Habeas

Fine-tune a multimodal Qwen3.8-27B (28B, Apache-2.0) on Modal/GCP free credits so it beats frontier models at Form I-9 compliance validation.

Form I-9 + presented identity/employment-authorization documents (scans) go in → an audit-ready report comes out: every field problem, document acceptability issue, timeliness violation, and reverification gap, each cited to the M-274 Handbook / 8 CFR 274a.2.

Fourth of four parallel repos (specula, suture, plumb, habeas) on one shared methodology (docs/methodology.md). Independent and operable in parallel.

Why this niche

• Static federal rules, zero drift. One jurisdiction, one form, two editions (2023-08-01, 2025-01-20 — the latter postdates most frontier training), one static rule set (M-274 + 8 CFR 274a.2). Every check is deterministic given the inputs — the oracle is a rules engine.
• Vision-native: the model reads actual presented documents (passports, EADs, DLs, SSN cards) — the exact regime where the multimodal 27B wins.
• Real money at stake: statutory fines per form (~$288–2,861 paperwork, up to ~$28,619 substantive per 8 CFR 274a.10, 2025-adjusted); ICE audit waves; 3-business-day deadline automation alone sells.
• Competition (research-verified): LawLogix / I-9 Advantage / Tracker are workflow SaaS, not LLM-native validation with citation-grade output.

The process (end-to-end, fully automatable)

1. Ingest employee Section 1 + Section 2 + Supplement A/B data, plus scans of presented documents.
2. Validate every field; classify documents against the fixed List A/B/C acceptability table; check category-code/expiry/habeasation consistency.
3. Enforce timeliness (Section 2 ≤ 3 business days from hire) and reverification (Supplement B when work authorization expires); apply the remote-examination (E-Verify) branch.
4. Emit audit report JSON: PASS | FLAG + {type, severity, field, observed, expected, cfr, correction}. Human signs off.

Benchmark contracts

See CONTRACTS.md — fixed violation taxonomy + severity weights, scoring, contamination, benchmark rules.

Directory layout

forge/      seeded I-9/document generator + M-274/8 CFR oracle + contamination
            monitor + golden benchmark (pkg habeas_forge)
model/      multimodal dataset builder, QLoRA SFT, GRPO RLVR, benchmark eval
cloud/      Modal app + Dockerfile + GCP spot scripts
eval/       deterministic golden harness
docs/       DECISIONS.md, BENCHMARK.md, HANDOFF.md, methodology.md

Training stack & cloud

Shared stack (docs/methodology.md): QLoRA 4-bit SFT → GRPO/Dr.GRPO (DAPO-style) RLVR against the I-9 oracle → ReST-EM → s1 curation → benchmark. Modal your-modal-profile primary, GCP fallback. Smoke LoRA first (DeltaNet gate).

Targets (golden benchmark, held-out seed, zero contamination)

Head-to-head vs Qwen3.8-2.4T-A95B, DeepSeek v4-flash, base Qwen3.8-27B.

Phases

P0 scaffold (current) → P1 rules engine + hand-verified cases → P2 generator + golden set → P3 dataset + SFT → P4 RLVR → P5 self-play/judges → P6 head-to-head + writeup.

Risks (honest)

• Heavy PII — synthetic-only training data, no real forms ever; encryption/access controls; validator positioning with human sign-off.
• Tamper / physical-document fraud out of scope (validator, not authenticator).
• DHS rulemaking churn is a tailwind (remote-examination procedure).