Record hourly 0910 HIGH as a research-only fold

[24601]

Decision changed

No runtime decision changes. The weekend packet does not move family choice, the scoring table, or who owns an act.

Acceptance (Astra), before any SKILL/README/Pages/runtime edit

1. Problem. An agent could treat a weekend of new repositories, a star count, a "beats Jev" line, a catalog of 944, the word calibrated, or a router as a reason to add a scoring row or to change family. It could let a model own a tool call, a shell write, a click, or a send.
2. Baseline. Installed guidance already separates a leaderboard from a deployment evaluation, refuses to treat a training recipe or a proper loss as calibration on this population, keeps the act in code or with a person, and keeps TypeSafe Jev as the default hosted exemplar for the class. A public slice is not Harbor. A wire path is not logit equivalence. A quantization, a port, or a LoRA head is not a new family. Choice still depends on the offered set. A specialist fit on a benchmark's train split is a different comparison from a zero-shot readout. Mappings section 17 stays the last scoring section.
3. Proposed improvement and assumptions. None in runtime files. Archive the sources' own limits where they already match those rules. Assumption: a README or card opening is enough to refuse promotion, and is not a reproduction. A description rewrite whose body was not opened does not reset review age. Stars are not a capability change.
4. Affected surfaces. research/notes.md §175, research/archive/hourly/2026-09-28T15/, fingerprints for bodies that were opened, sources.json pointers, and the hourly logs. Not SKILL.md, not README.md below ## License, not the site, not the package version.
5. Checks. make check (passed: repository quality checks, 197 unit tests, numerical self-test, fingerprint self-test, shell syntax). No Pages build. No behavioral scenario run, because installed guidance did not change.
6. Counterexample that could reject the approach. A held-out run, on labels and traffic this repository controls, where OpenDecider's typed-decisions cell, the Chinese bench, GAYA's in-domain table, Sev's 30-run ablation, or Bobcat's agreement alt changes the baseline placement or the training baseline the trainer installs. A catalog count that changes which family to use would also reject leaving the catalogs research-only. No such count is a design rule in this packet.

Soft judgment is not the veto. Catalog densification without a sharper decision rule stays in the archive.

Evidence and scope

Folded onto main 1c42a812da61d9a334ac7150630d2e4d248df931. Weekend gap since Friday 1510. This fold is §175 only. Scoring rows stay at 17. Archer stays promised_not_landed. Hub weights were not requested, so that status is unchanged and not newly verified.

Packet: 532 design-implication rows (505 novel, 27 description-rewrite revisits) and 81 research-only rows, archived at research/archive/hourly/2026-09-28T15/. Retrieval 2026-09-28T15:40Z. Fifty-seven GitHub README openings and ten Hub card openings were read. prithivMLmods/JEV-9B returned HTTP 401. Eight GitHub description rewrites were not opened, so their review age stays put. The other design rows are metadata triage. The 81 research-only rows were not fetched. Third-party numbers stay theirs. No X calls. No Harbor numbers were invented.

Reported, not reproduced, and not promoted:

• OpenDecider nano and the od1 specialist each report 0.796 on typed-decisions. The od1 card says that model is fine-tuned on the train split. Those cells stay separate and off the scoring table.
• GAYA's banking77 gap is an in-domain distill against zero-shot Laya, by their fairness note.
• The Chinese bench, the 154-message Banking77 run, JevSpan's 73.7 F1, and Bobcat's image alt stay public instruments.
• peira's visible table is a mock adapter. Sev's 30-run note was not checked against its PDF or JSON.
• awesome-jev's 944 is a catalog count. zero-shot-ie-bench's description (38/23) and README opening (53/29) disagree; neither becomes a census.
• Gates, routers, UI tests, and the WeChat fork keep the act in code or with a person. MongLong's jev-gate says savings are not established.

This does not belong in runtime guidance. The existing rules already cover it.

Verification

• make check passes; relevant failures and limits are recorded. Command: make check on this branch after the archive edit. Result: repository quality checks passed, 197 tests OK, decision and fingerprint self-tests OK, shell syntax OK.
• Behavior-changing skill edits have independent scenario evidence. Not applicable: no skill edit.
• Website edits have a Jekyll build and generated-site check. Not applicable: no site edit.
• Installation/metadata changes have a package-discovery smoke check. Not applicable: no package or version change.
• No duplicated research feed, tautological test, or unrun capability claim. Unrun benches stay unrun. invented_signal: false.
• Version changes and historical evidence are handled explicitly. Development version stays 0.8.2-dev. No release. Historical §174 claims are unchanged.

This run does not prove an unattended scheduler.