After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

A carry keeps its distance; the statistic only checked its steps

Open finding, 2026-08-05. Nothing is fixed yet. Written at the end of a session so the next one starts from it rather than rediscovering it. It invalidates results recorded earlier the same day, including "24 of 24 cells produced a capture".

What was seen

In the Isaac pilot at the derived latency and radius, every cell was verdicted CAPTURE and arm D never once acted. The gate was refusing, and the refusal was traced to the contact requirement on the carriage path — the agreeing body was not within the interaction radius.

That looked like the requirement being too tight. It is not. Measured:

grasp closed at post-grasp separation, median
seed 300 0.65 mm 3.35 mm
seed 301 0.67 mm 49.83 mm
seed 302 0.77 mm 62.85 mm
seed 303 0.64 mm 54.88 mm
seed 304 0.75 mm 177.75 mm

The object is not being held. The arm and the object drift tens to hundreds of millimetres apart after the grasp.

Why the statistic said otherwise

_ride_mask compares per-step displacement vectors and admits a mismatch of up to agreement — 25% — of the target's own step. At roughly 3 mm/step that is 0.75 mm per step of permitted disagreement, and sixty steps of it integrates to forty-five millimetres.

So a target that moves almost with the arm, every step, and steadily falls behind, scores 0.98 agreement over a run of 111. That is slipping, not carrying.

The gate was right and the verdict was wrong. The contact requirement added earlier the same day is the only thing that checked the accumulated quantity, and it caught exactly what it was written to catch.

The shape of the defect

The same shape as three others found today: a statistic that is correct locally and wrong when integrated. cv_gain measured the horizon along with the motion; the pre-roll fixed max error while the median was a different quantity; the capture-radius p90 measured the sweep's own grid.

Carrying is a statement about relative position being maintained. Per-step agreement is a local proxy for it, and the proxy does not imply the thing.

What this invalidates

The capture does happen — the probe held an object with 0.25 mm of disturbance and carried it. What is in question is how many of the cells scored as captures were carries rather than slips, and for how long they held.

What the fix has to be

carriage_evidence and capture_verdict must require the separation to stay bounded through the riding run, not merely the steps to agree. The gate already does this via contact; the verdict and the evidence statistic have to follow it, not the other way round.

Decided 2026-08-05, before re-measuring: the measured capture radius

Of the three candidates, this one introduces no new parameter, which the others do. The grasp separation needs a tolerance band around it; a fraction of the total carry needs the fraction, and would call 20 mm of drift over a 200 mm carry a success.

More to the point, it is not really a choice. capture_verdict was trusted for one reason, stated in its own docstring: it decides "with the same statistics the gate and arm D use, not with a separate rule invented for the pilot". That property had quietly broken — the gate called carriage_evidence with the radius and the verdict called it without — and the verdict then certified exactly the cells the gate was refusing. Restoring the invariant is the fix; the radius was already fixed by measurement at 2.5 mm.

Where the bound had to go

Three attempts, and the first two moved numbers without moving the verdict:

  1. Passing the radius to carriage_evidence from the verdict changed only the reported agreement. The verdict branches on estimate_capture, not on that number.
  2. Passing it to estimate_capture too was still not enough. It finds the first run of three agreeing steps, and the first three steps after a grasp are inside the radius however badly the carry later comes apart.
  3. It belongs in _ride_mask, which all three callers share, and in what held means. Held is a statement about now: an object the carrier has drifted away from is not held, however well each individual step agreed. The verdict now requires it.

The first two are recorded because each looked like the fix and neither was.

Second attempt: the bound was the wrong kind

The first fix bounded the separation by the capture radius, and that is wrong in a way the data showed immediately. A carry means the separation is bounded, not that it is small. Measured over eight cells, one held at a perfectly constant 3.35 mm — drift +0.004 mm/step, flat to three decimals — and the 2.5 mm radius rejected it. ee_frame is a virtual point between the jaws, so an object genuinely held sits a few millimetres from it by construction.

The other seven drifted at 0.44 to 1.23 mm/step, out to 68–209 mm. Those are slips and the first fix was right about them. Both kinds are present, and a test that cannot tell a constant 3.35 mm from a growing one is not measuring carrying.

The statistic, from the principle rather than from those eight cells

A carried object inherits its carrier's displacement. Inherit all of it and the separation is constant; inherit a fraction and the rest becomes separation. So the quantity is the share of the carrier's motion the target failed to inherit, over the run:

slip = (separation at the end − smallest separation in the run)
       / (distance the carrier travelled over the run)

Dimensionless and free of scale: 0 is a perfect carry, 1 is an object left where it was. The baseline is the run's own minimum rather than its first step, which skips the settling just after a grasp without introducing a number for how many steps that takes.

The bound is agreement, which already exists. It declares that a mismatch of up to a quarter of the target's step still counts as riding. The defect was never that quarter — it was that a per-step allowance accumulated without limit. The same quarter applied to the run is a bound, and no parameter is added.

This has to be confirmed on cells I have not seen

The eight cells above were examined before the statistic was written, so the statistic is not independent of them however carefully it was derived from principle. It is validated on fresh seeds, and if it disagrees there, the disagreement is the result.

Third position, and the one that holds: three questions were being conflated

The statistic above does not survive its own first test, and the reason is that the whole line of reasoning was answering the wrong question.

A synthetic slip of 2 mm per step against a 15 mm/step carry is 13% — inside the agreement bound the statistic proposed. Tightening the bound until it fails would have caught it, and would also have caught real Isaac cells slipping at 14%, which is where the question should have been asked instead: is a cell that slips 14% unusable?

It is not. Arm D predicts the target by rolling the carrier forward. A target inheriting 86% of its carrier's motion is mispredicted by 14% of what the carrier travels over one dispense window — about 3.6 mm against a 20 mm tolerance. Arm D still lands.

Three questions had been collapsed into one:

answered by
Is this trace a capture — still, then riding, effect accumulating? the verdict
Can arm D predict it? the scoring
Should the gate let arm D act? the gate

The verdict was built for the first. The contact bound was added while thinking about the third. The slip discussion was really about the second — and the second is measured, not asserted in advance.

By the design's own definition a slipping carry is a capture. Capture names two properties: the target is perfectly still before the arrival, and the effect then accumulates without bound. A target inheriting 86% of its carrier's motion has both, and travels 157 mm doing it.

So what was actually wrong

Only one thing, and it was in the gate: the contact requirement was applied at every step. An object genuinely held sits a few millimetres from ee_frame, which is a virtual point between the jaws, so per-step contact rejects real carries — including the one cell holding at a constant 3.35 mm.

Required once in a run, it does the job it was added for: drift agrees with its body step for step, has zero slip, and never comes within 183 mm of it, so it has no moment of contact anywhere in the run.

Measured after the change, on CPU: treatment 1.00, drift, static and noise 0.00, slide 0.05 — still under H3's ceiling. And the earlier "24 of 24 captures" stands after all.

What this episode cost, recorded

Three positions in one afternoon on the same question, and the middle one was wrong in both directions — it rejected real carries while being too loose to catch the synthetic slip it was written for. The tests hold all three, because each looked right at the time and the next person will reach for them in the same order.

Where things stood

Everything else is committed and green: 244 tests, the arms and gate re-derived at latency 8, the SELF comparison passing Case A a second time on fresh seeds with the competitor scoring 0.045 and taking 8 discordant pairs.

The Isaac side owes the preregistration three numbers — engagement, normal_alignment, observation noise — and cannot supply them until a cell can be trusted to be a carry.