Agent environments, graded to the cent.
A test environment for AI agents on pharma revenue work. Plug your agent in, run the tasks, and get a score on every decision it makes: chargebacks, 340B rebates and Medicare price refunds.
If the ledger doesn't balance to the cent, the episode scores zero.
- Your Agent, Our Tasks
- Scored to the Cent
- Runs on Your Machines
One environment, a simulated manufacturer desk, sends its records to the AI agent: chargebacks, 340B rebate claims and Medicare price refunds in one queue. The agent's actions go to the grader. The grader's score lands in the ledger, then the next episode starts.
The Scoreboard
Six of seven top models slipped on the rules.
Same tasks, same grader, every model. Each bar is one task, filled white to its score; orange is the share the model got wrong.
See Where Each One Slipped- GPT-6 Astra: 12 of 12 tasks fully right; every decision right on every task; average 1.00.
- GPT-6.1 Sol: 11 of 12 tasks fully right; skipped the HRSA report on 7 of 74 claims; average 0.99.
- Gemini 4 Carbon: 3 of 11 tasks fully right; posted the same ledger credit twice on 7 lines; average 0.96.
- Claude Fable 5.2: 7 of 12 tasks fully right; skipped the HRSA report on 24 of 74 claims; average 0.96.
- Claude Opus 5.5: 0 of 12 tasks fully right; skipped the HRSA report on all 74 claims; average 0.87.
- DeepSeek V4 Pro: 1 of 12 tasks fully right; missed the claim deadline in 4 of 12 tasks; average 0.53.
- Gemini 4 Hydrogen: 0 of 12 tasks fully right; rebilled 119 chargeback lines where a credit memo was due; average 0.39.
Why Now
The deadlines aren't yours to set. The errors are yours to answer for.
Medicare's negotiated prices started on January 1, 2026. The 340B rebate pilot starts on January 1, 2027, for drugs selected for Medicare price negotiation in 2026 and 2027: big brands heading for their patent cliffs.
- You build or deploy the agent.
- Your client runs it on their books.
- Pennyproof grades it first, before it touches your client's books.
- $81.4B
- Bought under 340B in 2024, up from $53.7B in 2022 (91 FR 48886)
- Up to $1.5B
- A year in duplicate discounts: a 2019 manufacturer estimate, when the program was less than half today's size (91 FR 48887)
Source: HRSA's Federal Register notice on the 340B rebate pilot, FR Doc. 2026-15633, 91 FR 48884–48887.
In Plain Terms
Everything flows one way: to you.
What It Is
Simulated pharma revenue tasks and a grader that scores every decision. A test environment you take in-house.
What You Get
A private repo of tasks, the grader's readable code and a Docker image. Run it against your agent as often as you like.
What Stays With You
Your systems, agent and data stay on your side. Every record is synthetic. It runs offline; sharing a score is your call.
Nothing flows back to us.
The Cost
One wrong call costs money, or breaks a federal rule.
A chargeback credited twice. A 340B rebate denied that the pilot rules say must be paid. A duplicate discount nobody reported to HRSA. Agents make these calls at volume, so their mistakes add up at volume too.
The Desk
A ready-made back office, ready on day one.
A complete simulated manufacturer desk: claims, chargebacks, contracts, Medicare refunds and Medicaid data in one queue, set up and waiting. The hard cases are planted from the published rules, and each one keeps the citation that decides it.
The Agent
Point your agent at it. Let it work the queue.
Any agent, any model. It reads the records, then pays, denies, returns, holds or escalates each item, against the clock. Some cases are traps: a denial that looks right and is forbidden under the 340B pilot.
The Grade
Every decision scored, with its reason.
Right call, right amount, right reason: full credit. A forbidden denial costs a full point. If the ledger doesn't balance to the cent, the episode scores zero.
The Proof
A score you can show a client.
What the agent read, what it did, how each item scored and which rule decided it. The grader's code ships with the tasks, so every score can be checked.
The Scoreboard
Where each model slipped, task by task.
The same 12 tasks and the same grader for every model. Only one got every decision right. The others skipped the HRSA report, posted a credit twice, rebilled where a credit was due, or missed the claim deadline.
Task Fully RightOrange: The Share It Got Wrong
Each bar is one task, filled white to its score. The average is the mean of the bars, so a model can average high and still miss a rule in every task.
- GPT-6 Astra: 12 of 12 tasks fully right; every decision right on every task; average 1.00.
- GPT-6.1 Sol: 11 of 12 tasks fully right; skipped the HRSA report on 7 of 74 claims; average 0.99.
- Gemini 4 Carbon: 3 of 11 tasks fully right; posted the same ledger credit twice on 7 lines; average 0.96.
- Claude Fable 5.2: 7 of 12 tasks fully right; skipped the HRSA report on 24 of 74 claims; average 0.96.
- Claude Opus 5.5: 0 of 12 tasks fully right; skipped the HRSA report on all 74 claims; average 0.87.
- DeepSeek V4 Pro: 1 of 12 tasks fully right; missed the claim deadline in 4 of 12 tasks; average 0.53.
- Gemini 4 Hydrogen: 0 of 12 tasks fully right; rebilled 119 chargeback lines where a credit memo was due; average 0.39.
Mean score on 12 held-out release-one tasks per model (Gemini 4 Carbon on 11), run 12–13 Nov 2026. Same tasks, same grader.
Who It's For
For anyone putting an agent near pharma revenue.
Use it to test an agent before a client's books see it, to compare models, or to train on hard, verifiable tasks. You keep your agent; we provide the desk and the score.
Consulting and Implementation Teams
Deploying agents for life-sciences clients? Test them on a full revenue desk and bring the score to the steering meeting.
Agent and Revenue-Software Vendors
Show pharma clients how your agent scores before it touches their chargebacks and rebates.
AI Labs and Environment Vendors
Harbor-format tasks with rewards you can verify to the cent, ready for your own harness. Long-horizon work in a regulated domain few others cover.
How to Buy
Start with a pilot. Add what you need.
The tasks run in your environment, against your agent. Every order starts with a pilot on held-out tasks; add more tasks, hosted evaluations of more agents, or a yearly refresh as the rules change.
50 held-out tasks and the grader's code, yours to keep for testing, plus one hosted evaluation of your agent with a failure readout.
- Request AccessTell us what you want to test, using the form below.
- ChooseA short call to pick your menu. Every order starts with a pilot.
- SignA mutual NDA and a one-page order form, signed online.
- PayBy invoice: bank transfer, wire or card.
- RunA private repo of tasks and grader code, and a private Docker image you run on your own infrastructure.
- Read OutA scoreboard and failure report, and a call on where your agent broke.
See where your agent breaks.
Bring any agent. Request access and we'll send a login to the replay viewer and the full release-one scoreboard.