The census says where the signal stops, and for one net it is not the head

[syncytium2]

Tony's ruling on the verify round (#674): items 1-3, and the census re-run.

Item 4 is the one that mattered, and it replaced the headline

census.json now stores the two fields the builder's own docstring had promised since
the tool was written — share_varying (the full eight-layer vector, not just its minimum)
and head_input_sd — so the page can finally tell a signal that died inside the head
from one that never reached it.

The re-run reproduced every pre-existing field bit-identically across all 974 fits,
controls and recording included. Nothing already on the page moved; the two fields are pure
addition. That is also a reproducibility check that could have failed and did not.

	collapsed fits whose head input is already constant	median head-input SD
chorus_norm	10 of 153 (7%)	0.081
chorus_gain_norm	32 of 75 (43%)	0.0019 — about twice the 0.001 called silent
working fits, either net	0 of 279 and 0 of 357	0.13 and 0.24

So "the signal stops in the head" was right for chorus_norm — the head is the site in 143
of its 153 — and wrong for nearly half of chorus_gain_norm's failures, where the votes
never varied and the head was handed nothing to pass on. The page says so, and says the
consequence: a repair aimed at the head cannot touch those fits, and the warm-up in §5
was tried on chorus_norm only. Carried into the todo too, since options 2 and 3 both act at
or after the head.

The new negative result is worth as much as the split: no working fit of either net is
handed a constant head input, which the builder now asserts rather than merely reports.

Items 1-3

1. Wrong numbers. The sign-rule sentence now derives both directions from the census
   instead of printing one net's count in a sentence about two: it calls 5 working fits
   dead (1 chorus_norm, 4 chorus_gain_norm) and misses 3 collapsed ones, where the
   variation test misses none — an argument for the page's choice, written that way.
   Table 2's column is retitled and its note corrected against the raw score archives, which
   I read myself: at the budgeted threshold both refits return n_detected = 0 on all 24
   recordings — they call nothing, a different failure from collapse. Three bounds now
   round outward by rule, so "at least 0.82" became "at least 0.81 logits" and cannot regress.
2. Straw man — restated, not withdrawn. §3 now says the sibling already attributed the
   overlap to shared configurations and seeds, that §2 adds arithmetic rather than a
   correction, and that its conclusion stands.
3. Table 3 — the long parenthetical moved into the note behind a †. Measured: 1000 px
   into 1000 px, zero clipped, outcome visible; independently re-confirmed at twelve
   widths from 320 to 2,560 px.

The round I ran

Not a fourth full round — the repairs are narrow and each was verified by measurement. But
the process says role 10 re-runs in full on any blind pass and the render is what changed, so
the craft gate ran blind alone, archived as 10-ship-it-round4.md and labelled so it
cannot be misread as eleven roles.

0 blocking, 2 major, 10 minor, against 1 / 2 / 10 from the same role last round. It also
caught a defect this round introduced: Table 3's note had begun narrating the page's own
editing history. Removed.

Both remaining majors are pre-existing figure craft, raised by the same role last round,
left because they are not what was ruled on: Figure 6 separates five series by colour alone
(two merge under deuteranopia — exactly the comparison it exists to make), and Figure 4's key
is a caption clause with no swatch for "gray, untrained".

Residual list: blocking 6 → 0. Both structural findings are closed.

Determinism

Byte-identical rebuild confirmed on Python 3.11, 3.13 and 3.14 before pushing, re-run after
every rebuild. Tests 3 passed; sapper clear; darkroom copies replaced and verified byte-for-byte
against the repo; claim released.

No GPU was used. census runs each fit on the CPU — CUDA is reached only by replay(),
which this task never calls — so there was no contention with 064's sweep and none was possible.

🤖 Generated with Claude Code