todo2code

Ticket 003: Residual changelog diagnostic audit

• ID: ticket-003
• Owner: tom-sapletta-com
• Status: DONE
• Workflow state: DONE
• Created: 2026-07-31

Goal and scope

Audit the CHANGELOG_WITHOUT_IMPLEMENTATION findings that remain after ticket-002, classify a deterministic cross-repository sample, and change the library only when the sample demonstrates one repeated false-positive class that can be removed without treating unsupported release claims as evidence.

The unchanged corpus is:

• semcod/code2llm
• semcod/domd
• semcod/pactfix
• semcod/code2logic
• semcod/code2docs
• semcod/redup
• subactor/platform

External inputs remain detached tracked-only worktrees at the commits recorded by ticket-002.

Acceptance criteria

• AC-01: A current deterministic run is recorded for all seven repositories using tracked 18cc21b plus the explicit ticket-002 diagnostic patch only.
• AC-02: A deterministic stratified sample covers every repository and at least 100 residual CHANGELOG_WITHOUT_IMPLEMENTATION findings.
• AC-03: Every sampled finding has a review label, rationale and enough source/target context to reproduce the classification.
• AC-04: A code change is attempted only for a false-positive class present in at least two repositories with at least 20 sampled examples; otherwise the hypothesis is rejected and the ticket closes without semantic changes.
• AC-05: A focused hard-negative regression is observed failing before any implementation change.
• AC-06: The unchanged corpus demonstrates an improvement in at least two repositories, with stable graph fingerprints and no loss in gold v2 quality.
• AC-07: Full verify, examples, smoke, dependency audit and Docker validation pass; the local Java skip remains allowed only because CI requires JDK.
• AC-08: Results, raw commands, changed files and the next ranked hypothesis are preserved under this ticket and summarized in docs/READINESS.md.

Non-goals

• Broad capability-topic linking for changelog prose.
• Suppressing old or unverifiable behavioral claims merely to lower counts.
• Using an LLM to label the primary audit sample.
• Mutating or reading untracked content from external repositories.
• Combining unrelated semantic heuristics in one A/B result.

Participants

• user-tom-sapletta-com.md
• ai-codex.md

Evidence

• preprompt.md
• audit.md
• sample.json
• scripts/research/audit-changelog-sample.mjs
• iteration-01.md
• iteration-01.json
• ai-codex-logs.txt
• changelog.md

Approval

• Decision: approved
• Evidence: user message kontynuuj, following the ticket-002 conclusion
• Date: 2026-07-31

Conclusion

The evidence supports one narrow correction: exact Update <file> bookkeeping without behavioral wording is not an unsupported implementation claim. The change removed 547 CHANGELOG_WITHOUT_IMPLEMENTATION findings and 188 secondary UNLINKED_RECORD warnings across five repositories. All seven graph fingerprints stayed identical, gold v2 stayed perfect and the full offline validation suite passed.

The 1,306 remaining findings are intentionally retained: 1,275 are substantive or unverified claims, 30 are roadmap entries and one is a file-summary entry. The next ranked hypothesis is to model unchecked roadmap entries through explicit lifecycle/extractor semantics in a separate ticket, rather than hide them with another changelog text filter.

This site is open source. Improve this page.