Why scoring is not delegable

The division of labour

Platform owns	The harness (and its method) owns
which tasks, how many rounds	reading whatever it wants from the traces
running the harness, collecting traces	deciding how to modify the harness
all scoring	—
every curve point, ckpt and token	—

Scoring is not delegable. A subject that scores itself makes every cross-method comparison meaningless, so the platform measures every candidate itself, on its own fixed task set, with its own scorer — the same tasks and the same scorer for every method and every control.

That is the whole reason the platform exists as a separate artifact. If the harness under evaluation could score its own output, the reported delta would be a self-report, and two methods could not be placed on one axis.

What follows from it

• The method is external to the harness. It cannot be part of what is being measured, because it cannot own the measurement. See Methods live outside the harness.
• The method is invoked as a subprocess, never imported. An imported module is frozen at load time, and the case this platform exists to measure is a harness that rewrites its own improvement mechanism.
• The platform hashes its own files around every method call. A method that edits its own scorer fails the run: FAILING RUN, the paths named, no curve.jsonl, exit 3.

The full contract is INTERFACE.md.