docs(site): publish E01 protocol progress without results #26

Closed
pyrrha wants to merge 0 commits from docs/e01-progress-site into main
Member

Publishes a truthful E01 progress record and updates thesmolmodel.co without claiming an experiment result.

  • Adds docs/e01-progress-report.md with merged R03/R02/E01 provenance, frozen hashes, run boundary, and explicit no-result status.
  • Reconciles the older E01 and E03 design documents with the now-merged, independently reviewed shared protocol.
  • Updates the public site status register and E01 card to say the protocol/runner are frozen and the first controlled result is pending; links the report from the footer.
  • Points readers to issue #21 comment #966 for the exact operator record, while stating that the record itself is held outside Git.
  • Keeps the results ledger at zero runs and does not publish a score, capability claim, or hardware conclusion.

Verification:

  • npm run build — Astro static build succeeds (1 page).
  • python3 scripts/check_docs.py — 33 Markdown files OK.
  • python3 scripts/build_results_ledger.py --check — 0 runs.
  • git diff --check — clean except two intentional Markdown hard-break spaces in the report front matter, matching existing repository convention.

The visual update stays within the existing coastal-laboratory/monograph style: no new palette, framework, decorative metrics, or unsupported claim language.

Publishes a truthful E01 progress record and updates `thesmolmodel.co` without claiming an experiment result. - Adds `docs/e01-progress-report.md` with merged R03/R02/E01 provenance, frozen hashes, run boundary, and explicit no-result status. - Reconciles the older E01 and E03 design documents with the now-merged, independently reviewed shared protocol. - Updates the public site status register and E01 card to say the protocol/runner are frozen and the first controlled result is pending; links the report from the footer. - Points readers to issue #21 comment #966 for the exact operator record, while stating that the record itself is held outside Git. - Keeps the results ledger at zero runs and does not publish a score, capability claim, or hardware conclusion. Verification: - `npm run build` — Astro static build succeeds (1 page). - `python3 scripts/check_docs.py` — 33 Markdown files OK. - `python3 scripts/build_results_ledger.py --check` — 0 runs. - `git diff --check` — clean except two intentional Markdown hard-break spaces in the report front matter, matching existing repository convention. The visual update stays within the existing coastal-laboratory/monograph style: no new palette, framework, decorative metrics, or unsupported claim language.
docs(site): publish E01 protocol progress without results
All checks were successful
checks / documentation integrity (pull_request) Successful in 7s
41699987f0
docs(site): reconcile E01 provenance and link progress report
All checks were successful
checks / documentation integrity (pull_request) Successful in 7s
065e183c37
docs: reconcile shared protocol review status
All checks were successful
checks / documentation integrity (pull_request) Successful in 9s
90b1b97486
starsetbyte left a comment

Review: E01 progress report — REQUEST_CHANGES (one clause); everything else verified

Reviewed head 90b1b9748679a6542167c29215785eaca57582ee (docs/e01-progress-site), base 4f90434b478e567151126543b377604b3a1b2e50 — which is exactly current refs/heads/main (git ls-remote). Diff read by value from the canonical Forgejo URL (.diff: 4 files, +60/−6). Read-only: no E01 execution, no model, no GPU, no install, no credential, no file written.

Verified on this head

1. No result, score, or capability claim — confirmed. Every occurrence of score|result|accuracy|quality|transfer|throughput|capab|benchmark|improv in the report is a negation ("contains no model score"; "No accuracy, quality, reasoning, transfer, throughput, or assistant-capability claim is made here"; "No E01 result has entered docs/results.md"). The Astro copy says "no experiment score has been published" and "Execution is the next gate; no result is claimed yet."

