@indexnetwork/protocol 5.2.2-rc.370.1 → 5.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +1 -0
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -13,6 +13,7 @@ See [STABILITY.md](./STABILITY.md) for the public-contract and tier definitions.
|
|
|
13
13
|
## [Unreleased]
|
|
14
14
|
|
|
15
15
|
### Added
|
|
16
|
+
- Enforced eval baseline compatibility and auditable update governance: exact comparability assessment over harness/schema version, model and judge IDs, selection/full-corpus status, corpus and scoring-config fingerprints, run protocol, and completeness — provably incompatible cohorts are never compared (exit `2`), strict-mode unprovable comparability fails closed (exit `3`), committed schema-v1 baselines keep comparing under the normal policy with explicit notes, `--update-baseline` now requires `--reason` plus a complete full-corpus unfiltered run at a clean identifiable Git revision, every update persists a deterministic reviewable `*.baseline.update.json` provenance/diff summary through the overwrite-safe artifact path, added/removed/skipped cases are reported explicitly, and rolling baselines aggregate only compatible complete full-corpus reports while reporting every excluded artifact with its reason; the beta-binomial comparison and Wilson intervals are unchanged (IND-445).
|
|
16
17
|
- Provider-free privacy-aware eval artifact viewer with explicit shared v1/v2 and HyDE-public adapters, allowlisted redaction, attempt-aware execution inspection, baseline deltas, accessible offline navigation, safe failure pages, and atomic read-only output (IND-446).
|
|
17
18
|
- Default-off Lens B outcome-question shadow: pure, outcome-blind trade-off hypothesis mining over a user's OWN explicit opportunity decisions, with one unique counterpart per captured opportunity, recipient-scoped counterpart deduplication, run-local candidate aliases (raw opportunity ids are never sent to the LLM), trimmed/unique/non-empty compared sides, conflicting classifier assignments excluded from support, at least five genuinely distinct independent examples per side, small-cell suppression, aggregate-only telemetry, and an `OutcomeOutbox` contract enabling transaction-held scope revalidation plus atomic same-transaction outcome capture in the winning owner-action transition (IND-434).
|
|
18
19
|
- Default-off Lens C negotiation-evidence shadow mining from future negotiation tasks with immutable intent snapshots, exact task-linked allowlisted evidence, strict participant/source verification, recurrence across at least five distinct opportunities, and aggregate-only telemetry (IND-433).
|