music-video-gen/flow-state/EPIC-4.md

Epic 4 — the song tells a story

Epic 3 asked what is on screen. This one asks a question none of the existing layers can answer:

Why is this the last drop rather than the first one?

Nothing in the generator knows. And a viewer who cannot tell the difference is watching a loop with good production design.

---

1. What is actually missing

The generator has four timescales, and ArcDriver's header lists them honestly: per frame, per shot, per section, whole song. Three of the four are cyclic or local. The fourth — the "whole song" one — turns out to be much thinner than it reads:

• _slowAxisFor is the only thing in the system that travels one way across the track. It moves one or two params per scene, journey = smoothstep(t/duration), in a direction picked by rng.bool(). It is real progression, and it is blind: it does not know where the drop is, it does not know a section from a boundary, and half the time it travels the wrong way for what the song is doing.
• paletteArc.underDrift — a bounded hue crawl, plus per-kind offsets.
• buildSlope — one bar of lookahead. Local by design.

Everything else that decides what a video looks like is keyed on section kind, and kinds recur:

assignRostersByKind()   roster per KIND — every drop cuts between the same visuals
KIND_ENERGY / biasFor() same energy, density, motion for every drop
rhythmFor(energy)       same cutting pattern for every section of that energy
derivePaletteArc()      kindHue.drop — the same hue offset at every drop
frameShot(style, ...)   framing from previous shot + energy; no notion of when

So the fourth drop is cast from the same roster as the first, biased to the same energy, cut at the same rate, tinted the same hue and framed by the same rule. The content of the video is a function of kind, and kind has no arrow of time in it. That is precisely a song structure without a story: recurrence without consequence.

This is not a bug in any of those modules. Kind-keying is what gave the video its identity (§assignRostersByKind), and it should stay. What is missing is the second coordinate.

Today a section is identified by what kind of thing it is. It should be identified by what kind of thing it is, and where in the story it sits.

---

2. The proposal, in one paragraph

A new pure module, look/Story.js, runs once per track between segmentation and the look generator. It reads the section list and the summary, picks a plot the way directors.js picks a director, locates the song's key moments, and emits a small narrative state per section and per frame. Everything that currently keys on kind keys on (kind, story) instead. No scene changes, and — in the first slice — no new uniforms: the story acts by rewriting inputs the whole library already consumes, so it applies to all 61 scenes on day one.

---

3. What a story is, here

Three layers, smallest to largest.

Position — where a section sits

Derived, not invented. From the sections segment.js already produces:

{
  index, kind,
  ordinal, ordinalOf,   // "the 3rd of 4 drops" — the single most useful missing fact
  act,                  // setup | development | turn | climax | resolution
  isMoment,             // one of the key moments below
}

ordinal alone unlocks most of the devices in §5, and costs a Map and a loop.

Moments — the frames the story turns on

Measured from the track, never imposed. A short list, and each one has a definition that falls out of data already on the FeatureTrack:

moment	definition
arrival	first section whose energy clears 0.6 × max — the video's first "here it is"
turn	largest energy fall between adjacent sections after the arrival
climax	max-energy section; ties broken toward the later one
resolution	first section after the climax whose energy stays below it

A two-section track collapses these onto each other, and that is fine — a track with no structure gets almost no story, which is the correct behaviour and not a degenerate case to defend against.

Plot — what the track does with them

One coherent narrative shape per track, chosen weighted-random with the audio tilting the odds, exactly the way pickDirector works, and for the same reason: a fixed mapping from measured features to narrative is how a library ends up with one story per genre.

plot	shape
emergence	almost nothing, then something. Reveal rises monotonically and stays.
escalation	each recurrence of a kind is further than the last. A ratchet, not a curve.
collapse	order → entropy. The climax is a breakdown of the thing, not a peak of it.
return	ABA. The outro rhymes with the intro, transformed by what happened between.
unveiling	the protagonist is withheld until the climax and then it is all there is.

