fix: triage blind holdout LinkedIn misses

[ftchvs]

Summary

• Adds a scoped derived LinkedIn professional-claim review path for softer professional-outcome language.
• Keeps existing hard LinkedIn promise signals (double your salary, guaranteed promotion, 10x productivity) high risk.
• Documents the AND-62 blind-holdout follow-up and the remaining conservative telehealth overcall.

Blind holdout impact

• Decision accuracy: 0.967 → 0.989
• Decision mismatches: 3 → 1
• Policy false negatives: 12 → 10
• Policy false positives: unchanged at 7
• Remaining decision miss: blind_telehealth_info_review (needs_review expected, high_risk actual) due expected Google health policy firing with conservative high-risk scoring.

Test plan

• make real-world-blind-ci
• make real-cases-ci
• make test
• make pr-preflight

Refs AND-62.

[linear-code]

AND-62 Review remaining blind holdout misses after PR #9

Why

PR #9 merged the AdLint reliability backlog and left the blind holdout in a known conservative state: 90 rows, rule-only decision accuracy 0.967, three documented decision mismatches, 12 policy false-negative notes, and 7 policy false-positive notes.

The next useful increment is to review the three remaining decision mismatches without broadening product scope or tuning blindly against the holdout.

Remaining decision mismatch row IDs:

• blind_productivity_claim_review
• blind_promotion_workshop_review
• blind_telehealth_info_review

Scope

Classify each remaining miss as one of:

• true rule gap
• label/adjudication issue
• acceptable limitation
• future model/extraction issue

Apply only narrow, reviewed rule or label/reporting changes when the evidence supports them. Do not change source rows just to improve the metric.

Acceptance criteria

• Start from main after PR feat: complete AdLint Linear backlog #9 merge commit 83d0355.
• Reproduce the current blind baseline with make real-world-blind.
• Produce a row-by-row triage note for the three remaining mismatches.
• If a rule change is justified, keep it narrow and explain affected policy IDs.
• Preserve existing gates: make pr-preflight, make test, make real-cases-ci, make real-world-blind-ci, make policy-coverage-validate.
• Update docs with before/after metrics and the holdout-boundary decision.
• Open a small PR; do not bundle unrelated Meta Ads, storage, scoring, or research-loop work.

Notes

Related shipped work: PR #9, AND-28, and AND-49. Deterministic rules remain the production baseline; local model quality remains manual/scheduled diagnostics.

Review in Linear