Post-training platform

Post-training data and tests for AI agents

Send your agent’s conversations. Simulate the ones that fail. Score every one. Train on what moved the number, with a test set kept apart.

$ pip install zeroproof

import zeroproof.simulations as zps
# 1. simulate: tools + prompt become a world
data = zps.simulate(tools=my_tools, system_prompt=my_prompt, mode="rl")
# 2. score: your rules, every conversation
scored = data.grade(judge=my_judge)
scored.pass_at # how often it passes
# 3. train: keep what teaches, publish
rows, report = zps.optimize(data, mode="rl")
data.push("support-v1", agent="airline-support", holdout=0.1)
5.0% to 30.0%
open 8B model on a public support benchmark, after one training set
$15
cost of that training set, no human-written example
75.3%
refund decisions right, from a 1B model, at 18 cents per million tokens
3 numbers
per dataset: passes, passes every time, room to improve

Results

Two public results. Everything open.

Open 8B model, customer-support benchmark

tasks passed

  1. No training data5.0%
  2. 1,057 rows written by experts17.5%
  3. 1,057 rows from our simulations, $1530.0%

Same number of examples. None written by a person. Method

A 1B model that checks refund requests

right answers vs cost

100500$0.10$1$10before trainingthe big modelszeroproof-ecommerce-1b75.3% at $0.18 per 1M tokens

Trained inside a simulated marketplace. Model, data and test set public. Method

Datasets

Ready to train. One line to pull.

Support agents, coding agents, and yours. Rows and pass rate on every card. A test set with each one.

import zeroproof.simulations as zps rows = zps.pull("ds_…") # no key needed

Two commands to an API key.

pip install zeroproof && zeroproof signup --email you@company.com