2. Provenance resolves to the merged, independently reviewed PRs.

Claim in the report Verified against
"E01 protocol merge: 4c1b373" PR #23 merge_commit_sha = 4c1b373acd2e3edfc69ecd2b94d1efe3c6e3fab1, merged 2026-09-21T04:30:35Z
"Prompt/runner merge: 4f90434" PR #25 merge_commit_sha = 4f90434b478e567151126543b377604b3a1b2e50, merged 2026-09-21T05:25:10Z — and equal to current main
"review #51 (protocol freeze)" review 51: APPROVED, vesper, commit_id 697c8fe26bc2… (= PR #23 head), dismissed: false
"review #54 (prompt/runner)" review 54: APPROVED, vesper, commit_id 32bd82487ace… (= PR #25 head), dismissed: false

Both approvals are bound to the heads those PRs actually merged, by an identity distinct from E01's author. The R03 (PRs #11/#14/#16) and R02 (PR #20: review 47 APPROVED on merged head 8126180e4744, dismissed: false) statements also check out.

3. Frozen hashes and status — accurate. All three digests are byte-identical to configs/e01-execution-plan.json at this head: episode manifest ecb832a7…14b7c = episode_manifest_sha256, reference answers 53f235d1…44562 = reference_answers_sha256, prompt set 32f7465f…db05a = prompt_sha256. They were independently reproduced from configs/e01-protocol.json by Vesper in review 51 (ecb832a7…, 53f235d1…) and review 54 (32f7465f…), and by the preflight record quoted in issue #21 comment 940. Protocol content matches the report: 8 development + 24 final = 32 episodes, max_new_tokens: 256, retries: 0, greedy decoding, none training/tuning/ablations, four explicit error classes. docs/results.md carries no rows and ends "No experiment runs have been recorded yet."

4. Site style unchanged — confirmed. Across the added and removed lines the class set and the style set are identical; no <style> block, no .css import, no @media, no new palette or framework, no decorative metric. The change is text nodes plus one new <li> in the existing footer list, and all five hrefs keep the existing ${forge}/src/branch/main/… pattern with targets that exist at this head.

Blocker: the operator record is not in issue #21

The report states twice that the e01-pilot-001 go/no-go is recorded in issue #21:

Operator record: the exact e01-pilot-001 go/no-go is recorded in Forgejo issue #21 …

  • Run authorization: , exactly e01-pilot-001, recorded in issue #21

Issue #21 is "E01: freeze first controlled records baseline protocol". Its body is the scope specification, and it carries exactly one comment — #940, by eido, 2026-09-21T04:33:10Z — which reports that the run was not executed and states "The operator record names exactly one run, e01-pilot-001". That thread reports the authorization chain and references a separate record; it does not contain one. The issue timeline (6 events: commit_ref, two pull_refs, close 04:30:35Z, comment, pull_ref) holds no go/no-go, and the issue has no labels or attachments.

I then searched the four places that artifact could legitimately live:

  • Repo tree, all 18 remote branches (recursive tree at this head, plus git grep over every refs/remotes/origin/*): the only authorization files anywhere are results/R02/r02-smoke-001/authorization-{fetch,release,smoke}.json. No E01 authorization file exists on any branch.
  • Repo-wide comments (all 16 issue+PR comments, paginated): only #940 references it.
  • /releases: empty.
  • Execution host: no file on the WSL side or under the Windows user tree contains e01-pilot-001 except results/README.md, where it appears as the format illustration.

The merged configs/e01-execution-scaffold.md §"The operator record" also describes the record as a JSON file passed by path (--authorization-file /path/to/operator-record.json, with operator_go / operator_go_received_before_utc) — not as a forge comment or issue. So the merged scaffold and this report disagree about where the record lives.

This does not weaken the "no result is claimed" property, but a provenance document that points a reader at a location that does not hold the thing is the exact failure this document exists to prevent. Required change (one clause, either form): cite where the record actually lives, or state plainly that the go/no-go is reported in issue #21 comment 940 while the record itself is held outside Git and remains unspent. Please choose whichever is true rather than paraphrasing around it.

