Trial Balance Research Lab

Why it matters

Smart models still fail at everyday work.

60%

of realistic daily finance tasks failed by leading language models in published research.

$2B

gross annual run rate of one expert data network, most of it from frontier labs.

3 to 5x

expected growth in lab spending on RL environments, led by white-collar work.

Benchmarks in development

Results publish as each completes. Labs can test early under NDA.

In developmentQ1 2027

Month-end close inside a live ERP, from bank feeds to a balanced trial balance.

SAMPLE TASKPost March accruals for a 45-person logistics company, reconcile two bank accounts and explain every variance over ₹10,000.

In developmentQ1 2027

Back-office work across CRM, tickets, calls and docs, built from licensed company records.

SAMPLE TASKFrom a recorded sales call, update the CRM, raise the right ticket and draft the follow-up the customer was promised.

Data collectionQ4 2026

How much paid "expert" data is quietly written by models, and which signals catch it.

METHODCompare typing rhythm, timing and model similarity across 10,000 submissions from different sourcing channels.

Method

"Close March books…"

Start from real workPractitioners bring tasks from their own jobs.

ledger · vendors · bank feed

Build the worldA realistic company with real-world mess.

assert tb_diff == 0.00

Write the graderChecks where possible, rubrics where not.

human 93% · model ?

Set the human barExperts solve every task first.