dejvino/music-video-gen
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Withdraw the separation figure: it is a ratio of two small differences

Browse Source 
Three runs of the post-migration song variety measurement give separation 0.314,
0.555 and 0.065 — mean 0.311, half-range 0.245. The 0.38 quoted from a single
run was meaningless.

The cause is structural rather than a sample-size accident. Separation is
(observed - floor) / (ceiling - floor), and here the numerator is 0.0089 while
floor and observed each carry an error bar of 0.015 to 0.019. Dividing one small
difference by another amplifies the noise in both. The metric cannot resolve the
thing it was built to report.

What survives: the ceiling now sits above the floor on every run, with zero
variance, where before the epic it landed underneath and the ratio was not
computable at all. That is a change in kind and it is solid. The magnitude is
not.

The decomposition is the measurement to steer by — noise floor of exactly zero,
effects an order of magnitude above any drift — because it compares renders
directly instead of dividing differences of aggregates. Documented as such.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Dejvino 2026-08-18 06:48:38 +02:00
parent c3b67f45bd
commit df1ce9a957
1 changed files with 24 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
flow-state/MIGRATION.md
View File

@ -206,21 +206,35 @@ looks because the scene it probes is picked from the migrated set and changes
between runs — Constellation, then Metaballs, then Floating Geometry — so read
it as "comparable to the container" rather than as a precise ratio.
End to end, song variety across twelve songs before and after the epic:
End to end, song variety across twelve songs — and a warning about how to read
it. A single run after the migration gave separation 0.38. Three runs give:
```
floor observed separation
before 0.1449 0.1508 not computable (the reference sat below the floor)
after 0.1185 0.1376 0.38
separation 0.314, 0.555, 0.065 mean 0.311, half-range +/-0.245
floor 0.1073 +/-0.0145
observed 0.1162 +/-0.0185
ceiling 0.1434 +/-0.0000
```
The floor fell and the gap widened, which is both halves of what the inversion
was supposed to buy. Every structural block moved with it — scale 40→59%,
orient 65→89%, layout 60→111%, texture 89→81%, motion 53→65% of reference.
**Do not quote the separation figure.** It is a ratio of two small differences —
`(observed - floor) / (ceiling - floor)` — and the numerator here is 0.0089
while each of its terms carries an error bar twice that size. The ratio is not
measuring the generator at this sample size, it is amplifying the noise in both.
The 0.38 that appeared in a single run was meaningless, as was reporting it.
Coupling did not move: -0.09, still indistinguishable from zero. Whether a song
LOOKS different in proportion to how it SOUNDS different remains unsolved, and
nothing in the migration addressed it.
What can be said: the ceiling now sits reliably ABOVE the floor, with zero
variance across runs, where before the epic it landed underneath and the ratio
was not computable at all. That is a change in kind and it is solid. The
magnitude is not.
The trustworthy measurement in this harness is the DECOMPOSITION, not the
top-line ratio. `decompose` has a noise floor of exactly 0.0000 and effects an
order of magnitude larger than any drift, because it compares renders directly
rather than dividing differences of aggregates. Steer by that.
Coupling did not move: -0.02 +/-0.03. Whether a song LOOKS different in
proportion to how it SOUNDS different remains unsolved, and nothing in the
migration addressed it.
Read that with its caveat: the stage set changed underneath the measurement, so
part of the jump is that Metaballs expresses a cast more strongly than