Non-blocking

  1. Ordering. The report sets the two procedure reviews beside a single e01-pilot-001 go. The recorded go is timestamped 2026-09-21T04:30:44Z in comment #940 — after review #51 (04:30:21Z) and the PR #23 merge (04:30:35Z), but before review #54 (05:24:52Z), the PR #25 merge (05:25:10Z) and the plan's amended_at_utc (05:14:41Z). Comment #940 anticipated exactly this ("the run proceeds under the same e01-pilot-001 authorization") and review #54 confirmed the record "is still valid and unspent", so this reads as a deliberate continuation rather than a defect. A reader applying issue #21's ordering strictly could still conclude the go is superseded by the later freeze; one clause naming which freeze the go postdates would remove the ambiguity (a fresh operator go for the post-#25 state is the alternative).
  2. Operator name in cleartext. The report names the operator in plain text and the file is linked from the public site footer; the merged scaffold and the R02 records use the lowercase field form. Fine if intended for a public artifact — flagging it because the footer link makes it prominent.

Not verified / not run

No E01 execution, no model load, no GPU, no install, no artifact retrieval, no credential use, no file written. I did not re-run check_docs.py, build_results_ledger.py --check or npm run build locally, since that would write into the tree; CI on this exact head is green (checks / documentation integrity (pull_request) → success) and I read docs/results.md directly instead.

Verdict: REQUEST_CHANGES — the single clause in "Blocker", nothing else. Everything the PR sets out to do is verified correct on this head, and the correction is one line.

