Concepts
Improve
The evaluation and correction loop behind the API. Not a /v1 host surface.
There is no Improve HTTP route on the local runtime host. Integrate study and receipts at /v1; route misses through the evaluation loop documented here and on Benchmarks.
run misses → mine → route (exclusive) → agent draft → blinded adjudication → freeze → pre-register → score → compare
Exclusive Routing
At mine time a clip feeds exactly one pool: the frozen bench, or the learning stream of glossary entries, rules, and correction pairs. Never both. The route is recorded before adjudication and enforced at clip level.
Memory Gate
Runs only propose. A different actor accepts, rejects, or revokes with a reason. Behavioral rules also require a pre-registered ablation pair on the same frozen pack. The gate has never accepted an entry on weak evidence.
0 entries accepted, 1 legacy entry quarantined
Scoring
One frozen pack, one scored run, and a prepped arm that scored below the cold control. The number is on the benchmarks page because the loop only works if misses are captured, not explained away.
1 pack frozen, 1 run scored, no model judge
Next: Offline Adapters
Today accepted memory conditions runs through retrieval. The declared next step exports adjudicated correction pairs into offline adapter training, under a contract fixed before any adapter exists:
- All learning is offline over journaled receipts. Nothing learns inside a run.
- The correction row schema is already the training row shape; provenance-clean pairs come only from the learning-routed pool, never from bench clips.
- Weight training waits until the adjudicated pair count clears roughly one thousand. Current yield is about five rows per run, so this step is measured in months.
- A trained adapter ships as a versioned, content-addressed policy artifact behind the same proposal-to-decision gate as a rule, accepted only through a pre-registered ablation pair scored on the frozen bench.
contract only: 0 adapters trained, 0 claims without receipts