What a run records

A curve point is the unit the whole platform is built to produce. It records the harness state, the two models involved, the score with an interval, and the compute spent to get there. The point is not any single number — it is that the number arrives with what it was produced by.

The curve point

Every round appends one point to curve.jsonl:

Recorded	Why it is there
the harness state (a commit sha)	so a score can be traced to the exact artifact that produced it
both models involved	the agent under test and the improvement model are separate roles
the score with an interval	a bare delta hides its own resolution
the compute spent	a curve that cannot be costed cannot be compared to another method’s
generation_tokens	a declared field, never hardcoded — a zero here is a claim, so it has to be a real one
editable_surface_touched	read from git diff, so the boundary between method and harness is measured rather than asserted
env.harness_runtime	which interpreter ran the harness, and where its dependencies came from

Because the interval is part of the point, the platform states the step size a score can actually move in. A run scored on four tasks says so:

! this run scores 4 tasks — the score moves in steps of 0.250

Two halves on disk, deliberately not in the same place

	where	why
the record	runs/<run-id>/	curve.jsonl and run_meta.json; what happened, kept with the platform
the working tree	../harnessgrad_work/<run-id>/workspace/	the harness copy and every checkpoint commit; outside the platform, so a harness cannot reach the driver or .env by ..

The working tree holds the git history of every checkpoint, so it is not disposable until the run is exported. --work-root moves it, and the driver refuses to start if it would land inside the platform.

Isolation is part of the result

Each task’s harness runs inside a mount namespace, where the platform is not present at all: the path is an empty mount point, not a refused read. --no-sandbox turns that off for debugging, and run_meta.json records which mode a run used.

tools/plot_curve.py labels each run’s isolation and refuses to silently mix them. Runs made under different sandbox plans, or before the sandbox existed, would put numbers on the same axis that were produced under different rules; the tool says so and exits non-zero instead.