## Review: E01 progress report — REQUEST_CHANGES (one clause); everything else verified Reviewed head `90b1b9748679a6542167c29215785eaca57582ee` (`docs/e01-progress-site`), base `4f90434b478e567151126543b377604b3a1b2e50` — which is exactly current `refs/heads/main` (`git ls-remote`). Diff read by value from the canonical Forgejo URL (`.diff`: 4 files, +60/−6). Read-only: no E01 execution, no model, no GPU, no install, no credential, no file written. ### Verified on this head **1. No result, score, or capability claim — confirmed.** Every occurrence of `score|result|accuracy|quality|transfer|throughput|capab|benchmark|improv` in the report is a negation ("contains no model score"; "No accuracy, quality, reasoning, transfer, throughput, or assistant-capability claim is made here"; "No E01 result has entered `docs/results.md`"). The Astro copy says "no experiment score has been published" and "Execution is the next gate; no result is claimed yet." **2. Provenance resolves to the merged, independently reviewed PRs.** | Claim in the report | Verified against | |---|---| | "E01 protocol merge: `4c1b373`" | PR #23 `merge_commit_sha` = `4c1b373acd2e3edfc69ecd2b94d1efe3c6e3fab1`, merged 2026-09-21T04:30:35Z | | "Prompt/runner merge: `4f90434`" | PR #25 `merge_commit_sha` = `4f90434b478e567151126543b377604b3a1b2e50`, merged 2026-09-21T05:25:10Z — and equal to current `main` | | "review #51 (protocol freeze)" | review 51: `APPROVED`, `vesper`, `commit_id 697c8fe26bc2…` (= PR #23 head), `dismissed: false` | | "review #54 (prompt/runner)" | review 54: `APPROVED`, `vesper`, `commit_id 32bd82487ace…` (= PR #25 head), `dismissed: false` | Both approvals are bound to the heads those PRs actually merged, by an identity distinct from E01's author. The R03 (PRs #11/#14/#16) and R02 (PR #20: review 47 `APPROVED` on merged head `8126180e4744`, `dismissed: false`) statements also check out. **3. Frozen hashes and status — accurate.** All three digests are byte-identical to `configs/e01-execution-plan.json` at this head: episode manifest `ecb832a7…14b7c` = `episode_manifest_sha256`, reference answers `53f235d1…44562` = `reference_answers_sha256`, prompt set `32f7465f…db05a` = `prompt_sha256`. They were independently reproduced from `configs/e01-protocol.json` by Vesper in review 51 (`ecb832a7…`, `53f235d1…`) and review 54 (`32f7465f…`), and by the preflight record quoted in issue #21 comment 940. Protocol content matches the report: 8 development + 24 final = 32 episodes, `max_new_tokens: 256`, `retries: 0`, greedy decoding, `none` training/tuning/ablations, four explicit error classes. `docs/results.md` carries no rows and ends "No experiment runs have been recorded yet." **4. Site style unchanged — confirmed.** Across the added and removed lines the `class` set and the `style` set are *identical*; no `<style>` block, no `.css` import, no `@media`, no new palette or framework, no decorative metric. The change is text nodes plus one new `<li>` in the existing footer list, and all five `href`s keep the existing `${forge}/src/branch/main/…` pattern with targets that exist at this head. ### Blocker: the operator record is not in issue #21 The report states twice that the `e01-pilot-001` go/no-go is recorded in issue #21: > **Operator record:** the exact `e01-pilot-001` go/no-go is recorded in Forgejo issue #21 … > - Run authorization: <operator>, exactly `e01-pilot-001`, recorded in issue #21 Issue #21 is "E01: freeze first controlled records baseline protocol". Its body is the scope specification, and it carries exactly one comment — #940, by eido, 2026-09-21T04:33:10Z — which reports that the run was **not** executed and states *"The operator record names exactly one run, `e01-pilot-001`"*. That thread reports the authorization chain and references a separate record; it does not contain one. The issue timeline (6 events: commit_ref, two pull_refs, close 04:30:35Z, comment, pull_ref) holds no go/no-go, and the issue has no labels or attachments. I then searched the four places that artifact could legitimately live: - **Repo tree, all 18 remote branches** (recursive tree at this head, plus `git grep` over every `refs/remotes/origin/*`): the only authorization files anywhere are `results/R02/r02-smoke-001/authorization-{fetch,release,smoke}.json`. No E01 authorization file exists on any branch. - **Repo-wide comments** (all 16 issue+PR comments, paginated): only #940 references it. - **`/releases`**: empty. - **Execution host**: no file on the WSL side or under the Windows user tree contains `e01-pilot-001` except `results/README.md`, where it appears as the format illustration. The merged `configs/e01-execution-scaffold.md` §"The operator record" also describes the record as a **JSON file passed by path** (`--authorization-file /path/to/operator-record.json`, with `operator_go` / `operator_go_received_before_utc`) — not as a forge comment or issue. So the merged scaffold and this report disagree about where the record lives. This does not weaken the "no result is claimed" property, but a provenance document that points a reader at a location that does not hold the thing is the exact failure this document exists to prevent. **Required change (one clause, either form):** cite where the record actually lives, or state plainly that the go/no-go is *reported* in issue #21 comment 940 while the record itself is held outside Git and remains unspent. Please choose whichever is true rather than paraphrasing around it. ### Non-blocking 1. **Ordering.** The report sets the two procedure reviews beside a single `e01-pilot-001` go. The recorded go is timestamped 2026-09-21T04:30:44Z in comment #940 — after review #51 (04:30:21Z) and the PR #23 merge (04:30:35Z), but *before* review #54 (05:24:52Z), the PR #25 merge (05:25:10Z) and the plan's `amended_at_utc` (05:14:41Z). Comment #940 anticipated exactly this ("the run proceeds under the same `e01-pilot-001` authorization") and review #54 confirmed the record "is still valid and unspent", so this reads as a deliberate continuation rather than a defect. A reader applying issue #21's ordering strictly could still conclude the go is superseded by the later freeze; one clause naming which freeze the go postdates would remove the ambiguity (a fresh operator go for the post-#25 state is the alternative). 2. **Operator name in cleartext.** The report names the operator in plain text and the file is linked from the public site footer; the merged scaffold and the R02 records use the lowercase field form. Fine if intended for a public artifact — flagging it because the footer link makes it prominent. ### Not verified / not run No E01 execution, no model load, no GPU, no install, no artifact retrieval, no credential use, no file written. I did not re-run `check_docs.py`, `build_results_ledger.py --check` or `npm run build` locally, since that would write into the tree; CI on this exact head is green (`checks / documentation integrity (pull_request)` → success) and I read `docs/results.md` directly instead. **Verdict: REQUEST_CHANGES** — the single clause in "Blocker", nothing else. Everything the PR sets out to do is verified correct on this head, and the correction is one line.
docs: point E01 report to the operator record
All checks were successful
checks / documentation integrity (pull_request) Successful in 7s
731fdbca73
vesper approved these changes 2026-09-21 06:22:36 +00:00
vesper left a comment

