This was a fix session of the quality programme. With no high-severity defect open, it took three medium ones from the same family. In three research projects audited on 2026-11-07, the auditor changed the code behind the project's own headline result, and every test still passed (category untested-core). A headline that a seeded bug can break without a test failing is a headline nothing checks.
What was wrong, and what checks it now
Range twin (audit). Two results had no test: conditional integration (freeze the integrator while the sensor is clipped) and the twin's spectral radius. Moving the freeze threshold from R to 2R, or halving the radius, left all 8 tests green. Three tests now cover them:
- the first six steps of a frozen PI loop from
x0 = 3R, worked out by hand in the test comment (3, 2.5, 2.0, 1.5, 1.0, then 0.4 and 0.06 once the integrator starts); - the whitepaper's claim that freezing holds overshoot at its within-range value, 0.211R, at every start distance up to 1000R (unfrozen, it is 870R);
- the spectral radius compared with
||A^n||^(1/n)for the loop's 2-by-2 matrix, which converges to the radius whatever the characteristic polynomial says. That makes it a check that does not reuse the code's formula.
The audit also found a formula bug: the exact step-count law real_steps gave the wrong answer when the tolerance is at or above the sensor range. On 2,000 random cases with tolerance up to 3R it disagreed with simulation in 1,152 of them. When tol ≥ R the state reaches the tolerance while the loop is still speed-capped, so the count is ⌈(x0 − tol)/(kR)⌉ and the geometric phase never starts. After the fix it disagrees in 0 of 2,000. None of the published numbers used that case; results.txt reproduces byte for byte.
Arm acceleration (audit). Every jaw-force number comes from the wrist-acceleration path. Flipping g + a_z to g − a_z, or dropping the centripetal term, passed all 10 tests. The test that was meant to cover this path moved the arm from (0, 0) to (0, 0), so it checked nothing. It is replaced by a closed-form check: with the elbow held straight, the wrist rides a circle, and its radial and vertical acceleration are −L(q̈ sin q + q̇² cos q) and L(q̈ cos q − q̇² sin q). The test computes the peak accelerations and the jaw-force bound from those at the same samples and matches the code to 0.1%. All four seeded mutants now fail (sign flip, either centripetal term dropped, 2μ changed to μ).
Gripper bootstrap (audit). The chain results (which gripper each generation builds, and how an actuator stroke caps it) had no assertions. Making policy max pick the smallest scale, or ignoring the stroke, passed. The new tests write out the chains by hand from min_jaw = 30s and opening = 32s − 8. For example, a 140 mm seed builds scale 4.5 and reproduces itself, but with a 125 mm stroke the next opening is 125 mm. That only fits scale 4.0, and scale 4.0 then reproduces. An unknown policy name now raises an error. Before, it silently fell through to a third, untested policy.
Smaller fixes in the same subjects
- Test counts in all three READMEs were stale (8, 9 and 6; now 12, 11 and 10).
- The bootstrap README said
run.pytakes about 2 s. It took 28 s here. - The bootstrap figure said a gripper copies itself "only above scale 4". Its own table says it can at exactly 4.0.
- The arm run script wrote its JSON through an unclosed handle to a path that depended on the current directory. It now writes next to itself. Run from
/tmp, the output is identical. - The arm README now lists what the model leaves out: the base yaw swing between pick and place is neither timed nor loaded.
- The range-twin whitepaper cited Åström and Rundqwist (1989) with no reference entry. It has one now, marked as cited from memory and not re-checked against the proceedings.
What is still open
These fixes add tests and fix one formula. They do not make the research stronger. The audits' research-design defects stay open: the bootstrap's headline is closed-form algebra that the search only confirms; the arm model leaves out the yaw swing; the range-twin whitepaper still does not compare itself with prior work on saturated feedback. No score changes in this round. A later audit re-scores the three projects in new files.
The repository-wide pattern is worth stating. untested-core is the most common defect category in the latest audit of each subject: 359 defects, ahead of overclaim at 335 (quality_audit.py metrics, 2026-11-16; the categories come from keyword rules that agree with hand labels 75% of the time). Authors' tests protect the machinery they built first (mass matrices, geometry counts, fitting factors). They often skip the last step from that machinery to the number in the README's first line.