Evaluation ledger

Measured evidence only.

This page distinguishes runnable invariants from future measurements. Targets do not become results until their cases, command, environment, and limitations are committed.

Rule outcome benchmarks use synthetic cases. The published daily quote artifact is separate normalization and refusal evidence, with provenance shown in working mode.

Implemented

Published quote derivation and normalization

Reproduce committed response-derived bytes and normalize the daily artifact with invariant hashes; reject an untrusted actor case before evaluation. This is not a rule benchmark.

Implemented

Row-order invariance

Shuffle the same events; expect one canonical hash.

Implemented

Literal golden hash

Pin a committed fixture to its literal canonical result hash.

Implemented

Exact duplicate tolerance

Insert an identical source row; expect unchanged canonical events.

Implemented

Identity-conflict rejection

Reject conflicting reuse of an event or source identity.

Implemented

Time-format equivalence

Normalize equivalent offset and Z timestamps to the same instant.

Implemented

Sub-millisecond order

Preserve supported precision and reject timestamps beyond it.

Implemented

Locale-independent order

Order canonical keys by UTF-16 code units without locale data.

Implemented

Volatile-metadata exclusion

Exclude collection metadata from the canonical result hash.

Implemented

Mixed-sequence policy

Fail closed on mixed sequence presence and define the all-absent order.

Implemented

Dialect convergence

Replay equivalent committed source dialects to one canonical result.

Implemented

Dataset-profile determinism

Keep dataset profiles stable across shuffling and source dialects.

Implemented

Mapping-approval binding

Bind approval to the validated proposal and its executed transforms.

Implemented

Record-set completeness

Reject omitted declared rows or approved columns before result hashing.

Implemented

Mapping agreement reporting

Report per-field agreement between mapped canonical events and each mapping application's review outcome.

Implemented

Reachable mapping review

Require a justified override for a flagged field while preserving a fully resolvable path.

Implemented

Scenario classification

Pin three synthetic cases to their declared rule outcomes.

Planned

Evidence completeness

Resolve each finding through eventId to rawRowHash.