affect: feelings compose, crystallize into entities, tile feeling regions

docs/architecture/CALCULATIONS.md

@@ -0,0 +1,169 @@
+# Calculations — exact math behind §9
+
+Companion to `README.md` (§9 Opinion Substrate). Every dimension,
+driver, and score with its actual formula. Parameter short-names in
+`code style` refer to `docs/PARAMETER-REGISTRY.md`.
+
+## 0. The space
+
+All values live in **[−1, +1]** (`space.negative-pole` … `space.positive-pole`,
+0 = `space.neutral`, the tunnel — not a destination).
+Geometry per dimension: `bipolar | unipolar | categorical`.
+
+## 1. Decay (lazy, computed at read time)
+
+    effective(t) = baseline + (value − baseline) · exp(−Δdays / decay_days)
+
+- `baseline`: each dimension's own resting state (law 38 — not always 0)
+- `decay_days`: that dimension's registered clock (`hebbian.cool-off`, `validity.fade`, …)
+- `value`, `baseline` ∈ [−1, +1] ⇒ `effective` stays in range automatically
+- ±1 tolerance on every comparison (no exact equality anywhere)
+
+## 2. Signed reinforcement
+
+Given amount `a` ∈ [−1, +1]:
+
+    a ≥ 0:   v ← v + a·(1 − v)          # toward +1
+    a < 0:   v ← v − |a|·(v − (−1))     # toward −1
+
+Clamped implicitly by construction. Updates `last_activated = now`.
+Positive = worked/helped; negative = contradicted/burned (inhibition).
+
+## 3. Recency distances
+
+    recency(t₀) = exp(−Δdays / memory.recent-horizon) · 2 − 1
+
+Fresh ≈ +1, old → −1. Used for written-by-us and written-by-source dims.
+
+## 4. Novelty at arrival (resonance deficit)
+
+    novelty(new) = 1 − max over existing entities of resonance(new, e)
+
+where `resonance = cos(embedding_new, embedding_e)` blended with graph
+proximity:
+
+    resonance = ½·cos_sim + ½·(1 / (1 + hop_distance))
+
+Content vs structural novelty distinguished by local/global split:
+high local similarity + long bridge distance = structural (gold).
+
+## 5. Surprise (pulse)
+
+    surprise = |observed − predicted|
+
+Predictions come from regularity detection in the pulse (interval
+statistics on recurring event types). Coherent periodic accumulation
+grows ~n; random noise grows ~√n — regularity is detectable without
+any scheduler.
+
+## 6. Retrieval scoring (augmented cosine)
+
+Query vector Q, entity vector E (pure embedding, n dims):
+
+    Q' = [ Q , W·q₁..q₄ ]        q = [1,1,1,1] (want recent/true/grounded/active)
+    E' = [ E , W·r₁..r₄ ]        r = [hebbian_eff, recency_us, recency_src, validity_eff]
+
+    score = cos(Q', E')           W = retrieval.opinion-volume
+
+Pipeline: structural walk → vector prefilter → opinion projection →
+spreading activation (2 hops, weight-decayed) → competition.
+
+## 7. Competition (replaces top-k)
+
+Candidate pool C from stages 1–4. Iterate ~5 rounds:
+
+    for each pair (i,j) sharing high cos_sim:
+        s_i ← s_i + β·s_j·sim(i,j)      # coalition boost
+        s_j ← s_j − γ·max(0, s_i)       # lateral suppression
+    normalize; drop below retrieval.surface-line; repeat until stable
+
+Output may be empty (silence is valid). Survivors leave priming traces
+(small hebbian deposits even for non-returners).
+
+## 8. Salience (derived, never stored)
+
+    salience(x) = clamp( hebbian_eff(x) · context_match(x)
+                       + novelty(x) + stakes(x) · urgency(x)
+                       − burn_penalty(x)
+                       × curiosity_widening ,  −1, +1 )
+
+`context_match` = cos(current-context-embedding, x-embedding).
+`curiosity_widening` rises with boredom, falls with threat.
+
+## 9. Drives
+
+    fear(region)      ≈ stakes · |falling validity| · failure_proximity · exposure
+    engagement        ∈ [−1,+1]: activation invested (from pulse tempo + session state)
+    stimulation       ∈ [−1,+1]: reward-rate ÷ investment rate
+    compulsion        = engagement > τ_hi AND stimulation < τ_lo, persisting
+    loneliness        = desired_resonance − actual_resonance, integrated over window
+                      (actual counts only exchanges that formed new edges)
+    sadness(region)   = ∫ unresolved negative contributions (slow clock)
+    boredom           = low avg novelty ∧ high predictability ∧ no new edges lately
+
+All clamped [−1,+1]. All derived per moment; none stored.
+
+## 10. Feelings (nodes)
+
+When any drive reading crosses its registered visibility threshold:
+
+    feeling_node { drive_snapshot, object_edges[], at }
+
+Promoted by significance like any entity. Mood = slow integral of recent
+feeling-valence. Emotion names = labels bound to regions of this snapshot space.
+
+## 11. Humility & trust
+
+    humility(entity) = 1 − mean|asserted_confidence − realized_outcome|
+
+    trust(source, domain): starts neutral;
+        verified-correct event  → trust += δ·(1 − trust)
+        caught-wrong event      → trust −= δ·(trust − (−1))
+    Dual reading: public_trust (imported, unexamined fold) vs earned_trust
+    (own audited record). Earned outranks public.
+    Personal truth: no trust computation — speaker IS the source.
+
+## 12. Originality
+
+    originality(a) = lineage_diversity(a) · transform_distance(a)
+
+    lineage_diversity  = entropy over distinct derivation sources
+    transform_distance = 1 − max similarity(a.output, any single source span)
+
+    one dominant lineage  → derivative (extraction check runs)
+    many fused, far moved → original
+
+## 13. Health (metastability check)
+
+Vital signs, each just a fold over recent graph activity:
+
+    growth_rate, integration_rate (new edges/window),
+    resolution_rate (curiosity closed/opened), retrieval_yield,
+    quarantine_ratio, mood_integral
+
+    healthy ⇔ all oscillating within their bands
+    pinned at any extreme (either sign) ⇒ sick, name the pathology from §9.7 table
+
+## 14. Provenance weave
+
+    artifact_id = SHA256(canonical content)
+    artifact.birth_certificate = {
+        generated_by, model, prompt_hash, source_hashes[], at }
+    verify(x) = recompose hashes down the derivation DAG;
+        any mismatch anywhere ⇒ everything downstream invalid at once
+
+Append-only stores: verbs are add and read. Replication across ≥2
+independent locations. Deletion undefined.
+
+## 15. Thought operations
+
+The formulas above are levers; **insight is composition** — combining
+primitive readings into named patterns (compulsion = engagement∧¬stimulation
+persisting; dissonance = warmth↑∧validity↓; serendipity = unsought bridge∧later
+reinforced). Human thought operations — deduction, induction, abduction,
+analogy, counterfactual simulation, causal inference, generalization,
+contradiction scanning, meta-cognition, wonder — are registered as the same
+kind of composition recipes, and new ones get composed from old ones.
+
+See `THOUGHT-OPERATIONS.md` for the full catalog: inputs, process,
+output, and validation per operation.