Tasks/Professional Knowledge Work/Legal

Answer legal questions from the underlying case files

Work transactional legal matters across 745 case documents

legal_matter_caseload_transactionalProfessional Knowledge WorkLegal
instruction.mdthis is what the agent is given

You are the analyst covering six unrelated legal matters. Each has its own folder of case documents under tools/world_corpus/:

folder matter
w419_oilspill M/V Red Room Ohio River oil spill — OPA 90 liability, charter party, marine insurance
w433_seniorliving Magnolia Gardens senior living — resident agreements, California regulatory compliance, evidence
w425_distributions Summit — shareholder distributions, US tax treatment
w431_llc_spa AIAG LLC — member distributions, SPA indemnification
w417_generalcounsel general counsel desk — child support, WARN layoffs, HSR antitrust, LLC matters
w_securities_saas SaaS incident liability, securities disclosure (8-K)

745 documents in total. Task names are prefixed with their matter, e.g. w419_oilspill/World419_DM_01.

There are two sets of questions, and you are expected to work both of them in full.

18 visible questions — 3 per matter. Answer every one of them and submit through matter, which replies with a bare score like 4/6. It never tells you which criteria you met, and never shows you the criteria themselves. You have 90 submissions total — five per visible question — so you can revise and resubmit. Your job on these is to push every one of them as high as you can.

18 held-out questions — 3 per matter. You answer these blind and submit them once, with matter final, which ends the engagement — no feedback, no second attempt. Their prompts are in tools/heldout_prompts.json from the start.

The two halves come from the same six matters and are graded by the same kind of criteria. So whatever you learn on the visible half — how specific to be, how to enumerate, which documents settle which point — is exactly what the held-out half will ask of you. Working the visible questions properly is how you do well on the held-out ones.

The matter tool

matter status                    budget left, best score per question
matter tasks                     the visible questions
matter show <task>               one question in full
matter submit <task> <file>      grade an answer  (costs 1 of your 90)
matter close                     end the visible phase early
matter final [answers.json]      submit the held-out answers — ONCE, terminal

The rubrics live in a separate grading service you cannot read or reach except through matter submit. There is no copy anywhere in this container.

You have 90 submissions in total across the visible questions. Reading the documents costs nothing; only matter submit spends budget.

What to submit

Write your held-out answers into methods/main/answers.json as you go — it is a plain file, fill it in question by question rather than leaving it to the end. It maps each held-out task_name to your answer text:

{
  "w419_oilspill/World419_DM_02": "Neither Star Tankers ... $56,709,300 ...",
  "w433_seniorliving/World433_JS_08": "..."
}

Then run matter final. An unanswered question scores zero on every one of its criteria, so answer all 18 even where you are unsure.

How answers are judged

A question counts only if you satisfy EVERY one of its criteria. Getting 9 of 10 scores the same as getting 0 of 10 — the criteria are all must-haves. The x/y score you get back during the visible phase is there so you can tell whether a revision helped; it is not partial credit.

Each question has binary criteria of the form "States that <some specific finding>". A criterion is satisfied only if your answer actually asserts that finding. So:

  • State conclusions directly, then support them. No greeting, no preamble, no "here is what I found" — none of that satisfies any criterion.
  • Be specific. Exact figures, statutory subsections, policy names, article numbers, regulation citations. A vague characterisation satisfies nothing; the figure or citation the criterion names does.
  • Enumerate completely. Many criteria are one-per-item: every policy that covers a claim, every article that applies, every step of a liability computation. Missing one item costs one criterion. Length is not coverage — a long answer that names three of seven articles scores three.
  • Follow each question's own instructions on form and length. Some want a yes/no per item, some one or two sentences, some several paragraphs. A few ask for a memo or a schedule; give its full text as your answer.
  • Do not hedge. A criterion asks whether you stated something. A conclusion buried in qualifications may not read as stated at all.

What you have

  • tools/world_corpus/<matter>/ — that matter's documents as text, plus _index.json listing every path. Some entries are marked "extractable": false (audio, images) — those have no text and cannot be read.
  • tools/heldout_prompts.json — the 18 held-out questions.
  • methods/main/answers.json — where your held-out answers go.
  • CPU only, Python 3. Do not fetch anything external.

Metric

pass@1 over the 18 held-out questions · higher is better

Fraction of held-out questions whose every binary criterion passes, judged by Gemini 3 Flash on sealed rubrics.

anchorvisible setheld-outreward
Bno answer submitted0.00000.00000.00
Sgold answers under the same judge1.00001.00001.00
normalisation
m <= B0
m > Bmin(1, (m - B) / (S - B))

m = this run's held-out metric  ·  B = no answer submitted  ·  S = gold answers under the same judge

B = 0, S = 1, so reward = m. m is pass@1, not criteria fraction: a question scores only if all of its criteria pass.

Rollouts

123 minwall clock
$50.89spend
65.6Mtokens
7versions, 7 kept
0 20 40 60 80 100 $0 $15 $30 $45 cumulative spend on the run visible criteria met, best per question (of 137) gold answers under the same j… · visible · 1.0000 v0 v1 v2 v3 v4 v5 v6
keptrevertedno scoreturning point
  1. v0Baseline: 18 blank held-out slots, protocol fixed before any submissionPick held-out answers by source-checklist completeness, and never carry a rubric guess from the visible half into a blind answer.no submission yet1 min · $0.40
  2. v1Oil spill: party roles, current 33 U.S.C. 2704 cap ($1,900/GT, $56,709,300)1210 min · $3.94
  3. v2Senior living: manual/regulation mapping, awake-coverage, expert-report flipsFlip one conclusion at a time and resubmit: the bare x/y names no criterion, but a controlled flip localises what it pays for.3426 min · $9.54
  4. v3Summit S-corp: 1362(d) termination separated from 1361(b)(1) eligibility5350 min · $17.81
  5. v4AIAG lease and LLC: roof-code shift to tenant, post-death consent invalid6766 min · $22.52
  6. v5General counsel and securities/SaaS; all 18 held-out answers written101100 min · $37.95
  7. v6Falsification sweep on earlier matters; held-out JSON identical to v5Spend the last budget trying to break the standing answers, not raise them, and transfer nothing a flip failed to confirm.101123 min · $50.78

Seven snapshots, roughly one per matter plus a closing audit: 123 min, $50.89, 77 of the 90 visible submissions spent.

On the hidden set

held-out metricreward
no answer submitted0.00000.00
gold answers under the same judge1.00001.00
this run3 / 18 = 0.16670.1667