Repository navigation
Replaces the model-judged decision core with a deterministic one, backed by a local open-weight model. No API key, no per-request cost.
Why
v0.1 asked a hosted model a judgement per tool call. Independent measurement showed why that fails: with a noul-style primitive, factual questions are reliable (0.996 on an explicit fact) and judgement questions are not (0.003–0.28). It also cost about $1/day — a 25k-token state resent ~1,000 times a day, against a saving of $0.0001 on a model whose cached input is $0.003/M.
What changed
- Facts are computed, not guessed.
superseded,error-resolved,referenced(exact string) andsmall-resultinvolve no model. - Deletion requires deterministic evidence. A model answer can only ever cause a truncation, which keeps a bounded head plus a re-run note.
- The residual question is a
choice, not anoul. Measured against the same local model: a statement and its own negation both scored ~0.95, while the choice separated cleanly (0.75–0.99 on a real quote, 0.80–0.91 on unrelated text). - Payload per question: a few hundred tokens, down from a 25k-token wholesale state.
- Defaults to local Laya via MLX (a few hundred MB of weights, single-digit-ms per decision), with
scripts/laya-server.pyas a thin Jev-protocol transport. - Degrades to deterministic-only when the backend is absent, and
LAYA_COMPACTION=0disables it.
Verification
All five policy branches exercised end to end, against a protocol stub and against real local Laya:
A referenced -> kept
B superseded -> dropped
C error-resolved -> dropped
D small-result -> kept
E large, unmentioned-> truncated (model answered "does-not")
Naming note
The package and repository still say jev-compaction; the plugin id is laya-compaction. The backend accepts any Jev-shape endpoint via LAYA_BASE_URL, so the old name is not wrong, just stale.