How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Written 2026-08-05, before the replacement was measured. States the defect, the fix, and the rule for re-deriving the threshold, so that the value cannot be said to have chosen the rule.
cv_gain is the gate's negative evidence: a relational claim is refused where a
constant-velocity model already accounts for the motion, which is what keeps
Paper 003 from collapsing into Paper 002. The ceiling is 0.30.
On a steadily-closing pusher — the control built for exactly this clause — the statistic decays monotonically with the horizon:
| horizon | 4 | 5 | 6 | 7 | 8 | 10 | 12 |
|---|---|---|---|---|---|---|---|
cv_gain |
+0.577 | +0.490 | +0.406 | +0.314 | +0.219 | +0.014 | −0.191 |
It crosses the ceiling between 7 and 8 while the motion is unchanged. The same trajectory reads as "a constant-velocity model explains this" at horizon 6 and "it does not" at horizon 8.
The cause is in the statistic. It takes the last one-step velocity and
extrapolates it horizon steps, then compares that error to zero-order's. On
any curving trajectory a longer extrapolation overshoots further, so the
constant-velocity model looks worse the further ahead it is asked, and the
number falls. Two different things are being measured at once: whether the
motion is smooth, and how far ahead the question was asked.
The threshold was calibrated at horizon 6 and inherits that horizon silently.
That was invisible while dispense_latency never moved.
The physical scene forces dispense_latency to 8 — the block carries at
2.86 mm/step at the tenth percentile, so six steps move it 17 mm against a 20 mm
tolerance and the task is trivial.
The derivation
So a parameter fixed by physics and a threshold calibrated at a different value of it came into conflict. Neither is wrong; the statistic is. A threshold should not change meaning because an unrelated parameter moved.
Ask the same question one step ahead:
cv_gain = 1 - mean|x(i+1) - (x(i) + v(i))| / mean|x(i+1) - x(i)|
How much of the next step's motion does a constant-velocity model remove, against zero-order. The horizon does not appear, so it cannot leak in.
The meaning is preserved: high where motion is smooth and persistent, near zero or negative where it is intermittent or reverses, which is what the clause has always been for.
The old 0.30 does not transfer — a different statistic needs its own value, and carrying the number across would be the mistake this document is about.
The threshold is derived the way min_proximity_contrast was, and that
precedent is explicit in the code: "Every value from 0.30 to 0.90 separates the
treatment from all three controls identically... 0.50 sits in the middle of the
plateau."
So:
drift, slide, static, noise — across a range of horizons, on CPU.No value is chosen by looking at which one lets a result through. If no plateau exists, that is reported rather than resolved by picking a number from the range.
The other gate thresholds, min_proximity_contrast and
min_post_contact_far_deltas and min_carriage_run, are untouched. The
tolerance, the commit window's form, and min_ride_steps are untouched. This
replaces one statistic that was measuring the wrong thing and re-derives one
threshold that belonged to it, and the preregistration's list of tuning that
will not be done holds in full.
Every result derived with the old statistic is re-derived, not carried over.