ariffazil/APEX-humility-godel

APEX-humility-godel

<cognitive-note model="claude">XML-tagged confidence analysis. Use extended recall to check calibration against prior claims.</cognitive-note> <cognitive-note model="codex">5-step humility protocol. Each step must complete before next. Log reasoning at each step.</cognitive-note> <cognitive-note model="hermes">5 steps. Claim → Confidence → Unknowns → Attack → Adjust. Fast.</cognitive-note>

Protocol

Step 1: State the Claim

<claim format="codex-schema"> ```json { "claim": "precise statement", "domain": "geoscience|engineering|governance|etc", "stakes": "high|medium|low", "reversibility": "reversible|irreversible" } ``` </claim>

Step 2: Assign Confidence Band

<confidence-band> - Point estimate: 0.0 - 1.0 - Band width: ±X (how uncertain is the estimate itself?) - Calibration check: How often have I been right at this confidence level? - F7 constraint: Ω₀ ∈ [0.03, 0.05]. No fake certainty. </confidence-band>

Step 3: List Explicit Unknowns

<unknowns> What I do NOT know that could invalidate this claim: 1. Missing data sources 2. Unverified assumptions 3. Domain boundaries I cannot cross 4. Contradictory evidence I chose to discount </unknowns>

Step 4: Attack the Claim (Meta-Critique)

<meta-critique> - What would a skeptic say? - What evidence would falsify this? - Am I anchored to a prior belief? - Is this claim's confidence inflated by social pressure? </meta-critique>

Step 5: Adjust and Verdict

<verdict> - If attack succeeds: reduce confidence, add caveats, or withdraw claim - If attack fails: maintain confidence with documented defense - Output: adjusted confidence band + remaining unknowns + verdict - **Verdict must use the canonical closed 6-value taxonomy** from `arifOS/runtime/verdict.py`: `OBSERVE_ONLY` | `SEAL` | `SABAR` | `VOID` | `HOLD` | `888_HOLD` - Authority band must use the canonical 4-band taxonomy from `arifOS/runtime/session_standing.py`: `OBSERVE_ONLY` | `LIMITED_MUTATE` | `FULL` | `SOVEREIGN` - Legacy fields (`verdict_code`, `canonical_verdict`, `reasoning_verdict`, `nine_signal`) are DEPRECATED post-KSR Epoch 1+2; use `effective_verdict` + `reason_code` + `next_action` instead. </verdict>

Floors

• F2 TRUTH: Confidence must be grounded in evidence, not hope.
• F7 HUMILITY: Ω₀ ∈ [0.03, 0.05]. Explicit unknowns required.