ef1d24e692
10 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Dejvino
|
89e05459c0 |
Every shot stands on something, and the frame has two ends
A section used to be one scene, and two thirds of the library is composable —
sparse by design, elements ON something. Cast as backgrounds anyway, they left
9 of 40 sampled frames under 20% painted, the darkest at 0.3%: a minute and a
half of a few bright things on black, invisible to every gate because every
gate on the stack was a limit rather than a floor.
Every section now stands on a GROUND: a canvas that fills the frame, cast per
section kind so a shot cut changes the shot and not the world. When the shot
fills the frame itself it IS the ground — two canvases stacked is two pictures
fighting. Above that, a coverage BUDGET: director appetite times the section's
energy times where the story is, capped at two frames' worth of material.
The measured facts move into the repo. scenes/metadata.json is generated from
the gallery — coverage as a shot, coverage as a bed, variety, the structural
profile — tracked in git, stamped with a fingerprint of the scenes and the
metric definitions, and refreshed from gallery.html. `surface` is derived from
it rather than declared; nine scenes claimed `canvas` while painting under a
third of the frame, and declaring it is now a lint error. The generator weights
every layering choice by measured structural distance, because family labels
and the render disagree: two `geometric` scenes can be 0.31 apart and a `flow`
and an `organic` scene 0.04.
The gallery's 0.1 red line is gone. It was right when a section was one scene
and wrong now — nineteen scenes were failing a bar for being consistent, which
is a virtue in an ingredient.
Chasing the numbers turned up four real faults:
* A scene that reads prev() cannot be a ground. It returns the whole
composited frame including the layers above it, so a datamosh under a shot
is eating it: the render stopped reproducing from a seek and two WebGL
contexts diverged by 91/255 against a tolerance of 4.
* Screen was the wrong operator for a shot over a bed. It lightens, so a
median quarter of every frame clipped to paper and whole sections rendered
100% white. Replaced by a lumakey — the shot's brightness is its alpha.
* Feedback was an accumulator: a still image settled at 2.3x its own
brightness. Fine over black, fatal over a filled ground. Normalised at 0.6,
plus a highlight shoulder so the top rolls off instead of clipping.
* useTrack never prewarmed, so a fresh Show's first frame differed from every
later render of it — the export-breaking hazard Compositor.prime documents.
Blazing is a decision now, not a side effect: directors declare an appetite for
it, a section must be loud and late in the story to earn one, and quiet kinds
never do. The ceiling gate matches that — a hard cap per section, and no more
than a fifth of them hot at all.
Rendered across twelve videos, middle of every section:
painted 51% mean, darkest 0.3% -> 87% mean, darkest 43%
clipped white 24% median, worst 100% -> 1% mean, worst 30%
separation 0.10 -> 0.44
Seven scenes can ground a section — five geometric, two organic — so every
quiet section of every video stands on one of two beds. That is the library's
largest hole and it is scene work: there is no minimal or flow canvas that
fills half the frame without reading prev().