Each plot is a set of curves over five story variables, evaluated per frame:

tension     0..1   how hard everything is pushed
reveal      0..1   how much of the song's identity has been shown
closeness   0..1   wide and distant → close and involved
population  0..1   sparse → crowded → alone
order       0..1   regular → broken (or the reverse; the plot decides the sign)

These are staged curves, not ramps: they hold flat inside a section and move at boundaries, with a step at the climax. That is what makes them read as a story rather than as a slow zoom — a story advances in scenes.

---

4. Where it plugs in

Every one of these is an existing call site gaining an argument.

site	change
biasFor()	tension modulates energy/density within the kind's envelope, bounded to ±0.2, so a breakdown at high tension is still a breakdown
sampleValues(…, temperament)	extremity scaled by tension — the ratchet for escalation
assignRostersByKind()	roster stays per kind; which member plays becomes a story decision (§5.1)
buildStack()	overlay chance follows population, not just energy
rhythmFor()	later acts pick from faster patterns; the resolution gets a held shot
frameShot()	closeness biases the size draw
derivePaletteArc()	new 'narrative' mode keyed on act rather than on kind
ArcDriver._paramsAt()	journey comes from story.journeyAt(frame); the slow axis gets its sign from the plot, not from rng.bool()
ArcDriver.update()	passes a story-shifted personality clone, memoised on rounded reveal exactly as _paletteAt memoises on rounded shift

That last one is how the story reaches Epic 3's content registers without a new uniform: setPersonality is already called every frame, and the identity uniforms are derived from the personality object. Scaling notchDepth, hollow, inkOutline, posterize and latScaleSpread toward their full values as reveal rises makes the song's cast literally arrive over the course of the video.

---

5. The devices, ranked by legibility per unit of work

5.1 The anchor is earned. Today roster[0] opens every section of its kind. Instead, reserve it: earlier drops play companions, and the anchor arrives at the climax. Same roster, same identity, and now the biggest visual in the video lands on the biggest moment. Roughly twenty lines.

5.2 Recapitulation. Under the return plot, the outro re-casts the intro's scene, with the story's parameters rather than the intro's. The oldest device in music video and the cheapest one here — the scene is already built and cached.

5.3 The ratchet. Under escalation, ordinal/ordinalOf scales temperament extremity and slow-axis travel per recurrence. The fourth drop is measurably further out than the first, on every param the scenes declare.

5.4 Reveal schedule. §4's identity scaling. Under unveiling the protagonist's u_cast* form is a near-circle until the climax, then snaps to full at a downbeat.

5.5 Punctuation. A budget of two or three single-use events for the entire video, spent at the moments in §3 — the only cut to black, the only symmetry-fold flip, the only feedback reset. Single-use is the whole point: a device used twice is a style, used once it is a moment. These are the only additions that need a flash-gate review.

---

6. How to know if it worked

The existing instrument can measure this almost unmodified. Build the section×section descriptor distance matrix that checks/variety already knows how to produce, and ask three questions of it:

• Direction. Does distance correlate with |i−j| beyond what kind explains? Today this is ~0 by construction: the matrix is kind-blocked, all drops mutually near, and time is invisible. A story makes it a gradient. This is the headline number.
• Recurrence. For a repeated kind, is d(first, last) > d(first, middle)? That is the ratchet, and it is the one a viewer names as "it kept going somewhere".
• Coherence bound. Adjacent-section distance must stay under the existing ceiling. A story that maximises Direction by shuffling scenes is the failure mode, and this is the gate that catches it.

Plus: each punctuation fires exactly once, on a downbeat, under the flash limit; and every curve is a pure function of frame, so seek-exactness and the determinism grep hold.

