For judges: Check Every 1stRun Claim in 60 Seconds

Watch the 2:54 film Check the evidence See v1 against v2 Open a real receipt

Each Claim Has Proof You Can Open

1The agent improved
2 of 4 quickstarts finished, then 3 of 4.
2The agent recovers from failures
Step 1 failed, then 4 of 4 passed clean.

The Agent Runs a Quickstart Start to Finish, and v2 Beat v1

Same 4 quickstarts, v1 against v2

Finished, of 4

v12

v23

+1

Failed steps recovered

v10

v23

+3

Fixes verified clean

v10

v21

+1

Cost per run

v1$0.0262

v2$0.0200

-24%

Recovery in one run

step.1failed: curl: command not found
recover.1installed curl
step.2failed: ~/.local/bin not on PATH
recover.2added it to PATH
verifyfresh container, 4 of 4, no model calls

Show details

Every part of a run points to code

Plan: copies each docs command. planner.py
Tools: Firecrawl, Docker and a judge model. classify.py
State: saved after every step. state.py
Context: reads only the quickstart page. planner.py
Run: one fresh container per run. executor.py
Recover: classify, patch, retry, 3 tries. recovery.py
Verify: second container, no model calls. verify.py
Ask once: a human is asked only below 0.7 confidence. recovery.py precedents.py

All six numbers

Metric	v1	v2
Finished, passed or verified	2 of 4	3 of 4
Failed steps recovered	0	3
Fixes verified clean	0	1
Tool calls per run	4.25	6.0
Cost per run	$0.0262	$0.0200
Run time, verify included	128 s	201 s

List price, not billed.

The 2:54 Film Shows the Problem, Doc-test and Doc-audit

2:54

Watch the film The 2:59 long cut Doc-audit film, 1:56

Real Receipts and Real Audits You Can Open

Upstream pull requests and issues from onestrun runs will be listed here.

76checkable receipts 7repos audited

Show details

1stRun receipt --check fails when a receipt or its evidence changed. A run receipt An audit receipt Receipts re-issued Oct 11 from saved run state.

Each Doc-audit board maps who works on what, from 14 to 247 people. dbt-core shows 15 stale pages and 72 features with no docs. dbt-core board

Older audits show an estimated cost, $0.0001 to $0.0055. How it is measured

Verify the Repo From Your Own Terminal

First commits and their dates

$ git clone https://github.com/ELiTE0005/1stRun && cd 1stRun
$ git log --reverse --date=iso --format="%h %ad %s" | head -3

Offline tests

$ pip install -e ".[dev]" && pytest -q

No keys needed. Live tests are skipped. Actions

Run it on any quickstart

$ pip install git+https://github.com/ELiTE0005/1stRun
$ 1stRun https://docs.example.com/quickstart

Python 3.11, Git, Docker, a Claude token. README