[dmccoystephenson]

Summary

Closes the feedback loop between research-grounded predictions and what we actually observe by:

• Adding an Implementations convention to RESEARCH.md's "How to use this document" section: when a finding gets implemented, record the PR + ship date + observed-effect placeholder under the finding.
• Adding a corresponding bullet to CLAUDE.md's "Grounding work in research" so the rule appears at the same point of friction (PR authoring) where forward citations are already required.
• Retroactively populating Implementations entries for findings that have already been acted on:
  • §1 (self-critique with external signal) → PR Reframe self-review as rubric-based and CI-grounded #15 (rubric-based self-review)
  • §4 (long-horizon decay, context rot) → PR Promote 'verify each Closes #N' to template Phase 4 #26 (verify each Closes #N before push) — a narrow instance of "re-state scope at boundaries"
  • §5 (reflection iterations plateau) → PR Reframe self-review as rubric-based and CI-grounded #15 (one-intrinsic-critique-pass cap)

The "observed effect" column starts at "pending" — replacing it with real numbers requires running multiple cycles, which only becomes routine once #24 (structured self-audit) lands. This PR establishes the data-recording habit so the column has somewhere to go.

Closes #25

Test plan

• RESEARCH.md "How to use" section gains the Implementations convention with format example
• CLAUDE.md "Grounding work in research" gains a matching bullet pointing at RESEARCH.md
• Three retroactive entries placed under §1, §4, §5 — confirmed by grep -n "^\*\*Implementations\." RESEARCH.md returning lines 44/104/124 (one per finding)
• No changes to create-dev-loop.md template; substitution table unchanged
• PR Reframe self-review as rubric-based and CI-grounded #15 ship date 2026-05-25 confirmed via gh pr view 15 --json mergedAt
• Per the rule shipped in PR Promote 'verify each Closes #N' to template Phase 4 #26, ran gh issue view 25 and confirmed issue title/body match this PR's work

Note on AC #4

Issue #25's fourth AC ("at least one updated outcome entry, i.e. not just 'pending'") cannot be honestly satisfied yet — none of the three implementations has been measured across enough cycles to replace "pending" with a real observation. Marking that AC explicitly as deferred rather than fabricating an outcome.

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — only CLAUDE.md (1 line) and RESEARCH.md (15 lines across 4 places) modified; all changes implement #25's AC #1–#3 directly.
• Tests-new: PASS — the convention introduced is prose; the three retroactive Implementations entries demonstrate the format being applied, making the convention self-exemplifying.
• Tests-fix: N/A — not a bug fix.
• Sibling structure: PASS — new How-to-use bullet follows the same **When ...** — ... shape as its three neighbors; Implementations subsections all use the same **Implementations.**\n- PR #N (summary): shipped DATE. Observed effect: ... shape.
• Sibling renames: N/A — no identifiers renamed.
• Docs: PASS — Phase 7 sources of truth for this repo are create-dev-loop.md and README.md; neither changed because this PR adds a process rule to RESEARCH.md/CLAUDE.md (themselves Phase 7 sources for the project, both updated together to keep them in sync).
• Issue resolution: PARTIAL but transparent — AC #1, #2, #3 met; AC #4 ('at least one updated outcome entry not just pending') explicitly deferred in the PR body since outcomes require multi-cycle data that doesn't yet exist. Calling this PASS-with-noted-deferral rather than FAIL because the AC is structurally impossible to satisfy in the same PR that introduces the tracking.
• Validation: PASS — grep -n '^\*\*Implementations\.' confirms the three entries land under §1, §4, §5; gh pr view 15 --json mergedAt confirmed PR #15 ship date.
• Placeholder-consistency: N/A — no template change.
• Escaping: N/A — the fenced code block added to RESEARCH.md is not inside the create-dev-loop.md outer markdown template; it's free-standing in RESEARCH.md, so no triple-backtick escape needed.

Adversarial probe: 'Should Implementations also appear in the Adding-a-new-finding template at the bottom of RESEARCH.md?' Considered and rejected — that template is for new findings, which by definition have no implementations yet; adding a blank subsection would be noise. Leaving as-is is consistent with the issue's 'optional subsection' framing.