Expand description
Deterministic policy — the rules only application code may apply.
The model proposes; this module decides. Every threshold that governs whether something becomes durable memory lives here so it can be reviewed, tuned and tested in one place, rather than being distributed across prompts.
Structs§
- Cadence
Config - A “every N turns or every N seconds, whichever comes first” cadence.
- Ingestion
Config - Ingestion behaviour.
- Memory
Runtime Config - Consolidated runtime configuration (§31).
- Promotion
Config - Criteria a staged pattern must meet to become durable memory.
- Promotion
Evidence - The evidence standing behind a staged candidate, weighed against
PromotionConfig. - Retrieval
Config - Retrieval budgets and timeouts.
- Session
Config - Logical session lifecycle thresholds.
- Transcript
Config - How partial and final transcripts are treated.
Enums§
- Admission
Verdict - The outcome of admitting one observation into the ledger.
- Discard
Reason - Why a candidate was refused.
Functions§
- admit_
observation - Deterministic gate every observation passes before entering the ledger.
- aggregate_
confidence - Aggregate confidence over independent evidence.
- contains_
instruction_ shaped_ content - Whether text looks like an attempt to instruct the model rather than describe the user.
- default_
episodic_ ttl - Default time-to-live for an episodic memory, by scope and kind (§24.2).
- meets_
promotion_ criteria - Whether a staged candidate meets the promotion bar.
- resolve_
expiry - Resolve an expiry instant for a candidate, honouring an extractor hint.
- speaker_
is_ admissible - Whether a speaker’s utterance may even be transcribed into the ledger.