Skip to content

feat(eval): SOTU gold fixture 2026-07-10 + span/attribution audit (gold-side)#23

Open
aRealGem wants to merge 1 commit into
mainfrom
claude/sotu-gold-fixture
Open

feat(eval): SOTU gold fixture 2026-07-10 + span/attribution audit (gold-side)#23
aRealGem wants to merge 1 commit into
mainfrom
claude/sotu-gold-fixture

Conversation

@aRealGem

Copy link
Copy Markdown
Owner

Integrates the architect's manually-verified 15-claim SOTU gold fixture (written verbatim) plus read-only hygiene tooling. Gold-side only — never enters model context. No merge (Jackie's call).

Deliverables (full detail in HANDOFF_sotu_gold_fixture_2026-07-10.md)

  • (a) Fixture writteneval/benchmarks/claim-set/sotu_gold_fixture_2026-07-10.json (verbatim). Contamination guard verified: no loader globs the claim-set dir into prompts.
  • (b) Taxonomy mapping interim-1 → VerdictLabel (6-label). Clean for 5 values; true_at_utterance/unverifiable_personal_testimony are lossy; normative_with_true_premise does not map (canonical is factual-only) — flagged for Precious.
  • (c) Span resolution 15/15 (resolve_fixture_spans.py + .offsets.json): 14 exact sentence-level; trump2026-03 paragraph-level ("1826" in adjacent sentence); trump2026-05 cross-speech → biden_2022:0325 (confirms the misattribution flag).
  • (d) Attribution audit 277/277 match (read-only). The Dominican-officers claim lives only at biden_2022:0325, correctly attributed — no repo metadata fix needed; the misattribution is fixture-side only.
  • (e) Citation hygiene: resolved canonical primary URLs for the resolve_url:true entries (statutes/treaty/congress/CBO/BEA/BMJ confident; a few flagged for exact-issue confirmation). Proposed in the HANDOFF, not written into the verbatim fixture.

Proposed (not created) follow-up cards: fixture↔PR#22 verdict_gold reconciliation; fixture r2 (re-key trump2026-05, fold URLs); wire the attribution audit into CI.

Suite unaffected (data + standalone gold-side script).
🤖 Generated with Claude Code

…ld-side)

Integrates the architect's 15-claim manually-verified SOTU gold fixture (written
VERBATIM) with hygiene tooling. Gold-side ONLY — must never enter model context;
verified no loader globs the claim-set dir into prompts (contamination guard holds).

- sotu_gold_fixture_2026-07-10.json — payload verbatim (15 claims, 7-value interim enum).
- resolve_fixture_spans.py + .offsets.json — read-only span resolution (15/15 resolved;
  trump2026-03 paragraph-level, trump2026-05 confirmed cross-speech misattribution ->
  biden_2022:0325) and a 277-set attribution audit (277/277 match; the Dominican-officers
  claim is correctly under biden_2022, so NO repo metadata fix is needed).
- HANDOFF_*.md — taxonomy mapping (interim -> VerdictLabel v0.2.0; normative_with_true_premise
  does not map cleanly), citation-hygiene table (resolved primary URLs), contamination-guard
  note, and proposed follow-up cards (not created).

No merges (Jackie), no kanban writes, no corpus mutations.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GNg63eCE6Q7dNJLm7hxSev
@cursor

cursor Bot commented Jul 10, 2026

Copy link
Copy Markdown

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

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.

2 participants