Poisson Tasks is out. Find where your agent breaks before your users do. Read the launch
Poisson Prove is out. It tells you whether a change to your agent really changed how often it fails. Read the launch
Poisson Labs
Talk to us

Poisson LabsRL and post-training

Train agents against checks they can't game.

Static task sets saturate, and policies learn to game weak rewards. The gym is in development; Tasks and Prove run today.

Get started Talk to us
calibration card · gpt-5-nano
DEPTH · dependent calls d1 0.55 d2 0.40 d3 0.24 d5 0.18 d8 0.87 DISTRACTOR · of 16 tools r0 0.92 r2 0.90 r4 0.91 r8 0.91 r14 0.96 0%25%50%75%100% PASS@1 · 60 TASKS × 5 ATTEMPTS PER LEVEL · TASK-RESAMPLED 95% ON STRICT
strict: reached the goal and stopped there leaky: reached the goal, and may have gone past it
calibration card · Qwen3-8B
DEPTH · dependent calls d1 0.00 d2 0.40 d3 0.82 d5 0.86 d8 1.00 DISTRACTOR · of 16 tools r0 1.00 r2 0.99 r4 0.99 r8 0.98 r14 1.00 0%25%50%75%100% PASS@1 · 60 TASKS × 5 ATTEMPTS PER LEVEL · TASK-RESAMPLED 95% ON STRICT
strict: reached the goal and stopped there leaky: reached the goal, and may have gone past it

8,400 seeded episodes, one dial at a time, on two models. These measure the tasks and the instrument: no RL training result exists yet, and none is shown. The cards

01The closed loop

Practice comes from where the last checkpoint failed.

The gym's architecture, built on the four tools. In development, nothing measured yet.

  1. Tasks

    Seeded worlds, the tools the agent acts through, and the hidden checks that give the reward.

  2. Break

    Searches for where the current checkpoint fails. Those failures become the curriculum.

  3. Trace

    Reads every recorded episode and says at which step the remaining failures happen.

  4. Prove

    Compares each checkpoint with its starting point on held-out conditions training never touched.

The gym page →

02What runs today

Start with the tasks and the comparison.

TasksAvailable

Seeded worlds with hidden checks.

One dial at a time, on its own domain or your own tools, with an export to Prime Intellect verifiers.

poisson-tasks · the Petstore API · gpt-5-nano
A Tasks card: strict pass at one, at faults 0, 1, 2 and 4, with 95% intervals.
One run on simulated replies: 10 tasks a level, 40 episodes, $0.015. An example, not a result about the model.

ProveAvailable

Checkpoint against checkpoint.

The same conditions on the same seeds, so a rising reward has to show up as a change you can defend.

shell · poisson-prove
pip install poisson-provepoisson-prove init --example# 1. sweep the conditions, before the changepoisson-prove sweep --yes# 2. sweep again after it, on the same seedspoisson-prove sweep --yes \    --target poisson_prove.examples.mock_agent:rollout_patched \    --pair-with latest# 3. compare the pair into report.md, heatmap.svg, receipt.htmlpoisson-prove report --event violated --x tool_fail_rate --y task_id# in CI: exit 4 if any condition came back significantly worsepoisson-prove check --event violated --fail-on worse

MATH ASINFRASTRUCTURE.

Talk to us