The chorus-collapse page's blind round: three of six repairs do not survive it

[syncytium2]

Closes the heaviest of the four flags #667 landed with: the round-2 repairs had
never been blind-reviewed. This is that round -- 11 of 11 roles, blind, against
the built page on main.

Verdict: the flag cannot be retired. Three of the six repairs hold, three do not.

round-2 repair	verdict
Figure 1 rebuilt as a matched pair	holds
Interior census and replays (400-frame trim)	holds as a measurement
Figure 5 merged	holds
The #596 test replacing the separation measure	does not hold
The selection split by rule (Table 2)	does not hold for the budgeted column
The replicate report's reading "corrected"	does not hold, and is new damage

The diagnosis itself survives, more firmly than before. collapse_table.json
was re-derived from the raw run archives without calling the builder (1,938 rows,
zero differences), every permutation test re-run with an independent generator at
up to 200,000 shuffles, and the nets rebuilt in torch to re-measure the untrained
controls. The page rebuilds byte-identically on Python 3.11, 3.13 and 3.14, so the
two cross-version determinism bugs that bit it before are closed, not dormant.

What failed, and why each is a different kind of failure

• Figure 4 narrowed a defect instead of removing it. Round 2 replaced a test
  that could not fail. The replacement can fail, and still misses 6 chorus_norm and
  13 chorus_gain_norm fits whose votes are constant on a real recording. Separately,
  the untrained comparator it is set against is the lowest of the grid's four encoder
  shapes by 14x, and contradicts this project's own README at the commit the page cites.
• Table 2's false-alarm-budget column is wrong under the page's own definition.
  Those two refits have n_detected = 0 at that rule's threshold -- they call
  nothing. The page quotes the sibling report's footnote verbatim, including the
  words "or called nothing", and then uses only the first half.
• The correction of the sibling report is a straw man, and round 2 introduced it.
  That report already attributes the overlap to shared configurations and seeds.

Underneath them, the finding that is not a wording fix. census.json stores only
the minimum across the eight head layers and never the head's input, so the page cannot
tell a signal that died inside the head from one that never reached it -- and in 10 of
153 chorus_norm and 32 of 75 chorus_gain_norm collapsed fits, it never reached it. The
builder's own docstring promises both missing fields. Closing this needs a census re-run
on a GPU, not an edit.

Why nothing was repaired. Blocking went 3 -> 4 -> 6. The process says a flat or
rising blocking count means escalate rather than patch, because patching converts a
fixable draft into a long tail nobody has reviewed together. So the page on main is
untouched and the record states its fingerprint as unchanged.

Ranked for Tony -- wrong numbers ("1 working fit as dead" is 5; Table 2's column;
three bounds rounded inward and literally false); the straw man; Table 3's outcome
column clipped 211 px at every desktop width with no affordance; and the census re-run.
Two items outside this claim: docs/goals/learned-model-family.md states 112.6/40.4
against this page's 111.5/39.6 (#664 owns that file), and Kosson et al. 2024 is uncited
prior art sitting under the page's novelty claim.

Both gates green: roster 11/11 with reports, grants 11 ok / 0 mismatch. Record and all
11 role reports also delivered to the claimed darkroom folder; the claim is released here.

🤖 Generated with Claude Code