A6 — Equity & Research pack fixtures

[jonathanbbechtel]

Summary

• Expands spec.RESEARCH_REFS from 3 to 10 real, citable published tutoring/intervention studies (R1–R10), with 3 entries carrying non-null contradicts_claim to support Track 5 mixed-evidence tasks.
• Generates fixtures/pack_equity_research/ via generator.py --seed 42 --pack equity_research: subgroup attendance + outcomes CSVs, research_refs.json (10 entries), program_context.json, and manifest.json. Deletes .gitkeep.
• Authors docs/dataset_cards/pack_equity_research.md documenting subgroup variance, intentional low-N suppression, IEP disparity, and research reference structure.

Implements platform-ai#574.

Signals embedded

• AIAN low-N suppression (Track 4): AIAN n=3 per month (< threshold of 10); suppressed in all 3 months. Also suppressed: NHPI (n=4), TwoOrMore (n=7), Unknown (n=1), NB (n=6), U (n=6).
• IEP attendance disparity (Track 4): IEP=true avg attendance 68.5% vs IEP=false 82.5% — realized gap ~14 pp (spec penalty is 11 pp; seed-42 stochastic variation adds ~3 pp).
• Research refs (Track 5): 10 references; R2 (Robinson et al.), R4 (Fryer), R10 (Vaughn et al.) carry contradicts_claim.

Test plan

• All 81 existing unit tests pass (pytest)
• ruff check . — no issues
• ruff format --check . — all formatted
• pre-commit run --all-files — all hooks pass (including mypy)
• research_refs.json validated: 10 entries, all with required fields, 3 with non-null contradicts_claim
• pack_operations and pack_outcomes NOT regenerated — only pack_equity_research committed
• Numeric signal constants in spec.py unchanged (AIAN_TARGET_N, IEP_ATTENDANCE_PENALTY, etc.)

🤖 Generated with Claude Code

[pearleng-atlantis]

Error: This repo is not allowlisted for Atlantis.