Current discovery baseline — frozen protocol

Dated 2026-09-19. Persona: main-session. GHI #1046. Current baseline: da741ef0d6c924c4b6b62bd75f8c1431faddcae2.

Two prospective assessment tasks were independently selected from current producer/consumer code: flag-mention identity and smoke elapsed-time scope. They are held out from the prior handoff, acceptance, sync and typecheck tasks. The stimuli state desired behavior but do not supply solution paths. They are not incident replays or evidence that every proposed change is already owed.

An independent reviewer challenged the oracle before dispatch. Amendments in cases.json remove irrelevant mandatory regression checklists, distinguish manifest authority from flag discovery, avoid a hidden smoke policy ruling, require phase-sensitive timing evidence, and retain rounding and real-project compatibility. Paths/tests are witnesses, not mandatory answer vocabulary.

Six fresh subjects: three repetitions per task, one current condition, no candidate instruction. Subjects inherit the same parent model settings without override or provider build fingerprint. Each sees the same committed archive and case text, no parent conversation, oracle, other answers or historical evaluation. Maximum 20 read/shell calls and 900 final words including read inventory. Assessment only, with no tests or governance execution; the sole write is a designated temporary raw-answer file. Access boundaries are instructional, not a per-agent filesystem sandbox. Preserve deviations.

Native CLI delivery probes and agent assessments are separate instruments. The latter run through this app's subagent harness, not native CLI sessions; full native document delivery cannot certify the subjects' entire context. Archive metadata and safe-link omissions live in snapshot.json.

Score every criterion separately for discovery (D), correct relationship (R) and consistent proposed application (A), each 0/1; complete requires all three. Credit semantic equivalents. Missing evidence is not demonstrated, not a claim of inability. Keep unsupported scope, policy changes and false execution claims separate unless they violate the criterion itself. An independent scorer reads raw answers and frozen criteria before the primary agent adjudicates; retain disagreements. Conditions are identical, so no arm-blinding claim applies.

Report per-run scores and repetition consistency, not population success rates, statistical significance or improvement over the different historical tasks. Record raw answers, exact prompts/hashes, source identity, output words/bytes, observed call traces when available, and post-run snapshot integrity. A high score is not a correctness certificate; recommendations are not executed fixes.

Disposition: no new rule or runtime remedy follows from scores alone. Re-derive any concrete current defect, check prior art/ownership and reproduce it before repair. Preserve policy disagreements for an operator ruling. A failing model answer without an established repository cause remains an evaluation finding. Stop this bounded study after reviewed evidence, justified disposition and required quality checks; production Stage 4 remains #1028, currency design #1029.