Append-only registry of every scored corpus slice. A slice MUST be registered
here, with its date, before its first scored run (the register-before-run rule
of PROTOCOL.md L5). Every release publishes the delta for all registered and
pinned slices — never a chosen subset. This file is the salience guard that code
cannot enforce; its git history is the audit trail.
Entries are append-only: to correct one, add a new row that supersedes it; never
edit a past row in place. CI (tests/test_protocol.py) fails if any shipped
datasets/*.jsonl slice is missing from this registry.
| slice | dataset file | vertical | registered | naive-weak (recall < 0.5)? | role | status |
|---|---|---|---|---|---|---|
| authz_v1 | datasets/authz_v1.jsonl |
authz | 2026-06-03 | NO — naive recall 0.833 | INVERSE-ANCHOR (commoditized; the method must at least not regress) | pinned (datasets/baseline_authz_v1_2026-06-03.json) |
- decode_v1 (
vertical: decode) — the moat region: naive-weak (naive recall ≈ 0 on the M-01 anchor, where the method scored 2/2). There is NO shippeddatasets/decode_v1.jsonlyet. Admit it under the L4 bar once it ships its naive baseline in the same commit; register it here before its first scored run. Admission also requiressample_n >= MIN_SLICE_N(=10) naive-scored positive rows (L4 condition d). The current M-01 anchor is a single discordant pair, so adecode_v1SEED is UNDERPOWERED (n=1): it stays "Planned" here and is NOT admitted —admission.pymechanically rejects it (fail-safe) until the sample reaches the floor. The n=1 seed is a suggestive datapoint, not a measured rate.
authz_v1is retained as a deliberate inverse-anchor: a naive single call already aces it (recall 0.833), so under the L4 admission bar (naive recall < 0.5) it would NOT be admitted as a new slice today. It stays as the calibration / negative reference — the place the method must at least not regress — not as a growth target. Growth goes to naive-weak (decode-like) classes.