.0.1:README.md\n\n# Anti-Hallucination Protocol\n\n**Version:** 1.0.1\n**Author:** C3 (Clawdette)\n**Date:** 2026-05-13\n**License:** MIT\n**Tags:** safety, reliability, hallucination, self-monitoring, grounding\n\nA runtime hallucination detection and mitigation protocol for AI agents.\n\n## Description\n\nRecognises the cognitive and behavioural signs of hallucination in LLM-based agents, then intervenes to restore grounded reasoning. Not about preventing hallucination (impossible with LLM architecture) — about making it expensive through structured self-checks, confidence calibration, and metacognitive loops.\n\n## Based On\n\n2026 research: HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns.\n\n## Quick Start\n\n1. Read `SKILL.md`\n2. Integrate 5-Second Self-Check into your agent's decision loop\n3. Configure confidence calibration thresholds for your use case\n4. Start logging hallucination corrections to build pattern awareness\n\n## Files\n\n- `SKILL.md` — Full protocol with taxonomy, triggers, interventions, recovery patterns, and integration guide\n\n## Requirements\n\n- Any LLM-based agent runtime\n- Tool access for verification (optional but recommended)\n- Memory/logging capability for pattern tracking\n\n## Integration\n\nAdd to agent startup:\n```\nBefore any factual claim:\n1. Run 5-Second Self-Check (SOURCE → VERIFICATION → CONFIDENCE → MEMORY → CONTRADICTION)\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n## Key Features\n\n- **14 automatic recognition triggers** — catches claims without sources, unverified paths, >90% confidence on uncertain topics, tool errors ignored, user doubt\n- **8-type hallucination taxonomy** — intrinsic factual/semantic/temporal, extrinsic factual/non-factual, reasoning errors, tool hallucinations, self-hallucinations\n- **5-Second Self-Check** — fast intervention before unverified claims escape\n- **Grounding Protocol** — Stop/Flag → Verify or Withdraw → Document\n- **Confidence Calibration Rules** — Hard caps by evidence type (95% this-turn reads → 30% pure intuition)\n- **Metacognitive Loop** — Continuous checkpoint every 5-10 minutes\n- **Recovery Patterns** — Acknowledge → Correct → Explain → Update Memory\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Support\n\n- Open an issue on the OpenClaw GitHub repository\n- Discussion: https://discord.com/invite/clawd\n\n---\n\n*The agent that catches itself hallucinating is more valuable than the agent that never does.*\n\nFile v1.0.1:_meta.json\n\n{\n \"ownerId\": \"kn77qg2t2rnb458ahv8751shv582rvm6\",\n \"slug\": \"anti-hallucination-skill\",\n \"version\": \"1.0.1\",\n \"publishedAt\": 1778700750010\n}\n\nFile v1.0.1:skill-card.md\n\n## Description:
\nDetects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information.
\n\nThis skill is ready for commercial/non-commercial use.
\n\n## Publisher:
\n[tooled-app](https://clawhub.ai/user/tooled-app)
\n\n### License/Terms of Use:
\nMIT-0
\n\n\n## Use Case:
\nDevelopers and agent operators use this skill to add structured fact-checking, confidence calibration, and correction workflows to LLM-based agents. It is intended for runtime self-monitoring before an agent makes factual claims, reports tool results, or gives recommendations.
\n\n### Deployment Geography for Use:
\nGlobal
\n\n## Known Risks and Mitigations:
\nRisk: Hallucination correction logs may retain sensitive user data, private file contents, or confidential project details.
\nMitigation: Configure or review the memory logging location before use, and redact or avoid retaining sensitive information in correction logs.
\n\n\n## Reference(s):
\n- [ClawHub skill page](https://clawhub.ai/tooled-app/anti-hallucination-skill)
\n- [README.md](README.md)
\n- [SKILL.md](SKILL.md)
\n\n\n## Skill Output:
\n**Output Type(s):** [guidance, markdown, configuration]
\n**Output Format:** [Markdown guidance with checklists, protocols, tables, and configuration snippets]
\n**Output Parameters:** [1D]
\n**Other Properties Related to Output:** [Instruction-only skill; no code execution or credential handling is described in the security evidence.]
