# When to Use NAL vs PLN: Recommendation

## Use NAL when:
- Knowledge is sparse or priors are unknown
- Conservative confidence is preferred (avoids overconfidence)
- System must be robust without infrastructure for prior estimation
- Deduction chains are long (confidence degrades gracefully)

## Use PLN when:
- Node priors can be reliably estimated from data
- Domain has rich statistical grounding
- More accurate strength estimates justify the infrastructure cost
- Single-step or short chains (confidence inflation less severe)

## Hybrid approach:
- Use PLN strength formula (prior-adjusted) with NAL confidence formula (conservative)
- Gets prior-adjusted strength without confidence inflation
- Revision is identical in both systems, no choice needed

## Root cause of PLN confidence inflation:
- c2w(0.9)=9, so two 0.9c premises give w=81, w2c(81)=0.988
- Multiplicative weight space explodes before w2c re-compresses
- NAL f*c product stays conservative: 0.8*0.9*0.7*0.9=0.4536

## Critical insight: NAL f*c product dual purpose
1. Anti-inflation: derived confidence never exceeds evidence chain strength
2. Correct silence on dead chains: if A-not-B (f=0, high c) and B-is-C (strong), the zero confidence result means NO EVIDENCE about A-C, not negative evidence. The chain is dead and contributes nothing. PLN w2c(w1*w2) would incorrectly preserve confidence on a broken chain.

## Key distinction:
- Zero confidence = no evidence (correct: chain is dead)
- Zero frequency with high confidence = strong negative evidence (would be WRONG here)
- A rock is not a bird, birds can fly: we have NO evidence about rocks flying via the bird chain
- Rocks cant fly for OTHER reasons, not because of the bird inference path