Skip to content

algorithm.md

Latest commit

 

History

History
144 lines (111 loc) · 6.01 KB

algorithm.md

File metadata and controls

144 lines (111 loc) · 6.01 KB
Raw

Algorithm and invariants

This file is the specification: what the kernel computes. For where each constant came from — published work, production tuning, or an arbitrary bounded choice — read foundations.md.

The checked-in fixtures are the behavioral authority. Neither language runtime is the oracle by itself. kernel_golden.json contains 101 affect, 16 retrieval, 69 appraisal, and 39 surfacing cases; continuity_traces.json adds three three-step sequences.

State domains

  • PAD mood and baseline valence: [-1, 1]
  • OCEAN traits, appraisal goals, trust, and attachment values: [0, 1]
  • OCC carry values: [0, 1]
  • relationship stage: stranger, acquaintance, friend, close, or intimate

Only Neuroticism and Extraversion currently affect inertia. Relationship stage affects the resting point, and attachment longing can add a state emotion. The default engine does not evolve personality, goals, relationship, or attachment.

Affect controls

A PAD value is neutral when all three absolute components are below 0.15. Otherwise the signs map to the standard eight labels:

PAD signs Label PAD signs Label
+++ exuberant -++ hostile
++- dependent -+- anxious
+-+ relaxed --+ disdainful
+-- docile --- bored

Sampling temperature is clamp(1 + 0.20 × arousal, 0.72, 1.18) with top_p=0.97. A missing mood produces temperature 1 and no top-p override.

The response-budget factor is 0.60 below arousal -0.30, otherwise 0.72 below valence -0.30, otherwise 1. Applying it can never increase the caller's positive integer budget.

Ambivalence requires the strongest positive and negative emotion to each be at least 0.15, with min/max >= 0.40.

Default conversational appraisal

The included intent and English expectation mapping is a non-habituating reference policy. It is not a universal event ontology. Applications can inject a synchronous appraisal policy while retaining the same transaction and state boundaries.

Relationship stages map to ordinals 1–5. Their resting-point weights are 0, 0.20, 0.40, 0.60, 0.70; unknown labels behave like stranger.

Mood inertia is:

phi = clamp(0.80 + 0.20 × (N - 0.5) - 0.10 × (E - 0.5), 0.62, 0.92)
rest = (stage_weight × baseline_valence, 0, stage_weight × 0.10)
next_mood = round4(clamp(rest + phi × (current_mood - rest), -1, 1))

The reference intent impulses are:

Intent PAD change Fresh emotion signal
ABUSE V-.35, A-.10, D+.25 reproach .95×respect, distress .65×rapport
APOLOGY V+.05 joy .20×rapport, gratitude .35×honesty
VULNERABILITY V+.15, A+.10 gratitude .70×honesty, joy .75×rapport
CURIOSITY V+.20, A+.15, D+.05 admiration .75×intellectual, joy .50×rapport
CHALLENGE V-.05, D+.10 admiration .45×intellectual, distress .25×rapport
NEGLECT V-.10, A-.05 distress .40×rapport
CASUAL V+.02 joy .05

For the ambiguous intents CASUAL, CURIOSITY, CHALLENGE, and APOLOGY, an already-polarized valence is then widened: when |valence| >= 0.20 it is scaled by 1.10 if negative and 1.04 if positive, then clamped and rounded to four decimals. The other intents carry an unambiguous sign of their own and are not amplified.

Slow baseline valence becomes round4(clamp(0.98 × old + 0.02 × appraised_valence, -1, 1)).

Prior OCC signals decay by 0.70 for reproach, 0.80 for distress, 0.85 for admiration, 0.88 for gratitude, 0.90 for joy, and 0.80 otherwise. Values at or below 0.05 are dropped. Fresh, carried, state-derived, and expectation-derived signals merge by maximum.

Retrieval scoring

The scorer uses cosine distance in [0, 2]:

similarity = 1 - distance / 2
age_days = max(0, now - timestamp)
recency = clamp(1 - age_days / 60, 0.40, 1)
salience = 1 + 0.03 × clamp(significance, 0, 1)
             + min(0.025, 0.006 × ln(1 + recall_count))
score = similarity × recency × salience + (0.05 when episode else 0)

Future timestamps receive maximum freshness. Missing, invalid, or unusable timestamps receive the 0.70 fallback. Candidate timestamps must be zoned ISO values at adapter boundaries.

Mood congruence applies only when enabled and absolute mood valence is at least 0.20. A same-sign memory receives 1.06 in a negative mood or 1.03 in a positive mood. Ranking deduplicates by ID, retains the best duplicate, then sorts by descending score and ascending Unicode code-point ID.

Presence surfacing

Presence is a pure state-to-payload transform. The compact line uses this fixed priority:

reflection pending → still reflecting
due intention      → holding a follow-up
carried thought    → carrying a thread
open thread        → holding a pattern
otherwise          → here with you

External cognition is injected. Numeric display values are rounded to four decimals, and carried-thought presence is derived from cleaned state rather than trusted from a caller flag.

Engine and trust invariants

  • Responding turns run gate → optional retrieval → ranking → appraisal policy → response controls → generation → one atomic commit.
  • Silent turns persist the user message without appraising or changing state.
  • Gate errors and malformed outputs propagate unless a caller explicitly opts into a respond/silent fallback.
  • Intents are trimmed, uppercased, and allowlisted. Custom labels must match [A-Z][A-Z0-9_]{0,63}.
  • A store-owned transaction serializes one conversation across engine instances. Any pre-commit failure leaves state and transcript unchanged.
  • Raw memory and carried-thought text never enters the default system prompt. It is bounded as untrusted evidence—2,000 characters per item and 8,000 total by default—and adapters must preserve that lower-trust channel on the wire.

Prompt prose, model output, embeddings, custom policies, application safety, reflection, and language-level API shapes are outside cross-runtime parity.