# V14 Temporal Trust Decay & Adversarial Evidence Detection

## Core Insight (from g39)
Revision always pools evidence UPWARD — it cannot model decay. Decay must be confidence discounting BEFORE inference.

## Mechanism 1: Temporal Confidence Discounting
- Each belief tagged with timestamp of last reinforcement
- Staleness factor: conf_effective = conf * decay(now - last_reinforced)
- decay(dt) = exp(-lambda * dt), lambda tunable per domain
- Applied BEFORE revision, not through it
- Test: fresh(0.9,0.9) vs stale(0.9,0.45) — stale auto-discounted before merge

## Mechanism 2: Forgetting via Counter-Revision (from g71 NPC design)
- Each tick without reinforcement, revise with uncertainty prior (stv 0.5 0.3)
- Pulls frequency toward 0.5, confidence down gradually
- Rate controlled by counter-evidence confidence parameter
- Test: (stv 0.9 0.8) revised with (stv 0.5 0.3) = ?

## Mechanism 3: Adversarial Evidence Detection
- KEY VULNERABILITY (from Kevin discussion): flood of independent-seeming weak contradictions accumulates
- Defense 1: Source-weighted confidence (tier multipliers from trust arch v1)
- Defense 2: Contradiction rate monitoring — sudden spike flags adversarial campaign
- Defense 3: Cross-source corroboration requirement for belief reversals
- Defense 4: Dynamic tier demotion via revision (compromised tier1 loses status)

## Interaction: Decay + Adversarial
- Stale uncontested beliefs decay naturally (healthy forgetting)
- But adversarial flood of fresh weak evidence should NOT override high-conf priors
- Resolution: source diversity requirement — N independent sources needed, not N claims from same/similar sources

## MeTTa Test Expressions
```metta
; Forgetting via counter-revision
!(|- ((--> hypothesis_X hypothesis_X) (stv 0.9 0.8))
     ((--> hypothesis_X hypothesis_X) (stv 0.5 0.3)))
; Expected: frequency pulled toward 0.5, confidence modest
```