Establish v6 baseline: validate H10 detection, scoring recalibration, and adversarial-30 regression

[cpitzi]

Summary

v6 shipped in #44 (Heuristic 10 journal-legitimacy flagging + the #43 scoring-formula fix) and is now the current production prompt — but, like v5 before its #40 validation, it has never been run. Two things changed that need empirical confirmation before v6 is treated as production-validated.

Mirrors #40 (which validated v5).

Tasks

• H10 detection — run v6 against test-sets/predatory-venues.md. Confirm refs 1 and 4 flag Elevated (H10 alone), ref 3 flags High (H10+H7), and the two clean controls stay Defensible: ref 2 (Faridvand et al., Int. J. Nursing Practice, DOI 10.1111/ijn.12659) and ref 5 (Vogel et al., Lancet Global Health, DOI 10.1016/S2214-109X(15)70094-X). These controls were corrected during feat(prompt): v6 auditor — journal legitimacy (H10) and scoring formula fix #44 review — verify v6 clears them under H10 (indexed journals must not flag).
• Scoring calibration — under the new formula (no D×3), clean real articles should now score high: re-run Ahmadinezhad and Patriksson and confirm they land near the top of the scale (they scored 5 and 22 under v5's D×3 formula). The dirty corpus (Amarnath, Madhukar) should still score low.
• Adversarial-30 regression — run v6 against test-sets/adversarial-30.md (corrected test-sets/adversarial-30: trap-index ground-truth errors surfaced by the v4 baseline run #37 key: 14 clean / 16 manipulated). H/E/M/D tier assignments should match v5 (sensitivity 100%, specificity 92.9%); H10 should fire on nothing (adversarial-30 has no predatory-venue traps), so any new H10 flag on a clean ref is a false positive to capture. Note the numeric scores will differ from v5 because the formula changed.
• Commit reports with -v6- filenames; do NOT overwrite the v5 baseline (reports/adversarial-30-v5-2026-06-19.html).

Notes

• Dispatch opus with live web search (H10 hits DOAJ api.doaj.org, PubMed, Scopus, WoS).
• Verify web-dependent claims against the source before trusting the report, per the Run fresh adversarial-30 test with v3 prompt and commit report #13/test-sets/adversarial-30: trap-index ground-truth errors surfaced by the v4 baseline run #37/Establish v5 baseline: run v5 against adversarial-30, temporal-impossibility, and real-articles #40 discipline — especially confirm the H10 controls clear (the journal-indexing lookups).
• Because both the heuristic set and the scoring formula changed, v6 metrics are not directly comparable to the v5 baseline numbers.

Origin

Surfaced by #44 (the v6 release). The H10 prompt logic and scoring fix were verified for correctness during PR review (and two broken clean-control DOIs were corrected pre-merge), but no audit was run; this issue tracks the empirical validation.