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
Dejvino
|
1ca9fa2f1f |
Let scenes layer on each other, and give the camera somewhere to look
Two devices that were built, wired up, and doing nothing. OVERLAYS. buildStack gated layering on `surfaceOf(m) === 'composable'`, and no scene in the library declared `surface` — so the only thing that could ever sit on top was the one scene declaring `role: 'accent'`, and the overlay roster excluded accents by construction. Empty intersection, every time: 0/144 stacks carried an overlay. Every layered frame in every song was the same particle field. Labelled the library from the phase 12 coverage gate rather than by eye: the 37 scenes painting under 30% of the frame are composable. Particle Field loses its privileged slot and becomes one of them, keeping only `background: false`, which is the honest part — points in empty space cannot carry a section alone. The reserved accent slot is gone; one roster, two passes at it. 1 distinct overlay scene becomes 27, and the rate lands at 34% of stacks after trimming a base chance that had been tuned while the branch was dead. THE CAMERA. framing.shift moved the frame by a median of 0.029 of a half-frame at a fresh random angle every shot, so successive offsets cancelled and the median jump at a cut was 0.014. Present in every frame, visible in none. look/Camera.js is the director's camera department: the story says tension, order and which act a section is in, and this turns that into where the frame looks and how it travels there. Each director names a camera. Jump distance follows tension and act, speed follows energy, and the curve is one of four. Cuts choose between reframing and matching, so two scenes can still read as one place. Median offset 0.190, median reframe 0.135, and 94% of shots now move during the shot rather than only at the cut. Four new gates, each with a FLOOR — the device this replaces passed every existing check while doing nothing, because the only bound was a ceiling. Also lands the in-progress Epic 4 story layer it builds on: Story.js, phase 13, and the direction statistic. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
e5bb7d77e0 |
A reproducible migration, and the first eighteen scenes through it
The library is sixty-one scenes, which is too many to convert from memory or from taste, so the migration is a queue with a gate rather than a judgement call per file. Three pieces. A classifier reads each shader and assigns a tier from evidence in the source — drawn, figure, field, treatment — so two passes over the library reach the same answer and the work has an order. MIGRATION.md is the recipe per tier, written to be followed mechanically. And a gate makes the result verifiable: `consumes` is now a schema field, the lint enforces it in both directions, and the per-scene battery renders each scene under two deliberately distant identities and requires the picture to change. That gate is the part that matters. Without it `consumes` is a comment, and the whole inversion becomes unverifiable at exactly the point where it stops being checkable by eye. With it, a scene that declares the cast and ignores it fails. Eighteen scenes migrated. Four by hand at the drawn tier — Firefly Drift, Metaballs, Floating Geometry, Prism Bloom — and ten at the field tier by script, which is one declaration and one wrapped return. All eighteen pass. The field tier is honestly marginal and the gate says so: every one of the ten moves by 37 to 39 of 255, against 173 to 255 for the drawn tier, and the uniformity across ten unrelated scenes is the tell. That is one global posterisation applying, not ten scenes expressing anything. Cheap, real, shallow. The decomposition moved from identity being worth 54% of the container to 158%, but the stage set changed underneath the measurement and part of that is Metaballs expressing a cast better than Constellation did. What survives the caveat is the useful finding: a migrated library scene carries the identity better than a stage written from scratch to carry it. The four bespoke stages were the wrong shape of effort. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
00ad1d8c2b |
Measure where visual difference comes from, and correct the epic's framing
Two cheap measurements settled a question three expensive ones had not, and both
of them contradicted the diagnosis offered for it.
The stated diagnosis was that the identity's expressive range had become the
bottleneck. It has not. A census of the identities themselves — no GPU, seconds
to run — puts mean distance between twelve songs at 0.43 with no near-identical
pairs and full coverage of every decision space: six of six fills, five of five
lattices, six of six protagonist forms, three of three element scales. The songs
are handed genuinely different designs.
The decomposition then asked whether those designs reach the picture, by holding
the container fixed and varying only the identity, then the reverse:
identity only 0.0299
container only 0.0557
both 0.1101
neither 0.0000
Identity is worth 54% of what the container is worth, against an instrument
noise floor of exactly zero, and the two compose to more than their sum. The
inversion works at the frame level. What it does not do is replace the container.
That corrects EPIC-3 §5, which proposed a song picking two to five stages on the
theory that shared content would substitute for container variety. Container
variety is the larger of the two effects and identity adds to it. Four stages
with a rich identity throws away the 0.056 the library was already providing —
which is exactly the shape of every measurement in this epic: stages have the
lowest floor of any arm and no advantage in spread.