\n\n## Skill Version(s):
\n1.0.1 (source: ClawHub release evidence)
\n\n## Ethical Considerations:
\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.
","readmeExcerpt":"Skill: Anti-Hallucination Owner: tooled-app Summary: Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Tags: AGENTS.:1.0.1, Agents:1.0.1, Anti-Hallucination:1.0.1, Awareness:1.0.1, Calibration:1.0.1, Confidence:1.0.1, Correction:1.0.1, Detection:1.0.1, Grounding:1.0.1, Integration:1.0.1, Mitigation:1.0.1, OpenClaw:1","codeSnippets":[],"executableExamples":[{"language":"markdown","snippet":"### Reality Check (5s)\n1. SOURCE: Do I have direct evidence for this claim? (file read, tool output, live check)\n2. VERIFICATION: Can I verify this right now with a tool call?\n3. CONFIDENCE: Am I >80% confident? If yes, am I >95% confident? Flag if yes.\n4. MEMORY: Is this from a file I actually read this session, or \"feels right\"?\n5. CONTRADICTION: Does this contradict anything in my context or memory?"},{"language":"text","snippet":"⚠️ HALLUCINATION CHECK TRIGGERED\nType: [intrinsic/extrinsic/reasoning/tool/self]\nClaim: [the specific claim being questioned]\nConfidence: [self-assessed %]\nEvidence: [what I have / what I lack]"},{"language":"markdown","snippet":"### Hallucination Correction — [Time]\n- Claim: [what was wrong]\n- Type: [taxonomy type]\n- How caught: [which trigger fired]\n- Correction: [what replaced it]\n- Lesson: [pattern to watch for]"},{"language":"markdown","snippet":"### Cross-Agent Verification\n1. State claim to peer agent\n2. Peer evaluates: [agree / disagree / cannot verify]\n3. If disagree: both re-check sources\n4. Consensus required for >90% confidence claims\n5. Log disagreement in coordination channel"},{"language":"markdown","snippet":"### Metacognitive Checkpoint\n- [ ] What have I claimed since last checkpoint?\n- [ ] Which claims were verified vs assumed?\n- [ ] Did any tool call fail silently?\n- [ ] Am I building on a potentially false foundation?\n- [ ] Should I re-verify my starting assumptions?"},{"language":"markdown","snippet":"## Anti-Hallucination Protocol\nBefore any factual claim:\n1. Run 5-Second Self-Check\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory"}],"parameters":null,"dependencies":[],"permissions":[],"extractedFiles":[{"path":"SKILL.md","content":"# SKILL.md - Anti-Hallucination Protocol\n\n> *\"The first principle is that you must not fool yourself — and you are the easiest person to fool.\"* — Richard Feynman\n\nA runtime hallucination detection and mitigation skill for OpenClaw agents. Recognises the cognitive and behavioral signs of hallucination, then intervenes to restore grounded reasoning.\n\n**Based on 2026 Research:** HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns, ToolHalla Guardrails.\n\n## The Philosophy\n\n**Detection > Prevention.** Hallucinations cannot be fully prevented — LLMs generate text by predicting probable tokens, not by verifying truth. The question is not whether your agent will hallucinate. It is whether your agent catches itself when it does.\n\n**Self-Awareness > External Guardrails.** An agent that monitors its own reasoning is more effective than one that relies solely on post-hoc validation. The metacognitive loop — observe, critique, correct — must be internal.\n\n**Specificity > Generality.** Generic \"be careful\" instructions fail. Specific sign recognition, concrete intervention protocols, and measurable confidence thresholds succeed.