Firm Memory

A 120B open-weights model beat a frontier model at drafting law. It was not the better reasoner. It was the better informed.

51.8

harness + firm memory

45.1

frontier model

Harvey Labs legal benchmark, 82 criteria. Two independent judges. The rubric never entered the system that did the work.

Run	Score	Cost	Time
Harness + firm memorygpt-oss-120b · open weights	51.8%	$0.033	13m 46s
Frontier modelhigh reasoning · same documents	45.1%	$2.50	18m 00s
Same model, no harnessone call · documents pasted	≈39%	$0.0037	1m 22s

75×

cheaper than the model it beat

+6.7

points, same rubric

What it stands on

2,115

negotiation episodes from the firm's own prior deals

8,262

edges linking position to provision

A pile of past deals is not firm knowledge. It becomes firm knowledge when every position taken can be found, cited and checked.

Why it compounds

71%

of the tokens we generate are reasoning

$0.88

of every $2.50 task, spent on thinking

You buy the reasoning. You keep the answer. The reasoning is thrown away.

A partner writes you missed 9.9, it bars sandbagging, that’s a deal-breaker. It teaches the model once, then dies with the thread. The next matter starts from zero and you buy it again.

Caught instead, that one sentence becomes a rule with a severity floor and a test that runs forever. The next associate never makes the mistake.

What the firm keeps

GraphHow the work is actually done, step by step.

StandardThe playbook nobody ever wrote down.

EvalsProof the next model is better before you switch.

DataTraining data owned outright, not rented.

Every model is a tenant. Swap it the week a better one ships and the firm loses nothing.

Where it lives

Claude, ChatGPT and Google Docs — the places the work already happens. Nothing new to learn, nothing to export.

Captured as it happens

Claude — reasoning Escrow at 9.7 is 5% / 12 months. Playbook floor is 7.5% / 15. Below walk-away.

Claude — reasoning Set-off barred at 9.12, so escrow is the only liquid source. Compounds with the shortfall above.

Lawyer — decision “Change 90 days to 90 business days — we never agree calendar on a cure period.” Associate, in chat

Google Docs — edit within ninety (90) ~~days~~ Business Days of the Closing Date Partner · suggestion · accepted

→ becomes a rule Cure periods are stated in Business Days. Calendar days are a High finding. Applies to every matter from now on

Three streams, one graph: what the model worked out, what the lawyer decided, and what the partner actually changed.