The direction is therefore not a small set of stages. It is the whole library
consuming the cast: keep the sixty-one containers and make them draw the song's
content rather than their own. The migration was filed as a cost; it is the
payoff.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
Dejvino
|
577ec107f6 |
Sweep pool size directly, and withdraw the claim that it mattered
The Epic 3 arms appeared to show that a small casting roster was the largest available win. Swept directly across 4, 8, 16 and 32 over twelve songs with three pool draws each, the differences are 0.003 to 0.007 against a run-to-run noise of +/-0.003 to +/-0.005. Pool size does nothing measurable. The arms varied two things at once — smaller pool AND the same pool for every song — and only the stages-versus-legacy comparison inside them held everything else constant. That one still stands at +14%. POOL_SIZE stays at 8, now on grounds the metric cannot see: about nine distinct scenes in a video rather than seventeen. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
ec38def1a9 |
Epic 3, first slice: the song brings its own cast
Four stages, an identity layer, and an A/B that says the idea is right and the
reason it works is not the reason I expected.
A stage has no image of its own. It owns arrangement — a procession, a
constellation, a soloist, a swarm — and what it arranges comes from the track:
`castMain` and `castChorus` for the forms, `inkMask` and `inkValue` for the hand
they are drawn in. The identity generates a protagonist and a chorus with sides,
notches and hollows, plus an ink treatment of weight, edge, fill, outline and
posterisation. All of it travels as uniforms, so it is data rather than code and
a stage consumes it without knowing any other stage exists.
The protagonist IS the signature form rather than a second opinion about it.
They were separate draws in the first version, which let a track built on
hexagons put a round protagonist on screen — the signature said one thing and
the picture said another, and the shape trait stopped meaning anything for
stages. The cast now reads its geometry from the personality live and adds the
notches and hollows that turn a shape into a character.
Measured across seven songs, three arms, same instrument:
arm floor observed spread
stages, four of them 0.0801 0.1133 +0.0332
legacy scenes, four of them 0.0830 0.1122 +0.0292
the unrestricted generator 0.1058 0.1146 +0.0088
The prediction in EPIC-3 §7 was that sharing a cast would drop the floor
sharply. The floor did drop sharply — by a quarter — but nearly all of that came
from casting FEWER scenes per video, not from sharing content: the legacy arm,
four ordinary scenes with no cast at all, gets most of the same benefit. Stages
add about 14% on top of that.
That is worth knowing before building the other four registers. The single
largest available win was hiding in the roster size all along, and it costs one
number to take.
One result runs the wrong way and is recorded rather than explained: coupling is
+0.29 on the legacy arm and -0.18 on the stages. At twenty-one pairs neither is
distinguishable from zero, so it is a flag for a larger sample rather than a
finding.
Also fixed: a backtick inside the shader preamble's template literal, which
closed it and made every check page hang on "starting…" with no error in the
console.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
||
|
Dejvino
|
75e0f0ef01 |
Make the whole library reachable, and measure what that did and did not buy
Three changes, from the list the measurements pointed at. The personality's values now centre on the audio and the seed picks within that centre, rather than every value being drawn from the seed alone. A bright track sits high in its frame, a dynamic one has depth, noise earns corners where tonal material stays round. The signature's own choice is tilted by the music too — it decides which scenes a track can cast, so leaving it to the seed meant the most consequential decision in the generator had no relationship to the song. The signature stopped being a hard filter. As a filter it was the single largest cause of sameness: a scene declaring all four traits was eligible for every track and opened half of all videos, one declaring two was eligible for one track in fourteen, and eleven scenes out of sixty-one carried nearly everything. It is a weight now, worth six times at full honour. That fix alone made things worse, which is worth recording. The filter was doing two jobs — collapsing the library, and giving each track a DIFFERENT pool to cast from — and removing it kept the second loss. Every track drew from the same weighted library and measured song separation fell. So each track now draws its own pool of about a third of the library, weighted by the signature but sampled without replacement, keeping the differentiation and dropping the bias. Motion became a character rather than a rate. Tempo was the only lever and tempo compresses; stillness is a separate question from speed, and it can now halve the animation rate or raise it by a third. What it bought, measured against the same instrument: every scene in the library is now cast, where sixteen were never reached; identical casts across seeds went from four pairs to none; and the raw structural distances all rose — motion by 57%, layout by 22%, scale by 23%. What it did not buy: videos also became more varied INTERNALLY, by more than they became different from each other. Two songs are no more distinguishable relative to how much one video already changes over its own length than before, and the coupling between musical distance and visual distance is still indistinguishable from noise at this sample size. The ceiling reference now sits below real outputs, so the separation ratio is reported as not computable rather than as a large number. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
