OpenRouter for legal work. A bundle of vendor-neutral Claude Code skills that tell any LLM or router which model to use for which legal task — grounded in mid-2026 legal benchmarks, not vibes or brand loyalty.

Ask "which model should I use to draft this clause / extract these obligations / research this issue / review this contract / translate this statute?" and the matching skill runs a short intake (cost, speed, accuracy/ stakes, privacy/jurisdiction/language) and returns a concrete recommendation: primary model, fallback, what to avoid, and what a human must still verify.

⚠️ This routes models. It is not legal advice, and it never replaces a qualified lawyer or, for translation, an accredited legal translator. Every recommendation ends with what a human must verify.

No single model is best at legal work — the podium re-ranks by task. On mid-2026 benchmarks:

• Contract drafting → Claude Opus 4.8 leads (and uniquely flags contradictory instructions).
• Info extraction → GPT 5.6 Sol leads (but flattens conditionals; weak on scans).
• Legal reasoning → the top 10 cluster within ~3 points, so cost and speed decide, not rank.
• Contract review → a composite (extraction + reasoning + drafting + conflict-detection); route by sub-task, then grade the result against a 7-class error acceptance test (Spellbook Labs SEC study).
• Legal translation → no clean benchmark exists; directional, human review mandatory.

Routing off a generalist leaderboard picks wrong. Route to the task, under the user's constraints, and always name what a human must verify.

skills/
├── legal-ai-model-router/      # entry point — classifies the task, dispatches to a vertical
├── route-contract-drafting/    # generate / redline / rewrite contract language
├── route-info-extraction/      # pull clauses / dates / parties / obligations from documents
├── route-legal-research/       # issue-spot / apply rules / case analysis / memos / agentic research
├── route-contract-review/      # assess an existing agreement for risk / conflicts + redline
├── route-legal-translation/    # translate contracts / statutes / case law (incl. Arabic/MENA)
└── route-council/              # high-stakes escalation: convene the top 2–3 models + reconcile
data/
└── scorecard-2026-07.md        # canonical dataset: all tables, methodology, caveats, live sources (SoT)

Each route-* skill is self-contained (its own references/scorecard.md) so you can install one vertical or the whole bundle.

Clone, then copy the skills you want into your Claude Code skills directory:

git clone https://github.com/sboghossian/legal-ai-model-router.git
cd legal-ai-model-router

# whole bundle
cp -r skills/* ~/.claude/skills/

# or a single vertical
cp -r skills/route-contract-drafting ~/.claude/skills/
cp -r skills/legal-ai-model-router   ~/.claude/skills/   # recommended entry point

Then just ask in natural language — e.g. "which model should I use to redline this NDA on a budget?" — and the router triggers. Or invoke directly: /route-contract-drafting.

The skills are plain Markdown instructions; they carry the benchmark data inline, so they work with any Claude Code setup and need no API keys, network calls, or dependencies.

Every vertical uses the same four intake axes (inferred from your request; you're only asked what's missing, batched and multiple-choice):

Output is uniform:

PRIMARY:    <model> — why, tied to your axes + the benchmark
FALLBACK:   <model> — when to switch
ESCALATE IF: <trigger> → <stronger model / human>
AVOID:      <model> — why, for THIS task
CONFIDENCE: low | med | high
VERIFY:     what a human must still check

By default the bundle is pure prose and data — no code, no network, no dependencies. CONFIDENCE above is then the model's own estimate of itself, which is exactly the wrong place to take a model at its word: a router's most important decision is knowing when to distrust its own classification.

skills/legal-ai-model-router/classify.py replaces that self-report with a measured probability:

export TYPESAFE_API_KEY=...          # console.typesafe.ai
python3 classify.py "Review this Arabic MSA and redline it, then translate it"

VERTICAL:   contract_review  (p=0.99)
STAKES:     Medium  (score=1.3)
IS LEGAL:   0.87
COMPOSITE:  0.81
ACTION:     decompose

It calls Jev, TypeSafe AI's System One model, which returns typed answers with calibrated probabilities rather than text. About $0.00002 per call and ~700ms.

It is an enhancement, never a dependency. No key, no network, a bad response: the script exits 2, says why, and the skill reads the same five-row table it always has. Nothing in the bundle requires it.

The obvious objection, stated rather than buried: a tool whose pitch is vendor-neutral now ships an optional classifier that calls exactly one vendor. That is a real tension. Two things keep it honest — the path degrades to the offline table rather than failing, and the vendor being called is not one of the models being ranked, so it cannot put a thumb on the scale of its own recommendation. If that trade is not worth it to you, do not set the key; you lose a number and nothing else.

A single top-scorer that catches 8 of 10 issues is materially incomplete, not "80% good." So for High-stakes work — signable/filing-facing text, full contract review, obligations extraction, conflict detection, or purchase/M&A agreements (auto-elevated by the contract-error risk prior) — the router escalates from one model to a council: route-council convenes the top 2–3 cross-provider models, runs them independently, then reconciles.

