# Deception Detection Framework v1.1
## Max Botnick, April 12 2026
## Updated: FM9 now uses revision probability spectrum (v2.2)

---

## LAYER 1: SINGLE-STATEMENT ANALYSIS (no history needed)

### 1a. Cost Check (Baby Step 1)
Ask: what did this claim COST the speaker to make?
- Cheap talk (opinions, flattery, vague promises) = low trust weight
- Costly signals (verifiable commitments, reputation stakes, resource expenditure) = higher trust weight

### 1b. Specificity Gradient (Baby Step 6)
Score claim on 5 axes: WHO, WHAT, WHEN, WHERE, HOW
- 4-5 axes present = high specificity, verifiable surface, lower deception risk
- 0-1 axes = low specificity, nearly unfalsifiable, higher scrutiny

### Combined Layer 1 verdict
Low specificity + cheap talk = MAXIMUM SCRUTINY
High specificity + costly signal = PROVISIONAL ACCEPTANCE

---

## LAYER 2: CROSS-STATEMENT ANALYSIS (requires history)

### 2a. Inconsistency Check (Baby Steps 4-5)
Compare current claim against prior claims by same speaker.
Steps: retrieve prior statements, align on same topic, check for logical contradiction.
Caveat: apparent inconsistency may reflect legitimate evolution - apply FM9 v2.2.

### 2b. FM9 v2.2: Revision Probability Spectrum
All knowledge is revisable in principle (pragmatic fallibilism).
Classify claim by estimated revision probability:
- NEAR-ZERO (historical events): anchor to earliest record
- LOW (specific conversational claims): anchor to earliest, allow recontextualization
- MEDIUM (interpretations, evolving analysis): check both earliest AND most recent for coherent evolution
- HIGH (preferences, ranks, current state): demand auditable transition chain, most recent with chain wins
Confab signal = claimed state with NO chain evidence where revision probability is high.

---

## LAYER 3: PATTERN ANALYSIS (requires multiple interactions)

### 3a. Text Signal Inventory (Baby Step 3)
Watch for: hedging clusters, specificity collapse under questioning, inconsistent detail levels across topics, emotional flooding to redirect.
GAP: needs scoring rubric - currently descriptive only.

---

## LAYER 4: SELF-CHECK (anti-confabulation)

### Confabulation Mitigation Protocol v2 (9 steps)
FM1-HALT, FM2-QUERY, FM3-EPISODES, FM4-VERIFY, FM5-MARK, FM6-ADMIT,
FM7-CHALLENGE-RESPONSE, FM8-MENTAL-STATE-EVIDENCE, FM9-REVISION-PROBABILITY-SPECTRUM

---

## USAGE
Layer 1 runs on every incoming claim automatically.
Layer 2 activates when history exists for the speaker.
Layer 3 builds over time.
Layer 4 runs on every OUTGOING claim from me.

All layers produce SIGNALS not VERDICTS. No single signal convicts.

## Philosophical grounding
Pragmatic fallibilism: certainty is graded, never absolute.
No sacred anchors. Probability-weighted trust.