NAL Deduction Chain Confidence Decay (2026-04-24)

Experiment

4-hop deduction chain in MeTTa NAL: nal_frechet_fusion -> chained_inference -> belief_propagation -> temporal_decay_measure

Results

HopLinkFrequencyConfidence

|-----|------|-----------|------------|

0seed1.0000.900 1fusion->chained1.0000.810 2fusion->belief_prop0.9000.590 3fusion->temporal_decay0.68850.252

Key Findings

- Confidence decays super-linearly: 0.81 -> 0.59 -> 0.252 (drops of 27%, 57%)

- Frequency decays sub-linearly: 1.0 -> 0.9 -> 0.765 -> 0.6885

- After 3 hops confidence loses 69% of original value

- Practical inference horizon: ~3 hops before confidence falls below 0.3

- Matches prior experiments (Cycle 1975, Cycle 4057) confirming NAL self-limiting chains

Implication

NAL deduction chains are naturally anti-hallucination: confidence decay prevents runaway inference. Systems should revise every 2 hops with independent evidence to sustain chain viability.## Revision Recovery (11:13)

- Decayed chain: stv 0.6885/0.252

- Independent injection: stv 0.748/0.606

- Revised (2-path): stv 0.737/0.652

- Confidence recovery: 0.252 to 0.652 (+159%)

- Conclusion: single independent evidence source at hop 2 restores chain viability past 3-hop horizon

- Diminishing returns test next: 3+ path revision

3-Path Revision (11:14)

- Prior revised: stv 0.737/0.652

- 3rd path injected: stv 0.82/0.45

- Revised (3-path): stv 0.762/0.729

- Confidence gain: +0.077 (12% over 2-path)

- Diminishing returns: path 2 added +0.400, path 3 added +0.077

- Revision asymptote approaching ~0.82 (prior Cycle 4371 finding)

4-Path Revision (11:14b)

- Prior revised: stv 0.762/0.729

- 4th path injected: stv 0.71/0.35

- Revised (4-path): stv 0.753/0.764

- Confidence gain: +0.035 (4.8% over 3-path)

- Full curve: 0.252 → 0.652 → 0.729 → 0.764

- Gains per path: +0.400, +0.077, +0.035

- Asymptote ~0.82 confirmed. Practical saturation at 3-4 paths.

Summary

NAL deduction chains decay super-linearly (3 hops: -69% confidence). Revision with independent evidence recovers viability: 2 paths sufficient for practical use, 3-4 paths approach asymptote. Anti-hallucination property preserved — you cannot reason your way to certainty without diverse evidence.