A prediction, stated in advance: the seed-variety floor — how far a video travels from itself — will rise, because that is what progression is. Epic 3 spent its effort pushing that number down. Both are correct, and the instrument is what is wrong: it measures distance and calls it drift, with no way to tell wandering from travelling. The fix is one extra statistic, not a retreat from the feature — split the self-distance into an ordered component (monotone with time; a story) and an unordered residual (a shuffle). Ship that statistic before the feature, or the first honest measurement of Epic 4 will read as a regression and be argued about instead of read.

---

7. Risks

The story overrides the song. A plot that declares a climax where the track is quiet is worse than no plot. Mitigation is structural: moments are found in the audio (§3), never placed by the seed, and tension is bounded inside the kind envelope so the quiet-kind coupling in directors.js — the one that keeps an intro off a strobing scene — still holds absolutely.

Every video tells the same story. The exact failure directors.js was written to fix. Same mitigation: five plots, weighted, with seeded curve parameters inside each.

It becomes a slow zoom. If the curves are smooth ramps, this is an effect, not a narrative. Staged curves with plateaus and a step at the climax are load-bearing, not a refinement.

Short tracks. Under three sections, most of this has nothing to work with. Degrade to the current behaviour explicitly rather than letting the curves do something arbitrary.

---

8. The smallest experiment worth running first

Do not build five plots on a prediction. Three changes, no new uniforms, no scene edits:

1. Story.js with position and moments only — no plot templates, one hardcoded escalation curve.
2. Wire it to exactly two sites: the slow-axis sign/magnitude in _paramsAt, and the anchor reservation in §5.1.
3. Add the Direction statistic to the variety report and run the song bank.

The prediction is specific: Direction moves off zero and Coherence holds, while the between-song distance is unchanged — the story should differentiate a video from itself in time, and have no opinion about other songs. If Direction does not move, the story is not reaching the image and the rest of the epic is worth nothing until it does.

---

11. What was built, and what it measured

Built, in look/Story.js plus one argument at each call site listed in §4:

• position (ordinal/ordinalOf/act), the four moments, five plots with seeded curves;
• the anchor is earned (§5.1), the ratchet on temperament (§5.3), recapitulation (§5.2), the reveal schedule on the identity uniforms (§5.4), story-driven cut rate, framing closeness, and a narrative palette-arc mode;
• the slow axis now takes its direction from the track rather than a per-scene coin flip, and its journey from the staged story curve;
• Phase 13 (checks/phase13.js, 8 checks, no GPU) and the direction statistic in the variety report (checks/variety/signature.js).

Punctuation (§5.5) was not built — it is the only part that needs a flash-gate review, and it is worth doing after the numbers below are understood rather than before.

The first direction measurement

checks.html?variety=1&library=0&seeds=6, one song, against the arcless single-scene reference:

floor      0.1697        direction   -0.14
observed   0.1594        arcless ref  0.01
ceiling    0.1996

The prediction in §8 was that direction moves off zero upward. It did not. Three readings, in the order they should be checked:

1. The recapitulation is fighting the statistic, by construction. Roughly two videos in five recap, and return — the plot most likely to — is an arch that comes back. Its first and last probes are deliberately similar, which is exactly what a negative rank correlation between time separation and distance means. The statistic as written cannot tell ABA from no story at all; it may need to be measured against the journey curve rather than against clock time.
2. Six seeds of one song is a small sample, and the probe count (5) makes each video's correlation rest on ten pairs.
3. The arc may not be reaching the image, which is the reading that matters and the one §8 was written to expose. If 1 and 2 are controlled for and direction stays at zero, the story is moving parameters that do not change the picture — the same failure the slow axis had before scenes declared slowAxis, and the fix would be the same: name the levers rather than guessing at them.

The no-story control arm (the same report with the story layer bypassed) did not complete — the run hung in the browser after the first arm, so the floor and separation figures above are not yet attributable to this work either way. That comparison is the next thing to run, and it should be run before any conclusion is drawn from the numbers.

