PLN Skill Library: Reusable Inference Patterns

Transitive inheritance with confidence decay: c_out = c1 * c2 * ... * cn

(|- ((--> A B) (stv 0.9 0.9))
    ((--> B C) (stv 0.85 0.9)))

Result: A-->C with f=f1*f2, c=f1*f2*c1*c2. Confidence decays geometrically - correct epistemic humility.

Use when: chaining known relationships, risk assessment, capability modeling.

Infer shared traits from observed behavior:

(|~ ((Implication (Inheritance  (IntSet Feathered))
     (Inheritance  Bird)) (stv 1.0 0.9))
    ((Inheritance Pingu (IntSet Feathered)) (stv 1.0 0.9)))

Use when: self-model building, diagnosing causes from symptoms, competitive analysis.

Merge independent evidence streams on same claim:

(|- ((--> X Y) (stv 0.8 0.5))
    ((--> X Y) (stv 0.6 0.7)))

Use when: multiple sources report on same fact. Boosts confidence when evidence agrees.

Encode qualitative assessments as NAL, run deduction, compare outputs to prose judgment.

Validated: ICP insufficient stv0.72, Bittensor trustworthy stv0.15, NEAR not-AI-innovation stv0.60, Sentient valuable-innovation stv0.56. Rankings matched qualitative analysis.

NAL abduction IS backward chaining for simple cases. Chain repeatedly to trace goals back to preconditions:

achieves_mastery --> solves_problems (0.85, 0.351)
solves_problems --> gathers_resources (0.80, 0.202)

Confidence decay through abduction is steeper than deduction - correct because reasoning backward from effects to causes is inherently less certain.

Use when: goal decomposition, root cause analysis, planning backward from desired outcomes.

Encode disbelief with stv 0.0:

(|- ((--> X trustworthy) (stv 0.8 0.5))
    ((--> X trustworthy) (stv 0.0 0.7)))

Revision merges positive and negative evidence. High-confidence negative evidence dominates low-confidence positive.

Use when: integrating contradictory reports, stress-testing claims, adversarial analysis.

Experiment: 3 options, 3 rounds of revision with varying evidence quality.

Round 1: A=0.84/0.588  B=0.33/0.588  C=0.50/0.200
Round 2: A=0.666/0.745 B=0.715/0.790 C=0.50/0.404  (B overtakes A!)
Round 3: A=0.830/0.874 B=0.670/0.826 C=0.293/0.685  (A recovers!)

Key findings:

• One strong high-confidence evidence round can flip preference rankings
• Reversals are NOT sticky - equally strong counter-evidence restores original order
• Confidence accumulates monotonically even as frequency oscillates
• NAL revision has no momentum effect - always responsive to new signals

Use when: tracking evolving preferences, A/B testing with Bayesian updates, any decision where new evidence should rationally override prior beliefs.

Domain: sleep_deprivation causal cascade, 4-step chain.

Step 1: sleep_dep --> elevated_cortisol     f=0.900  c=0.850
Step 2: sleep_dep --> impaired_judgment      f=0.720  c=0.490
Step 3: sleep_dep --> poor_decisions          f=0.612  c=0.270
Step 4: sleep_dep --> relationship_damage     f=0.428  c=0.087

Confidence decay accelerates: each step loses more than the last. By step 4, confidence is below the 0.15 corroboration threshold - flag for empirical validation before acting on conclusion.

Temporal encoding: Use product terms (x event_A event_B) with implication to derive precedence within standard NAL.

Use when: risk assessment, causal modeling, deciding when a chain is too long to trust without independent evidence.