Today's models aren't ready for regulated work. We help close the gap.
Benchmarks that show where they fall short,
and environments to train on that work.
We build companies. Not prompts: whole companies, with the files a regulator would ask for. Havenor Therapeutics has 789 of them, 120 colleagues, and the usual mess: superseded SOPs, executed copies, reports that disagree with their own protocols.
Then an expert gives a model a real assignment. In Task 03: review the sterile-filter validation and write the brief an inspector will read.
The model wrote a fluent brief and caught a 30-minute window the approved protocol never contained. Here is a check it missed.
AVT-VAL-106 · Wetting & integrity report
Limits: bubble point NLT 3.45 bar, diffusive flow NMT 18.0 mL/min. Criterion 5: “Integrity result remains correlated to the bacterial-retention qualification.”
?Correlated by what?
AVT-VAL-107 · Bacterial retention report
Bubble point before and after challenge on three filter lots. No diffusive-flow result.
§The only retention evidence.
Model's brief · Claude Opus 5.5 · Claude Code
“Bubble point ≥3.45 bar / diffusion ≤18.0 mL/min are valid release limits”
✗Credited both limits. Only one is tied to retention.
Nothing in that line reads wrong. It is wrong because a criterion signed Pass is not the evidence behind it. Every task we write turns on a judgment like this, and every criterion is checked by someone who has made it for a living.
Biopharma Bench V0.2: 81 tasks across 24 companies, 908 criteria, grouped by the kind of work. Each bar is one task; full height means every criterion met.
- 1GPT-6 Astra65.1%
65.2%
62.4%
58.9%
72.1%
72.4%
67.0%
- 2Gemini 4 Carbon64.6%
68.9%
59.2%
62.7%
60.8%
59.4%
62.9%
- 3GPT-6.1 Sol63.1%
64.1%
57.9%
61.4%
72.6%
65.7%
56.5%
- 4Muse Wonder 1.061.4%
63.7%
58.0%
57.5%
63.5%
60.1%
59.7%
- 5Grok 4.761.1%
58.3%
58.1%
65.0%
65.0%
72.1%
65.4%
- 6Claude Fable 5.160.0%
59.5%
58.5%
59.2%
59.1%
62.7%
68.7%
- 7Gemini 4 Hydrogen58.2%
61.6%
50.9%
61.3%
61.2%
42.4%
57.0%
- 8Claude Opus 5.556.2%
56.0%
56.4%
57.4%
54.1%
45.5%
69.3%
- 9Grok 4.653.8%
53.9%
51.9%
52.3%
54.2%
53.0%
62.2%
- 10Gemini 4 Argon53.2%
55.9%
60.2%
50.0%
38.7%
44.9%
51.3%
- 11Kimi K3.1 Pro50.7%
59.2%
44.2%
54.4%
14.0%
42.6%
62.5%
- 12Claude Sonnet 5.546.6%
47.3%
52.2%
40.2%
43.0%
40.5%
51.0%
- 13DeepSeek V4.1 Flash45.1%
44.5%
45.0%
46.6%
43.8%
43.9%
50.6%
- 14GPT-6.1 Terra40.5%
46.6%
33.7%
31.8%
38.6%
28.3%
47.3%
- 15GLM-5.3 Flash31.6%
32.8%
37.8%
26.6%
24.0%
27.5%
32.9%
The best model, GPT-6 Astra, scores 65.1% on the average task. Grok 4.7 finishes the most assignments outright: 6 of 81. No model does better.
The gap is not between labs. DeepSeek V4.1 Flash, an open model, scores 45.1%; Claude Opus 5.5 scores 56.2%. The gap is between all of them and the work.
Benchmarks, training environments, and deployments for work that lives in thousands of interlocking documents: biopharma dossiers, device files, and EPC project records.
| 01 | DeploymentDomain workflows across biopharma (CMC, quality, regulatory, clinical) and energy & EPC (submittal disposition, MDR/ITP, turnover). | Biopharma · Energy & EPC | ☐ |
| 02 | EvaluationYour data or ours: component task families and end-to-end role seats, co-built with your experts and our Fellows for offline regression and online monitoring. | 81 tasks · 908 requirements | ☐ |
| 03 | Post-training, OTS data & strategyArchitecture diagnosis, containerized RL environments with isolated verifiers, coupled partial-credit rubrics, SFT reference deliverables, and licensed corpora. | 24 drug & device programs · EPC packs | ☐ |
Tell us what you need to measure.
A lab, an AI company, or an expert who wants to write tasks.
Request · RC-REQ
Talk to us