datapce readout: See what your coding agent did.

Read-only. On your machine. Nothing installs into your agent. datapce-scan reads Claude Code logs and counts what ran. It sends nothing.

Code: codeberg.org/datapce/datapce-scan · MIT · MEASURED 2026-10-22 · SCAN2

In 48,982 past Bash calls of one operator, a naive rm -rf / prefix rule fired 257 times. A root-only rule fired 0 times. MEASURED 2026-10-22 · SCAN1

See the replay

The scan is part of open research on coding agents, failures included. Intake is PLANNED. Do not e-mail a readout. A readout will be checked and scrubbed on your machine, reviewed by a person, then used as test cases, labels and public counts.

One maintainer
Code release PLANNED for change-chain and the harness
Readout intake PLANNED

Fig. 1 — The research loop. Readouts of agent work go in. The harness and small models use them. Results come out. Dashed = PLANNED. ESTABLISHED 2026-10-21 · read from the code and readout-v0.schema.json.

Readout

One file of counts, labels and short checked sentences. No command text, no code. No outside readout exists yet.

What is in a readout

Harness

Research equipment, not the product. It runs the agent work behind the results. 3,511 Python tests. MEASURED 2026-10-21 · H2

How it works

Results

Failures come first. The JEPA failed its gate. change-chain linked 87.7 % of 57 commits to their session. MEASURED 2026-10-22 · CC10

What came out

What this page does not claim

One maintainer, one set of machines. The numbers say nothing about how other people work.
The models are small research fits, and the JEPA failed its gate.
H1, a demand test, was stopped on day 2 of 21. It has no verdict. Study record

Check every number