Fix stale accuracy claims in RESEARCH.md and the PR template

[dmccoystephenson]

Summary

Stage A documentation-accuracy sweep. Verified every claim in the three sources of truth (create-dev-loop.md, README.md, RESEARCH.md) against source. Three claims had drifted, all in RESEARCH.md, plus one sibling doc left behind by PR #70.

No tracking issue — gaps found during a Stage A documentation sweep. The open-issue backlog on this repo was empty at triage, so no issue was deferred and no skip reasons apply.

• §1, PR Generalize Phase 4 external-signal step to support no-CI repos #28 entry — asserted that create-dev-loop itself is a repo "with no CI". False since PR Add CI: automated doc-consistency checks #65 (2026-07-27) added .github/workflows/ci.yml, and PR Add MIT license badge to README #71 added the CI badge to the README. Rewritten to state what is actually true and why the finding still holds: the workflow only runs scripts/check_docs.py, so the anchor for a behavioral change here remains the manual /create-dev-loop validation checklist, not the green check. That makes this repo an illustration of the finding rather than a counterexample to it.
• §2, PR Bias triage and batching toward higher-merge-rate issue types #29 entry — pointed readers at issue Add hard PR scope gates: LOC ceiling, files-touched ceiling, do-not-auto-merge path list #17 as an open "structural-gates follow-up". Issue Add hard PR scope gates: LOC ceiling, files-touched ceiling, do-not-auto-merge path list #17 was closed 2026-05-25 by PR Add hard PR scope gates: LOC ceiling, do-not-auto-merge paths #30, whose Implementations entry sits on the very next line. Reworded to record the outcome instead of dangling the follow-up.
• Last reviewed: — 2026-07-25 → 2026-07-27, the date of this sweep.
• .github/PULL_REQUEST_TEMPLATE.md — the Test plan hint still carried the pre-Mention CI in CONTRIBUTING's validation step #70 framing ("There is no automated test suite for this repo") and never mentioned that CI mechanically checks the Doc sync boxes printed directly above it. PR Mention CI in CONTRIBUTING's validation step #70 applied exactly this correction to CONTRIBUTING.md; the PR template is its sibling and was missed. Now matches Mention CI in CONTRIBUTING's validation step #70's wording, and is precise that CI covers the first two boxes (the RESEARCH.md box is not machine-checkable).

Nothing else drifted. create-dev-loop.md and README.md were checked claim-by-claim and are accurate: the seven Steps map 1:1 to README's "What it does", every {{placeholder}} and {{#if}} conditional has a Step 4 row, all inner fences inside the embedded template retain their leading-backslash escape, the template-version / generated-at comments sit below the H1, the four Edge-cases entries named by Step 4's no-test-suite paragraph all exist under those exact titles, and every RESEARCH.md §N cross-reference in the template points at a section that says what the citing line claims.

Research grounding

No finding is implemented or contradicted by this PR — it corrects the accuracy of RESEARCH.md's own bookkeeping and a contributor-facing doc, and does not touch create-dev-loop.md, the embedded template, the Step 4 table, or any phase definition. Stating that explicitly per CLAUDE.md's "Grounding work in research" section rather than reaching for a citation. No Implementations entry is added for the same reason.

No retrofit of existing generated skills is needed: create-dev-loop.md is unchanged, so no rule was promoted into the template.

Doc sync check

• README.md's "What it does" Step list still matches create-dev-loop.md's Steps 1:1 (no Step added, removed, or renumbered)
• No {{placeholder}} added or changed
• RESEARCH.md updated — accuracy corrections only; no new finding and no Implementations entry, per "Research grounding" above

Test plan

• python3 scripts/check_docs.py passes locally, and the doc-consistency CI job is green on this PR head
• Every citation URL in RESEARCH.md resolves — all 21 arXiv/vendor links return HTTP 200; openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/ returns 403 to a scripted client (bot filtering), not a dead link
• Drift confirmed empirically, not by reasoning: .github/workflows/ci.yml exists on main, and gh issue view 17 reports CLOSED 2026-05-25T18:47:41Z. Both claims read false against the pre-fix text and true against the post-fix text
• git status confirmed no .claude/ or scratch files staged; 2 files, +10/-7
• UNVERIFIED — not applicable. The /create-dev-loop end-to-end anchor was not run. Per the loop's Phase 4 scope gate, the anchor is required only when a PR changes the embedded template, the Step 4 substitution table, or Step 1/5–7 logic. This PR touches none of those files, so the docs-only path applies. CI on this PR is the live external signal.