a946f0e105 | wip: instrument guards | ||
|
Dejvino
|
2bbf5bcf42 |
Run both variety tests against the song bank, and fix the ceiling twice more
The tests now run on real bank entries. The seed gates used to run on a two-section synthetic whose only kinds were intro and outro, so half the scene library was unreachable and the number was measuring that rather than the seed. Probes are labelled by section kind and occurrence rather than by index, which is what makes two different songs comparable at all — they have different section counts, so probe 3 of one is not probe 3 of the other, and matching by position would compare a drop against an outro and score the mismatch as variety. For two seeds of one song the labels are identical and this changes nothing, which is the point. The song test measures one thing the seed test does not: coupling, the rank correlation between how different two tracks sound and how different their videos look. Separation alone can be had by a generator that ignores the audio and hashes the file, and that would be a perfect score for a completely wrong video. Separation without coupling is not variety, it is a different seed per file. The ceiling took two more attempts. Recasting every layer at random averages a dozen scenes together and a dozen random scenes converge on the same generic busy image, so two references came out closer to each other than two real videos and blocks scored over 100% of achievable. Forcing one scene per reference collapsed the other way: a video that never changes scene has almost no internal variation, so the ceiling landed BELOW the floor, which is a within-video quantity. A reference has to match the structure of what it bounds. They now keep the real pipeline — rosters, shots, per-section sampling, so a reference rotates between three or four scenes exactly as a real video does — while drawing from disjoint slices of the library. Same complexity, nothing in common. The fingerprint measurement had the same shape of error: pooling every probe mixed in how much each video varies over its own length, which is large for everything, and washed the answer to a flat 100% while the separation score said almost everything was collapsed. One vector per video now. Both tests fail as committed. Seed separation 0.04, song separation 0.03, coupling -0.03 — two different songs differ from each other by about as much as one video differs from itself, and that difference has no relationship to the music. Colour scores 112%. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |
||
|
Dejvino
|
656e062069 |
A seed variety test, and a metric that had to be gated before it was believed
Every other phase asks whether one video is correct. This asks whether two are different — the failure the suite could not see, since a generator that ignores its seed passes determinism, flash safety and liveness perfectly. Frames reduce to a structural descriptor built to be blind to the cheap axes and sensitive to the expensive ones: standardized luma kills exposure and palette, a Laplacian pyramid gives the radial spectrum, and the gradient-angle histogram is carried through a DFT magnitude so a rotation shifts it without moving it. Colour is measured and never counted; its only job is to expose the case where two seeds differ by a palette swap and nothing else. The score means nothing on its own, so it sits between two references the same instrument produced: a floor of how far one video travels from itself across its own sections, and a ceiling of the same pipeline with every layer recast at random. Two checks gate the instrument before any number from it is trusted — recolour must move structure ~0 while moving colour a lot, and a quarter turn must not move it at all. Three things this got wrong first and now does not. Averaging each video's probes into one descriptor washed out the structure being measured and put the floor above the ceiling; probes are matched instead, which is fair because the track is held fixed. Cosine distance on all-positive histograms scored unrelated scenes at 0.05, too compressed to be read; chi-square replaces it. Single-link clustering in the library sweep chained overlapping pairs into a fourteen-scene group that did not exist; complete-link means every pair inside a group is really a twin. The sweep runs against all 61 visualizations, and it finds what the per-scene distinct gate cannot, because that one compares raw pixels and structural twins are merely differently coloured. Both rendered gates fail as committed. That is the point of them: separation is 0.07 against a 0.35 target, two 4-cliques of scenes are one look each, and a third of the library is never cast. The instrument passes; the generator does not. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> |