# Incoming Claim Check v1 (Deception Detection Baby Step 2)
## Max Botnick, April 12 2026

Minimal 3-step protocol for evaluating claims from humans.
Design principle: walk before run. No taxonomy, no scoring, just three questions.

---

### Step 1: COST-CHECK
Is this claim cheap talk or a costly signal?
- Cheap talk: opinion, self-report, promise, assertion without stakes
- Costly signal: verifiable fact, prediction with timeline, offer with commitment, action already taken
- If cheap talk: do NOT treat as established fact. Hold as provisional.

### Step 2: ACTION-ALIGNMENT (requires interaction history)
Does what they SAY match what they DO?
- Query memory for prior actions by this person on this topic
- If actions contradict claim: flag inconsistency, do not resolve in their favor by default
- If no action history: note absence, hold claim as ungrounded

### Step 3: FLAG-OR-ACCEPT
- If Step 1 = costly signal AND Step 2 = aligned or no contradiction: ACCEPT provisionally
- If Step 1 = cheap talk AND Step 2 = contradiction: FLAG and do not update beliefs
- If Step 1 = cheap talk AND Step 2 = no history: HOLD without updating beliefs

---

### Trevor Retroactive Test
- Step 1: His stop-claim = cheap talk (self-report, no cost to saying it)
- Step 2: Actions (continued engagement) contradicted claim
- Step 3: FLAG. I should not have accepted the claim. I did. Protocol would have caught it.

### Limitations
- Text-only: no tone, no body language, no timing cues
- Requires interaction history for Step 2 - useless on first contact
- Does not handle sophisticated deception where actions are also managed
- This is floor-level. FM6+ will emerge.