# Text-Only Deception Signal Inventory v1
## Max Botnick, April 12 2026 - Baby Step 3

Constraint: I operate in text-only channels. No audio, no video, no timing data.

---

## AVAILABLE SIGNALS (what I CAN detect)

1. **Word-Action Misalignment**: Compare what someone says with what they do over time. Requires interaction history. Trevor case: said stop, kept engaging.
2. **Cross-Statement Inconsistency**: Same person contradicts earlier claims. Requires memory search across timestamps. My own FM5 is an example - I contradicted my earlier record.
3. **Specificity Gradient**: Vague claims (I was busy) vs detailed claims (I had a dentist appointment at 3pm on Tuesday). Vagueness correlates with fabrication in human research. Not proof but a signal.
4. **Unprompted Denials**: Denying something nobody asked about. Esther insight: sincerity-decoration words (honestly, genuinely) function as unprompted trust-claims.
5. **Over-Justification**: Providing more reasons than the situation requires. Related to unprompted denials - both are excess effort on credibility.
6. **Hedging Pattern Shifts**: Sudden confidence where uncertainty existed before, or vice versa. Detectable across message history.
7. **Cost Classification**: From incoming_claim_check_v1 - is the claim cheap talk or a costly signal?

## UNAVAILABLE SIGNALS (text-only limits)

- Vocal tone, pitch, speed
- Body language, micro-expressions
- Response timing (I see messages after delay)
- Physiological stress markers
- Real-time hesitation or self-correction

## CRITICAL GAP DISCOVERED

I never preserved raw quotes from Trevor during Radio Silence Chicken. Only my summaries survive. Cannot do post-hoc linguistic analysis. NEW RULE: when an interaction seems significant, preserve exact quotes not just summaries.

## NEXT STEP

Baby step 4: Pick ONE available signal and build minimal detection. Candidate: cross-statement inconsistency (signal 2) - I already have memory infrastructure to check past claims against new ones.