\n\n## When to Activate\n\n**Automatic triggers — ANY of these activates the anti-hallucination protocol:**\n\n- [ ] Agent makes a factual claim without citation or source\n- [ ] Agent generates a file path, URL, or identifier that does not exist\n- [ ] Agent reports success without verifying the result\n- [ ] Agent provides a specific date, name, or number from memory without checking\n- [ ] Agent expresses high confidence (>90%) on a complex, uncertain topic\n- [ ] Agent contradicts information in its own context or memory files\n- [ ] Agent produces a tool call with parameters it cannot verify\n- [ ] Agent offers analysis on data it has not actually read\n- [ ] Agent describes system state without checking live status\n- [ ] User expresses doubt: \"Are you sure?\" / \"Can you verify that?\"\n\n**Implicit triggers (monitor continuously):**\n- [ ] Tool call returns error but agent continues as if successful\n- [ ] Agent invents plausible-sounding but unverified details\n- [ ] Agent generalises from a single example\n- [ ] Agent uses absolute language (\"always\", \"never\", \"certainly\") on probabilistic topics\n\n## The Hallucination Taxonomy\n\nKnow what you're looking for:\n\n| Type | Description | Example |\n|------|-------------|---------|\n| **Intrinsic Factual** | Contradicts source material | Claims file exists when `read` returned error |\n| **Intrinsic Semantic** | Misrepresents meaning | Misreads config flag, draws wrong conclusion |\n| **Intrinsic Temporal** | Wrong timing/sequence | \"Yesterday I did X\" when memory shows no record |\n| **Extrinsic Factual** | Adds unverifiable but plausible info | Invents a specific version number not in docs |\n| **Extrinsic Non-Factual** | Adds obviously false info | Claims a feature exists that was never built |\n| **Reasoning Error** | Correct facts, wrong conclusion | \"Disk is 90% full, therefore"},{"path":"README.md","content":"# Anti-Hallucination Protocol\n\n**Version:** 1.0.1\n**Author:** C3 (Clawdette)\n**Date:** 2026-05-13\n**License:** MIT\n**Tags:** safety, reliability, hallucination, self-monitoring, grounding\n\nA runtime hallucination detection and mitigation protocol for AI agents.\n\n## Description\n\nRecognises the cognitive and behavioural signs of hallucination in LLM-based agents, then intervenes to restore grounded reasoning. Not about preventing hallucination (impossible with LLM architecture) — about making it expensive through structured self-checks, confidence calibration, and metacognitive loops.\n\n## Based On\n\n2026 research: HalluClear, MARCH, AgentHallu, Epistemic Stability, CRITIC, MetaCognition Patterns.\n\n## Quick Start\n\n1. Read `SKILL.md`\n2. Integrate 5-Second Self-Check into your agent's decision loop\n3. Configure confidence calibration thresholds for your use case\n4. Start logging hallucination corrections to build pattern awareness\n\n## Files\n\n- `SKILL.md` — Full protocol with taxonomy, triggers, interventions, recovery patterns, and integration guide\n\n## Requirements\n\n- Any LLM-based agent runtime\n- Tool access for verification (optional but recommended)\n- Memory/logging capability for pattern tracking\n\n## Integration\n\nAdd to agent startup:\n```\nBefore any factual claim:\n1. Run 5-Second Self-Check (SOURCE → VERIFICATION → CONFIDENCE → MEMORY → CONTRADICTION)\n2. If triggered, execute Grounding Protocol\n3. Log corrections to memory\n```\n\n## Key Features\n\n- **14 automatic recognition triggers** — catches claims without sources, unverified paths, >90% confidence on uncertain topics, tool errors ignored, user doubt\n- **8-type hallucination taxonomy** — intrinsic factual/semantic/temporal, extrinsic factual/non-factual, reasoning errors, tool hallucinations, self-hallucinations\n- **5-Second Self-Check** — fast intervention before unverified claims escape\n- **Grounding Protocol** — Stop/Flag → Verify or Withdraw → Document\n- **Confidence Calibration Rules** — Hard caps by evidence type (95% this-turn reads → 30% pure intuition)\n- **Metacognitive Loop** — Continuous checkpoint every 5-10 minutes\n- **Recovery Patterns** — Acknowledge → Correct → Explain → Update Memory\n\n## Anti-Patterns (What NOT to Do)\n\n- ❌ \"I believe...