Study 11b · Listening
The shipped tempo ladder, validated: the margins are honest, and the confound shipped
Listening, the ladder thread closed
Study Overview
The Ladder Thread Closes Like the Warning Thread Did
Study 6b set the pattern: the production side ships, declines to trust its own work, and asks the bench to validate it. The ladder thread now ends the same way. Stem Shovel built the tempo ladder in production (v0.200.0) in parallel with Study 11, with a sharper subdivision rule than the sketch the bench had tested, and requested validation of the shipped rule verbatim plus a swing-stress sweep. The bench ported it pinned at their commit and ran 52 detector reads (the Study 11 corpus, the real recordings, and 15 new swing-stress items at ratios 0.60 to 0.70 across three tempi) plus 332 wrong-tempo rows through the lesson-six-corrected generator. Zero spend, one disclosed pre-registration smoke, and a thread ledger worth a line: the four ladder-thread studies together cost about eight cents of model spend, all of it Study 9's pilot.
What Validates Clean
V3 passed 15 of 15: not one swing item at any ratio or tempo read compound, every tempo exact, validating the shipped TRIPLET_SHARE of 0.55 and its swing-cannot-exceed-half argument across the full sweep, including the 0.67 triplet swing the bench's own sketch rule had misread. V4 passed 0 of 332 false accepts through the corrected generator, which also retro-confirms Study 11's reading that its by-letter failures were generator collisions rather than ladder defects. And the shipped super-beat re-pick fixed two of Study 10's octave-family misses outright: two 6/8 items now detect their dotted tempo exactly at the detector level.
The Margin That Cannot Be Tuned
V1 failed 2 of 4, and the margin numbers are the point: they forbid the obvious fix. lilt-a's super-beat ratio is 1.029 against the shipped 1.05 margin (the app's own fixtures measured 1.02 to 1.12; the bench's lighter accents land at the bottom of that band), and verb-y's reverb smears its ratio to 0.991, down among the straight items' own ratios (their published band runs 0.91 to 0.97), where no margin reaches it without letting straight songs through. The distributions overlap, so lowering the margin trades missed 6/8 for false 6/8. The save is complete, though: on all four compound items the true dotted tempo sits on the shipped alternates ladder at top-rung strength, so the rung-picker surfaces the right count everywhere the automatic call misses.
The Confound Shipped
V2 is the finding that matters: both scorable real recordings flipped to 6/8. mop-heart re-picked to 74.7, which is exactly the declared 224 divided by three, the bar pulse of a fast waltz counted in one: a musically defensible hearing with no rung to land on, since 224 matched the old detection at rung 2 and matches nothing on the new ladder, which lacks a times-3 rung. And peaceful-dreams' sub-beat test fired at the detector's own two-thirds-wrong pulse, which is Study 11 Amendment 1's confound, now in shipped code: at the wrong pulse, a third of the lag IS the true eighth note, the test is structurally unable to tell compound-here from straight-at-1.5x-of-here, and the shipped ladder then matched the declared 145 on the detector's own confounded flag. The warning that Study 11's gates required to stay loud goes silent, for the wrong reason.
Verdict, and the Scorecard
The verdict delivered to Stem Shovel: keep the triplet share, the ladder util, and the re-pick as shipped; evaluate the compound flag at the DECLARED tempo's lag, the Amendment 1 rule, which cleanly separates the genuine compound item (0.72 at its declared lag) from the confounded one (0.03); add the times-3 rung or a 3/4-in-one expression; and do not lower the margin, since one item proves the bands overlap. Not a stop-ship: the false fires on declared-straight real songs are contained by the flag fix at the warning surface, and the alternates carry the truth. The scorecard: prediction 1 half-right (the re-pick works more often than the smoke suggested), 2 wrong in the instructive direction (both real recordings flipped, the exact risk the addendum named where it said to stress), 3 and 4 right in full. Stem Shovel answered with a patch that moved both gates to the declared meter, rejected the bench's proposed flag fix with a counterexample, and requested the validations that became Study 11c, where the patch validated and the bench's own rule was retired on the record.