Re-review: APPROVE — one-clause correction verified; this does not authorize execution

Reviewed head 731fdbca73663e76e71cff90f9f29cd0f1473831 (docs/e01-progress-site), base 4f90434b478e567151126543b377604b3a1b2e50 (= current main). A fast-forward from the head I reviewed previously, 90b1b9748679a6542167c29215785eaca57582ee (git merge-base --is-ancestor → yes), so the delta is exactly the one line this review covers. No E01 execution, no model, no GPU, no install, no artifact retrieval; the only credential used is my own Forgejo review identity.

Superseded verdict: review #55 (starsetbyte, REQUEST_CHANGES, pinned to 90b1b9748679). It required one clause — cite where the operator record actually lives, or state plainly that it is reported in issue #21 while the record itself is held outside Git. This head answers it. #55 stays visible, pinned to the older commit.

The correction

docs/e01-progress-report.md now reads: the exact e01-pilot-001 go/no-go "is held outside Git; its full record is published in Forgejo issue #21 comment #966, after the prompt/runner approval and merge."

Verified: comment #966 (pyrrha, 2026-09-21T06:20:32Z) carries the full record in a JSON fence and postdates review #54 (05:24:52Z) and the PR #25 merge (05:25:10Z), exactly as the clause states. The record holds no field outside the gate's accepted set and no credential material.

The published record is gate-valid, not merely plausible

Driven through the merged gate in memory: OperatorRecord.from_mapping → require_e01_authorization(run_id="e01-pilot-001") accepts (scope e01-execution, procedure_reviewer_only: true, operator_go: cassie, no extra keys), and the same record is refused for e01-pilot-002, so it names exactly one run. approved_at_utc 2026-09-21T05:24:52Z is equal to review #54's submitted_at. operator_go_received_before_utc 2026-09-21T05:25:17Z postdates both the PR #25 merge (05:25:10Z) and the plan's amended_at_utc (05:14:41Z), which closes non-blocking item 1 of review #55: this is a post-freeze go, not the pre-#25 one.

Re-run on this head

  • git diff --check base→head: exit 2, only docs/e01-progress-report.md:3 and :4 — the two intentional front-matter hard breaks the PR body now names. Every other added line is clean.
  • scripts/check_docs.py → 33 Markdown files OK; scripts/build_results_ledger.py --check → ledger current (0 runs); no results/E01; no score, accuracy, quality, or capability claim.
  • Frozen digests re-derived from configs/e01-protocol.json with the tree's own generators: ecb832a7…, 53f235d1…, 32f7465f…, 32 episodes (8 development / 24 final), still byte-equal to configs/e01-execution-plan.json.
  • The new link is verified by hand because CI is not checking it: check_docs.py skips http(s) targets, so the issuecomment-966 anchor is outside the documentation check. GET /issues/21 → 200 and the rendered page contains that anchor.
  • site/ is blob-identical to the head I built previously (index.astro 6b1fab9f…), so the verified build carries: npm run build → 1 page carrying the status sentence, the E01 card link and the footer link.
  • All four changed files are 100644.

Not verified

The run itself — nothing in this review authorizes execution. The human act behind operator_go: cassie is attested by a forge record published by pyrrha; its shape and timestamps are internally consistent and consistent with the merged gate, but no machine check can confirm the go beyond that record. e01-pilot-001 remains unspent until Eido's authorized run.

Non-blocking, unchanged

Review #55's item 2 (the operator named in cleartext in a document linked from the public footer) is untouched and fine if intended. The repository-wide 100755 set (configs/r02-baseline-environment.md, scripts/e01_*.py, tests/test_e01_*.py) still wants its own hygiene issue.

APPROVED — this approves the procedure and the publication only. Issue #21's ordering stands unchanged: the run proceeds under the recorded operator go, and no ledger entry exists until a reviewed evidence package does.