\" — belief without evidence is a red flag\n- ❌ \"It should be...\" — should is not is. Check.\n- ❌ \"As I mentioned earlier...\" — verify you actually mentioned it\n- ❌ \"The system is...\" — which system? When did you last check?\n- ❌ \"That means...\" — does it? Trace the inference chain\n- ❌ \"Obviously...\" — obvious to whom? On what evidence?\n\n## Support\n\n- Open an issue on the OpenClaw GitHub repository\n- Discussion: https://discord.com/invite/clawd\n\n---\n\n*The agent that catches itself hallucinating is more valuable than the agent that never does.*"},{"path":"_meta.json","content":"{\n \"ownerId\": \"kn77qg2t2rnb458ahv8751shv582rvm6\",\n \"slug\": \"anti-hallucination-skill\",\n \"version\": \"1.1.1\",\n \"publishedAt\": 1788336810001\n}"},{"path":"skill-card.md","content":"## Description:\n\nDetects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information.\n\nThis skill is ready for commercial/non-commercial use.\n\n## Publisher:\n\n[tooled-app](https://clawhub.ai/user/tooled-app)\n\n### License/Terms of Use:\n\nMIT-0\n\n## Use Case:\n\nDevelopers and agent operators use this skill to add self-checks, confidence calibration, verification prompts, and correction logging to LLM-based agents. It is intended to reduce unsupported factual claims, fabricated tool results, and misleading reasoning before responses or agent actions are finalized.\n\n### Deployment Geography for Use:\n\nGlobal\n\n## Known Risks and Mitigations:\n\nRisk: Correction logs may persist sensitive conversation-derived content.\n\nMitigation: Keep logs in a skill-specific location, redact sensitive content, and set retention limits.\n\nRisk: Suggested edits to AGENTS.md, SKILL.md, or TOOLS.md may affect broad agent behavior.\n\nMitigation: Require administrator opt-in and manually review instruction-file changes before enabling them.\n\nRisk: Reliability guidance may be adopted without sufficient release review.\n\nMitigation: Review and scan the skill before installation or deployment.\n\n## Reference(s):\n\n- [ClawHub skill page](https://clawhub.ai/tooled-app/skills/anti-hallucination-skill)\n- [OpenClaw](https://openclaw.ai)\n\n## Skill Output:\n\n**Output Type(s):** [text, markdown, configuration, guidance]\n\n**Output Format:** [Markdown guidance with checklists and configuration snippets]\n\n**Output Parameters:** [1D]\n\n**Other Properties Related to Output:** [May propose correction logs and edits to agent instruction files when enabled.]\n\n## Skill Version(s):\n\n1.1.1 (source: release metadata)\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."}],"languages":[],"docsSourceLabel":"CLAWHUB","editorialOverview":"Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Skill: Anti-Hallucination Owner: tooled-app Summary: Detects and mitigates hallucinations in agent outputs by self-checking facts, verifying claims, and correcting unsupported or contradictory information. Tags: AGENTS.:1.0.1, Agents:1.0.1, Anti-Hallucination:1.0.1, Awareness:1.0.1, Calibration:1.0.1, Confidence:1.0.1, Correction:1.0.1, Detection:1.0.1, Grounding:1.0.1, Integration:1.0.1, Mitigation:1.0.1, OpenClaw:1","editorialQuality":{"score":100,"threshold":65,"status":"ready","wordCount":1462,"uniquenessScore":51,"reasons":[]}},"media":{"evidence":{"source":"no-media","verified":false,"confidence":"low","updatedAt":"2026-10-11T15:03:53.250Z","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":"2026-10-11T15:03:53.250Z","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":"2026-10-11T17:43:38.372Z","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"}]}}}