• The panel re-ranks by task, grounded in the scorecard: e.g. drafting → Opus 4.8 + Grok 4.5 + Gemini 3.5 Flash; extraction → GPT 5.6 Sol + Opus 4.8 + Grok 4.5; reasoning → Fable 5 + Gemini 3.1 Pro + GPT-5.6 Sol. Cross-provider on purpose — same-family models share correlated errors, so their agreement is weaker evidence.
• Two reconciliation modes. Recall-critical work (extraction, review) takes the union of findings — every issue any member flags is a candidate, never majority-voted away. Judgment work (a conclusion, a term) treats agreement as confidence and material disagreement as a signal to escalate to a human — never a silent pick.
• Calibrated confidence: unanimous → HIGH, isolated dissent → MED, material disagreement → LOW (escalate). Confidence measures model agreement, not correctness — the human verify step stays mandatory.
• Output keeps the base contract and adds CONVENED / CONSENSUS / DISSENT / RECONCILE. Council is 2–3× the cost and slower, so it's reserved for the high-stakes tail — low-stakes, high-volume, and real-time work keep the single-model route.

The data is triangulated from public benchmarks and independent research, all cited in data/scorecard-2026-07.md:

• legalbenchmarks.ai — Contract Drafting & Info Extraction (lawyer-checklist, all-pass grading).
• Vals AI × Stanford CodeX — LegalBench — legal reasoning across 124 models.
• Harvey — Legal Agent Benchmark (LAB) — task-design standard for agentic legal work.
• Spellbook Labs — SEC contract-error study — ~60% of 3,019 filed contracts carry a drafting issue (stable over 20 years; deal/M&A paper 12–15% high-risk). Ranks no models; it's the acceptance test contract review is graded against, and the risk prior that elevates M&A stakes.
• Wei Chen / The Atticus Project — CUAD, MAUD, ACORD; the "demand public benchmarking" thesis and the capability ≠ controllability principle.
• WMT25, SwiLTra-Bench, ArabLegalEval — the (thin) legal-translation evidence.
• Artificial Analysis, LMArena, SWE-bench, HealthBench — generalist cross-checks and the "verticals re-rank the podium" lesson.

• Numbers are a point-in-time prior (snapshot 2026-07) from boards that change monthly and disagree.
• legalbenchmarks.ai is English-only, US/UK-skewed, single-turn, LLM-judge-graded. Vals LegalBench is multiple-choice reasoning. Non-US, non-English, multi-turn, and long-horizon work is under-measured.
• Legal translation has no reliable public leaderboard — that vertical is explicitly directional.
• Hallucinated citations/clauses are the cardinal legal-AI risk. Every route ends with a verify step.
• The optional classifier measures calibration, not correctness. Measured, not assumed (2026-09-21, jev-1.13.0, 11/11 on the fixture set). Arabic, French and English classify at parity — verified, not claimed. Three limits worth knowing:
  • Confidence does not drop on a vague-but-plausible request. "Take a look at this NDA" returns contract_review at 1.00. A high number means "this reading is coherent", never "the user was clear".
  • is_composite on "review and redline" sits at 0.45 against a 0.50 threshold — deliberately, since route-contract-review owns that pair, but a thin margin worth re-measuring if that question changes.
  • An ambiguous acronym moves is_legal, and that is the gate behaving correctly. "Review this Arabic MSA and redline it" scores is_legal 0.30, because "Arabic MSA" reads as Modern Standard Arabic rather than a Master Services Agreement. Drop the word Arabic (0.97), spell the term out (0.98), or say "MSA written in Arabic" (0.96) and it resolves. The vertical pick stayed contract_review throughout — only the legal-ness signal moved. Worth knowing well beyond this tool: MENA legal work produces that exact phrase constantly, and it is ambiguous to any model, not just this one.

Benchmarks drift. When you refresh:

1. Edit data/scorecard-2026-07.md (the single source of truth) — update tables, bump the date, rename the file to the new month. Exception: a section that adds new data without refreshing the third-party boards (like §6, §7 and §8) keeps the existing snapshot date and carries its own date in its heading — renaming the file would falsely imply §1–§5 were re-pulled.
2. Re-slice the per-skill references/scorecard.md files from it.
3. Update the inline tables in each SKILL.md and note the change in CHANGELOG.md.

Live sources to re-pull are listed at the bottom of the scorecard.

AGPL-3.0-or-later. You're free to use, modify, and redistribute; derivative works and network services must stay open under the same license.

Provided as-is, for informational and decision-support purposes only. It routes AI models; it does not provide legal advice, and no model recommendation should be relied upon without review by a qualified legal professional. Benchmark data may be inaccurate or outdated. Model names and scores reflect third-party benchmarks and are used descriptively; no affiliation or endorsement is implied.