Module confidence 

Bayesian confidence model for relationship edges.

Uses a Beta-Binomial conjugate prior. The pseudo-counts (alpha, beta) are persisted on the edge, so evidence accumulates: the posterior after fifty corroborations is a different distribution from the posterior after five, and its variance is smaller. The stored confidence field is the posterior mean — a derived value kept in sync with the counts on every write, so read paths can keep scoring on the mean alone.

A new edge (or an edge from a store predating evidence persistence) starts at PRIOR_CONCENTRATION: the mean is preserved and the concentration is honestly low.

Structs

EdgeEvidence

Everything one edge persists about why it is believed: the Beta counts that move confidence, and the coherence counter that must not.

Evidence

The accumulated evidence for one relationship: the pseudo-counts of its Beta posterior.

ProvenanceWeights

Evidence weight of one observation, by provenance class.

Enums

ExtractionContext

How a relationship was established — determines initial confidence prior.

Provenance

The authorship class of an episode, and of every confidence-moving observation drawn from it.

Constants

DECAY_FLOOR

Minimum effective confidence floor — decay never goes below this.

DEFAULT_EVIDENCE_WEIGHT

Evidence weight of a single observation whose provenance is not modelled: one observation, one count.

DEFAULT_HALF_LIFE_DAYS

Default half-life for temporal decay (days). At 90 days without reinforcement, effective confidence halves.

DEFAULT_WEIGHT_EXTERNAL

Default evidence weight of an observation from an independent source.

DEFAULT_WEIGHT_SELF

Default evidence weight of the agent restating itself.

DEFAULT_WEIGHT_USER

Default evidence weight of an observation authored by the human.

PRIOR_CONCENTRATION

Total pseudo-count of the Beta prior an edge starts from. ~10 observations to overwhelm the prior.

Functions

effective_confidence

Compute effective confidence for a relationship, using last_reinforced or valid_from as anchor.

path_confidence

Compound confidence along a multi-hop path.

temporal_decay

Compute effective confidence after temporal decay.