datapce readout: The datapce harness

Research equipment, not the product. It runs the agent work behind the results on this site.

2 · What happens to it

The datapce harness

The harness is research equipment, not the product. The datapce open research group runs its coding-agent work through it. Today the group is one maintainer, so every result on this page comes from one set of machines.

datapce harness

Run an agent task on the model you pick. Keep a record of what happened.

  • Run. Send one task to the model you choose, hosted or local.
  • Check. Ask a second model family to check the result, when you want it.
  • Record. Keep counts, timing and errors for each task.
  • ESTABLISHED Built. Runs on the maintainer's machines.
  • PLANNED Public code release and the dpce command.
  • MEASURED 3,511 tests collected 2026-10-21 · pytest --collect-only
  • MEASURED 74/74 calls ran on the requested model 2026-10-21 · conformance.jsonl
task inone agent task datapce harness 1 run on the model you pick2 cross-check, when asked3 count time and errors resultfrom the model recordcounts, timing readoutsPLANNED back astest cases
Fig. 3 — One task in. A result and a record out. The harness runs the task on a chosen model and can ask a second model family to check it. It writes counts, timing and errors. Readouts made from those records, and their return as test cases, are dashed (PLANNED). ESTABLISHED 2026-10-21 · read from the code. Solid = built, dashed = planned.
What the harness does not do yet
  • It does not learn the best model for a task. No route was promoted (0 in 3 route tables). MEASURED 2026-10-21 · route_table.*.json
  • No agreement rate between the checking models has been measured.