The oracle
the verifier compiled from one recorded step · every hash checkable
The demonstration’s fourteen recorded steps contained exactly one moment that changed what the software does. The world model’s job was to find it, rank the steps, prune the reads, the failed attempts, the wandering, and point at step six. Replay’s job was to recover, byte for byte, what that step wrote. What you are looking at below is what fell out: a verifier. An oracle for this task, built from the work itself, with no human authoring a check.
The checks
Replay recovered the exact lines the fix added and deleted. Their content hashes are the verifier, four checks, acceptance threshold 1.0:
class-access-add add boltons/cacheutils.py 61b251087294… instance-access-add add boltons/cacheutils.py fe4d219645d8… class-access-delete del boltons/cacheutils.py bc0ce8273796… instance-access-delete del boltons/cacheutils.py fa79aff48ccf…
An attempt scores by producing those line-level effects, wherever and however it edits. The do-nothing candidate pays 0 (its diff matches no check). The recorded fix, the oracle candidate, pays 1.00.
The oracle run
The oracle candidate was replayed through Harbor, the open agent-rollout harness, inside Docker with the network off:
harbor 0.16.1 · oracle agent · docker · network none 1 completed trial · reward 1.00 · 0 errors model: none · cost: none
model: none is the line that matters. The world model ranked the corpus and
chose the moment; it never grades. Reward comes from replayed truth only, which is
why a model miss (the demo shows two, deliberately) costs a candidate task, never a wrong
reward.
Five artifacts make up the task (the environment, the instruction, the oracle patch, the task config, the verifier); each is hash-pinned, so a rerun either matches the receipts or refuses. Multiply by a corpus: the model ranks a million recorded steps for consequence, replay verifies the ones that matter, and environments like this one fall out of work that already happened, no task authors, no grader engineers, no QA team reviewing both.