Fast Harness

Reviewing every step is accurate — but expensive and redundant.

In the baseline loop, the teacher drives the robot chunk by chunk: every short action segment is a fresh, costly model generation. Cost scales with the number of steps and never goes down — the thousandth repetition of a solved task costs exactly as much as the first. And most of that spend is redundant: a familiar task repeats the same easy, already-mastered stretches, yet each one is re-reviewed at full price, no differently from a genuinely uncertain moment.

The missing capability is judgement about when the teacher is actually needed. The baseline loop has no way to tell a trivial, well-practised step apart from a risky one, so it plays it safe and reviews everything. What if the harness could instead learn, per task, which steps it can safely release on its own — spending GPT-6 only where it truly matters, and cutting calls, tokens, and wall-clock time while holding task success? That is exactly the loop we build next.

An experience-driven execution loop.

The harness loop sits between the student policy and the teacher. Each chunk, it takes a fresh student proposal and decides — from accumulated experience — whether this chunk needs a review or can be released on a zero-call fast path. Only unfamiliar, risky, or previously-failed moments pay for a review.

↔ scroll to see the full diagram

Inside that loop, two cooperating mechanisms do the real work, and they are clearer looked at apart. One learns to review less over time; the other decides what to do when a review catches a failure.

Getting faster: learning which reviews to skip

The efficiency gain rests on a single principle: reuse verified experience rather than re-verifying it at every step. The harness decomposes each task into a small set of recurring stages — approach, grasp, lift, place — and maintains, for each stage, an empirical estimate of how reliably the student policy has completed it. Stages with a consistent record of success are permitted longer unsupervised execution; stages without such a record are not.

At every chunk the harness resolves a single question: whether the current step requires teacher review, or can be executed on a zero-call fast path. For a stage with an established record, the interval between reviews is allowed to grow. This interval is governed not by the raw success rate but by a conservative estimate of it — one that discounts for how few times the stage has been attempted — so that a handful of favorable outcomes does not, on its own, widen the interval.

These stages are not labelled in advance. The harness recognizes one by resemblance: it keeps a small set of remembered examples for each stage and matches the current observation to the nearest of them. A step that resembles no remembered stage — or sits ambiguously between two — is treated as unfamiliar and reviewed regardless of any record, so a genuinely new situation cannot inherit another stage's permission to skip.

Three mechanisms make this improvement cumulative rather than transient, so that repetition yields improvement and not merely speed:

Skipping is nonetheless the exception rather than the default. A fixed set of conditions forces review regardless of a stage’s record: an unfamiliar or newly entered stage, a contact or alignment moment, proximity to a past failure, a stalled trajectory, or a bound on the number of consecutive unreviewed steps; a small fraction of chunks are also reviewed at random as an audit. A review is skipped only when none of these conditions holds.

Recovering when something goes wrong

Skipping is safe only if the failures that do occur are detected and corrected. When a review reports a failure — an execution error, or an action inconsistent with the intended goal — the harness leaves the fast path and reviews every chunk until the fault is cleared, after which normal skipping resumes. Correction escalates only as far as it needs to, within a bounded step budget; if the budget is exhausted without the task getting back on track, the episode is ended rather than left to run on indefinitely.

It gets faster with experience.

Across repeated episodes of the same task, the trend is consistent: the number of teacher calls falls while the review-skip rate rises, and both level off at a stable value. The first episode is reviewed in full by design, so every curve begins at a skip rate of zero.

↔ scroll to see all four tasks

Fewer reviews, fewer tokens.

A skipped review is a teacher call that is never made, so the token cost recorded per episode decreases with experience — by roughly two-thirds to three-quarters across the four tasks.

↔ scroll to see all four tasks

Success is preserved.

The purpose of skipping is to reduce cost without reducing reliability. Over the same episodes in which reviews fall by roughly two-thirds, task success is maintained, and the harness remains competitive with published methods on every task the benchmark reports.