Skip to content

Tighten diagnose-issue evidence rules after the OK #8911 incident#1085

Open
PavelMakarchuk wants to merge 1 commit into
mainfrom
tighten-diagnose-issue-evidence-rules
Open

Tighten diagnose-issue evidence rules after the OK #8911 incident#1085
PavelMakarchuk wants to merge 1 commit into
mainfrom
tighten-diagnose-issue-evidence-rules

Conversation

@PavelMakarchuk

Copy link
Copy Markdown
Collaborator

Why

Two of our own filings encoded false or mis-scoped primary-source claims:

  • pe-us #8830 → PR #8911 (reverted via #9009/#9016): asserted the 2021 OK Form 511 omitted IRC §401 pensions while citing the exact page that lists §401 first. No verbatim quote anywhere; the "2021 list" was the real list minus the hypothesized additions; tests were derived from the premise (circular); merged with zero reviews.
  • pe-us #8831 → PR #8875 (fixed same-day by #8902): quotes were real but attributed to the wrong Schedule IN-EIC section, spawning an unsupported childless carve-out.

A systematic audit of all June–July filings (16 items, 4 agents, verbatim re-extraction of every cited source) found no other fabrications — but both incidents share a gap the existing skill didn't close: citations were required, extraction wasn't.

What

.claude/commands/diagnose-issue.md:

  • Critical Rules 5–9: verbatim mechanical extraction required; negative claims need the full section extracted; year-change claims need both years diffed; quote scope (section/heading) verified, not just existence; engine outputs are hypothesis generators, never evidence; test expectations must trace to an external ground truth (TaxAct PDF / binary / official worked example), never the premise.
  • Step 7: concrete extraction how-tos (pdftotext path, HTML-as-PDF trap, archive.org fallback for repealed statutes, NBER historical-forms mirror).
  • Step 9: five-point pre-filing gate + mandatory Evidence section in drafts.
  • Common Root Causes 12 (our own prior fix was wrong — re-verify its premise) and 13 (parameter backward-leak, from CT #9008).

CLAUDE.md: session-wide six-rule summary so the requirements bind even outside /diagnose-issue.

🤖 Generated with Claude Code

Adds hard filing requirements to /diagnose-issue (Critical Rules 5-9,
Step 7 extraction how-tos, Step 9 pre-filing gate, root causes 12-13)
and a session-wide summary in CLAUDE.md:

- verbatim mechanical extraction required for every primary-source claim
- negative claims need the full extracted section; year-change claims
  need both years diffed
- quote scope (section/heading) must be verified, not just existence
- engine outputs are hypothesis generators, never evidence
- test expectations must trace to external ground truth, never the premise

Born of pe-us #8830/#8911 (fabricated 2021-OK-form claim, reverted in
#9009) and #8831/#8875 (real quote, wrong section, fixed by #8902).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@vercel

vercel Bot commented Jul 13, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
policyengine-taxsim Ready Ready Preview, Comment Jul 13, 2026 3:13pm

Request Review

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant