What actually ran

Published evaluation results

The committed DraftBench scorecard (run 9/29/2026) — skips are honest gaps awaiting credentials, never counted as passes.

264/277 passed· 0 failed · 13 skipped

• clauses (deterministic assembly fidelity) — 44✓
• cite-check (format + existence honesty) — 12✓ 2○
• grounding (cited-or-flagged validator) — 10✓
• red-team (adversarial guardrail checks) — 13✓ 3○
• gateway — 14✓
• drafting-gates — 10✓
• matter-os (deterministic extraction) — 12✓
• court-packet (readiness guidance honesty) — 6✓
• attorney network & paralegal mode (finalization-never) — 14✓
• small firm console (deterministic firm playbook) — 12✓
• argument war room (framing invariant + grounding) — 13✓
• source-locked assistant (boundary routing invariant) — 11✓
• upl capability engine (fail-closed matrix + ledger) — 58✓
• accessibility program (census law + token truths) — 12✓
• foundations — 10✓
• paired-methods — 7✓ 8○
• evidence-ledger — 6✓

Model-request ledger (counts only)

Grounding verdicts: n/a 4 — answers the validator could not tie to the document or a cited source are flagged, and the flags are counted here.

Error kinds: credit-balance 22 · timeout 3

What ran, separated

Four different things, never one number: model requests (every attempt), successful deliverables (drafts, revisions and redlines that completed), substantive review (review, opposing-counsel and counsel calls that completed) and attorney adjudications (decisions on findings). Each carries its denominator; excluded and not-run categories are named.

• Model requests: 45 attempts (every attempt of every kind, including refusals and errors).
• Successful deliverables: 7 of 32 deliverable attempts; 0 synthetic canary rows excluded.
• Substantive review: 13 of 13 review attempts completed.
• Attorney adjudications: 2 decisions recorded (0.0% not accepted as proposed).
• Not run in the period: revise, draft-edit, compare, grounded-redline, review, grounded-review, grounded-oc-pass, assistant, research, evidence-analysis, moot-bench.

Attorney adjudication of AI findings

Attorney reversal share (revised + rejected): 0 of 2 (too few to state a percentage; 30 are needed) — we publish this number because a supervised tool should be able to show how often the supervisor overrides it. Aggregates only; no reviewer identities, no matters.

Opposing Counsel Pass findings, as decided

No Opposing Counsel Pass finding has been decided on the trail yet; the share publishes here with its denominator once one is.

What each figure counts

• All-time governed ledger: every governed model request since the ledger began, counted by surface and outcome; never dollar figures, never content.
• The last 30 days: the same ledger over a rolling thirty days, shown beside the all-time figures and never merged with them.
• Trial cohorts: requests made under a trial, reported per cohort once a cohort exists; none is published yet.
• Fixtures and demonstrations: the sample matter, the test fixtures and the live-proof runs; never counted in any figure on this page.

Refusals written down, each with an enforcing symbol and a negative test: 253 by rule, 22 while a dependency is not live, 10 for input shapes never accepted (the catalogue).

No keyed live proof is recorded yet for any of the 4 programs; keyless runs never count.

Complaints, refunds, corrections & accessibility

Every case opened on the support page or the accessibility page is counted here — including the ones we haven't resolved. Counts only; never content or identities.

Resolution share: 100.0%.

The verbatim law corpus

Sections of primary law held VERBATIM with source provenance and a normalized-text fingerprint — the store engines cite from instead of remembering. Counts are derived from the corpus table at render, never hand-typed.

The permission matrix, as enforced

This table is generated by calling the same capability() function the routes enforce — the page cannot claim a posture the code doesn't implement. L5 is disabled for everyone: the platform never files, serves, negotiates, or appears.

15 surfaces are classified in the registry; unregistered surfaces and unknown states fail closed. Hard gates currently enforced: 14 of 14 registered (jurisdiction, coverage, claim-state-exact, attestation, access, attorney-finalization, citation-label, execution, upl, required-facts, conflict, currentness, court-form, upl-capability).