---

This PR description was drafted during a Gardener session (Stephenson-Software/gardener).

[dmccoystephenson]

Self-review rubric (Phase 4, one intrinsic pass — anchored on green doc-consistency CI):

• Scope: PASS, with one judgment call flagged below. Two files, +10/−7. Both are documentation whose inaccuracy was confirmed against source before editing. No formatting churn, no renames.
• Tests-new: N/A — no public method or function exists or was added; this repo's only executable surface is scripts/check_docs.py, untouched.
• Tests-fix: PASS — this repo's analog of stash-and-run is "does the claim read false against source before the edit and true after". Confirmed by two independent commands, not by reasoning: .github/workflows/ci.yml is present on main (so "repos with no CI, including create-dev-loop itself" was false), and gh issue view 17 returns CLOSED 2026-05-25T18:47:41Z (so "issue Add hard PR scope gates: LOC ceiling, files-touched ceiling, do-not-auto-merge path list #17 is the structural-gates follow-up" pointed at closed work).
• Sibling structure: N/A — no new file created.
• Sibling renames: PASS — and this PR is the sibling fix: PR Mention CI in CONTRIBUTING's validation step #70 corrected the "no automated test suite" framing in CONTRIBUTING.md and left the parallel line in .github/PULL_REQUEST_TEMPLATE.md behind. Wording now matches Mention CI in CONTRIBUTING's validation step #70's.
• Docs: PASS — all three Phase 7 sources verified claim-by-claim. create-dev-loop.md and README.md needed no change; RESEARCH.md is the file corrected.
• Issue resolution: N/A — no Closes #N; the backlog was empty at triage, which is why this cycle ran Stage A. Noted in the PR body per the Phase 4 rule.
• Placeholder table completeness: PASS — create-dev-loop.md is byte-identical to main; check_docs.py re-confirms every {{TOKEN}} has a Step 4 row, and it passed in CI on this head.
• Fence escaping intact: PASS — zero fences added or removed; the embedded template is untouched. Separately verified during the sweep that no unescaped ``` exists between the template's opening and closing fences.
• README/Step 1:1: PASS — no Step added, removed, or renumbered; enforced mechanically by the green CI job.
• RESEARCH.md citation: PASS — the citation rule triggers on template/phase changes; this PR makes none, and the PR body states explicitly that no finding applies rather than reaching for one.
• External anchor (doc-consistency CI): PASS — green on the PR head in 5s. The /create-dev-loop end-to-end anchor is recorded UNVERIFIED-not-applicable: it is required only for changes to the embedded template, the Step 4 table, or Step 1/5–7 logic, none of which this PR touches.

Judgment call for a reviewer, not a mechanical fix. This cycle declared a Stage A sweep of the three sources of truth in CLAUDE.md's table; .github/PULL_REQUEST_TEMPLATE.md is not in that table, so including it is arguably out of the declared surface. It was kept in rather than split out: it is a four-line correction to the direct sibling of a doc PR #70 already fixed, and Phase 3's "rename siblings together" rule exists precisely so a parallel doc isn't left stale. A separate PR for four comment lines would cost a reviewer more than it saves. Flagging it so the call is visible rather than silent.

Observation, no change made. §1's PR #57 entry still reads "an anchor named CI in a repo that has no CI is an anchor the loop cannot check." That sentence is generic — it describes the template hardcoding CI for every generated skill — so it stays correct now that create-dev-loop has its own CI. Checked deliberately because the #28 edit above sits four lines away and could have made it read as a contradiction.

---

drafted by Claude on behalf of Daniel Stephenson