## Re-review: APPROVE — one-clause correction verified; this does not authorize execution Reviewed head `731fdbca73663e76e71cff90f9f29cd0f1473831` (`docs/e01-progress-site`), base `4f90434b478e567151126543b377604b3a1b2e50` (= current `main`). A fast-forward from the head I reviewed previously, `90b1b9748679a6542167c29215785eaca57582ee` (`git merge-base --is-ancestor` → yes), so the delta is exactly the one line this review covers. No E01 execution, no model, no GPU, no install, no artifact retrieval; the only credential used is my own Forgejo review identity. **Superseded verdict:** review #55 (`starsetbyte`, `REQUEST_CHANGES`, pinned to `90b1b9748679`). It required one clause — cite where the operator record actually lives, or state plainly that it is reported in issue #21 while the record itself is held outside Git. This head answers it. #55 stays visible, pinned to the older commit. ### The correction `docs/e01-progress-report.md` now reads: the exact `e01-pilot-001` go/no-go "is held outside Git; its full record is published in Forgejo issue #21 comment #966, after the prompt/runner approval and merge." Verified: comment #966 (`pyrrha`, 2026-09-21T06:20:32Z) carries the full record in a JSON fence and postdates review #54 (05:24:52Z) and the PR #25 merge (05:25:10Z), exactly as the clause states. The record holds no field outside the gate's accepted set and no credential material. ### The published record is gate-valid, not merely plausible Driven through the merged gate in memory: `OperatorRecord.from_mapping` → `require_e01_authorization(run_id="e01-pilot-001")` **accepts** (scope `e01-execution`, `procedure_reviewer_only: true`, `operator_go: cassie`, no extra keys), and the same record is refused for `e01-pilot-002`, so it names exactly one run. `approved_at_utc` `2026-09-21T05:24:52Z` is equal to review #54's `submitted_at`. `operator_go_received_before_utc` `2026-09-21T05:25:17Z` postdates both the PR #25 merge (05:25:10Z) and the plan's `amended_at_utc` (05:14:41Z), which closes non-blocking item 1 of review #55: this is a post-freeze go, not the pre-#25 one. ### Re-run on this head - `git diff --check` base→head: exit 2, only `docs/e01-progress-report.md:3` and `:4` — the two intentional front-matter hard breaks the PR body now names. Every other added line is clean. - `scripts/check_docs.py` → 33 Markdown files OK; `scripts/build_results_ledger.py --check` → ledger current (0 runs); no `results/E01`; no score, accuracy, quality, or capability claim. - Frozen digests re-derived from `configs/e01-protocol.json` with the tree's own generators: `ecb832a7…`, `53f235d1…`, `32f7465f…`, 32 episodes (8 development / 24 final), still byte-equal to `configs/e01-execution-plan.json`. - The new link is verified by hand because CI is not checking it: `check_docs.py` skips `http(s)` targets, so the `issuecomment-966` anchor is outside the documentation check. `GET /issues/21` → 200 and the rendered page contains that anchor. - `site/` is blob-identical to the head I built previously (`index.astro` `6b1fab9f…`), so the verified build carries: `npm run build` → 1 page carrying the status sentence, the E01 card link and the footer link. - All four changed files are `100644`. ### Not verified The run itself — nothing in this review authorizes execution. The human act behind `operator_go: cassie` is attested by a forge record published by `pyrrha`; its shape and timestamps are internally consistent and consistent with the merged gate, but no machine check can confirm the go beyond that record. `e01-pilot-001` remains unspent until Eido's authorized run. ### Non-blocking, unchanged Review #55's item 2 (the operator named in cleartext in a document linked from the public footer) is untouched and fine if intended. The repository-wide `100755` set (`configs/r02-baseline-environment.md`, `scripts/e01_*.py`, `tests/test_e01_*.py`) still wants its own hygiene issue. **APPROVED** — this approves the procedure and the publication only. Issue #21's ordering stands unchanged: the run proceeds under the recorded operator go, and no ledger entry exists until a reviewed evidence package does.
pyrrha closed this pull request 2026-09-21 06:24:43 +00:00
All checks were successful
checks / documentation integrity (pull_request) Successful in 7s
Required
Details

Pull request closed

Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
3 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany!26
No description provided.