Evolution Research

Your traffic in.
A better endpoint out.

Evolution Research post-trains the model behind your AI agent on your own traffic, then hands back a better endpoint and the eval results that show what changed.

Start a pilot
We
Collect

Your app reaches its AI model through one web address: OpenAI, Anthropic, or your own model server. You swap that one address for ours. We pass every call through to the same model, so nothing changes for your users, and we record each task from start to finish, not as separate calls.

base_url = "https://api.openai.com/v1"
base_url = "https://our-endpoint/v1"
We
Diagnose

We look through those trajectories for failure modes and show them in a dashboard, with how often each one happens and real examples.

You
Decide

You review the issues and confirm which ones to fix.

We
Improve

We build evals for each confirmed issue from your own data, then post‑train the model or optimize the harness against those evals.

You get
Ship

You get a new endpoint, plus the eval results that show what changed.

Case study · Financial-model auditing

A better auditing endpoint for a production AI spreadsheet agent

We ran the loop on the agent's production traffic and post-trained the open-weight model behind it. Every comparison below runs both models on the same tasks, in the same environment, graded by the same rubric grader.

Collect

Real conversations from the agent's investment-banking and private-equity users, as full trajectories.

Diagnose

Auditing stood out: users asking the agent to find and fix the errors in a financial model.

Improve

Audit evals and training tasks built from those conversations and from certified financial models; the open-weight model post-trained against them.

Ship

The new endpoint went live to the agent's paying users; later checkpoints keep improving.

1 · Auditing, held-out tasks · latest checkpoint

It audits better, at the level of frontier models

+0.0400.546 → 0.585

On 20 held-out audit tasks, the trained model beats the base model and lands level with Opus 5.5 and GPT-6.1 Sol. It improves or holds on every task family, and it makes the fewest unwanted edits of any model tested.

20 tasks × 3 runs · p = 0.03 · higher on 11 tasks, lower on 5
Why this score holds up
  • Real banking and private-equity audit work: 4 tasks are real users' own requests, the other 16 are built by our task pipeline from certified financial models
  • Varied: three-statement models, LBOs, DCFs, fund cash flows, valuation packs and lender models
  • 3 runs per task, averaged, so one lucky run cannot move the score
  • The same kinds of work we train on, but every task held out from training and checked by workbook fingerprint
2 · The operator's production audit benchmark · shipped endpoint

It carries over to an external audit benchmark it never trained on

+0.190.536 → 0.724

On the operator's own audit benchmark, built separately from a different task distribution, the shipped model finds and fixes more of the planted errors than the base model.

3 tasks × 3 runs · p = 0.03 · 5 of 6 trained checkpoints beat base
Why this score holds up
  • Built by the operator's in-house finance experts and vetted by hand
  • Hard by design: the gate a model must pass before release to enterprise customers who pay $14M
  • Our training data is a different kind of audit (checking against attached documents); this benchmark is blind error-hunting
3 · The operator's general benchmarks · shipped endpoint

General ability goes slightly up, never down

+0.31/ 10 on broad modeling work

On the operator's 66-task financial-modeling benchmark and its 20-task release gate, the audit-trained model scores at or above the base model. Specializing on audit did not cost other skills.

66 tasks · 9.01 → 9.32 · p = 0.05  |  20 tasks · 8.69 → 8.71
Why this score holds up
  • The operator's own release benchmarks, built by its in-house finance experts and vetted by hand
  • The same gates every model passes before release to enterprise customers who pay $14M
  • 66 + 20 tasks across modeling, valuation and reporting work
Base modelAfter our post-trainingFrontier reference
How we measured

Evaluation tasks are reserved from training and checked against every training set by workbook fingerprint. Both models run in the same agent environment; one frozen rubric grader scores every run, and full traces are kept for every rollout. p-values are one-sided paired tests; taken together the three results give p = 0.006 across 100 tasks.

Start a pilot

Try it on one of your agents

A pilot covers one agent. We agree the success test with you before any training, then report the before and after on your own evals.

Or write to hello@evolution-research.com