---

12. The second direction measurement — reading 3 was right

Same invocation, checks.html?variety=1&library=0&seeds=6, after the two fixes in §12.1:

                §11          now
direction      -0.14       +0.24        arcless reference -0.02
floor          0.1697      0.2130

§8 predicted direction moves off zero upward and that the floor rises, because that is what progression is. Both happened, and the epic's headline hypothesis is confirmed.

Of the three readings offered in §11, reading 3 was the correct one: the arc was not reaching the image. The story layer was never the problem — it had nothing to speak through. Two of the channels a narrative would have to travel down were inert, and neither failure was visible from the source:

The overlay path was structurally dead. buildStack gated layering on surfaceOf(m) === 'composable', and no scene in the library declared surface — so the only scene that could ever sit on top was the one declaring role: 'accent', which the overlay roster excluded by construction. Empty intersection, every time: 0/144 stacks carried an overlay. population had been wired to the overlay chance per §4 and moved a number nothing read. Fixed by labelling the library from the phase 12 coverage gate — the 37 scenes painting under 30% of the frame are composable — and collapsing the reserved accent slot into one path. 1 distinct overlay scene became 27, at 34% of stacks.

The camera was inert. framing.shift moved the frame by a median of 0.029 of a half-frame at a fresh uniform angle every shot, so successive offsets cancelled and the median jump at a cut was 0.014. closeness reached shot SIZE per §4; the recentre got a per-track constant and a die roll. Fixed by look/Camera.js — see §12.1. Median offset is now 0.190 and the median reframe 0.135.

This is the same failure mode the slow axis had before scenes declared slowAxis, exactly as reading 3 predicted, and it had the same shape: a lever that existed, was wired, was covered by gates, and moved nothing. Both passed every check that existed, because every bound on them was a ceiling. A device doing nothing clears a ceiling comfortably.

The lesson worth carrying out of this epic: a gate on a device needs a floor, not only a limit. Three of the four new camera checks are floors for this reason.

Readings 1 and 2 are still uncontrolled — the recap arch and the small sample both still apply, and the honest statement is that direction was measured under them both times, so the change is attributable even though neither absolute figure is clean.

12.1 Built since §11

• look/Camera.js — the director's camera department, and the answer to "who translates story into imagery". Story says tension, order and act; this turns that into where the frame looks and how it travels there. Each director in directors.js names a camera. Jump distance follows tension and act, speed follows energy, curve is one of four, and a cut chooses between reframing and matching so two scenes can still read as one place. 94% of shots now move during the shot — the recentre is no longer per-shot constant, which is a deliberate departure from the rule in framing.js (that rule was right about size and wrong about where the camera is looking).
• surface declared across the library, role: 'accent' retired in favour of surface: 'composable' plus background: false, and one layering path instead of two.
• Four camera gates in phase 11, each with a floor.

---

13. What is left

Ranked. Three of the four are measurements, and they all report through an instrument that is currently broken — hence the ordering.

13.1 The seed-variety instrument does not report. Not in the original plan, and now the first thing to fix:

separation — not computable: the reference landed below the floor
floor 0.2130 · ceiling 0.2179 · all six seeds read "thin"

The ceiling — videos built from casts sharing no scenes — is 0.2179 against a floor of 0.2130. Real seeds already differ by about as much as maximally-unrelated ones, so separation divides by a near-zero interval and comes back NaN. This is the separation NaN in the standing phase 12 failure, and it has been there across every run including the pre-work baseline.

§11 anticipated this precisely — "the instrument is what is wrong" — and proposed the direction split as the fix. But direction was added alongside the broken separation rather than replacing it, so the gate now leads with an unreadable number while a working one sits underneath. The decision to make: repair the ceiling, or retire separation and read (floor, direction) as the pair.

13.2 The coherence bound (§6). Never built. Optional while direction sat at zero; load-bearing now that it is +0.24, because nothing currently stops future work from buying direction by making adjacent sections incoherent — the named failure mode in §6.

13.3 Recurrence (§6). Never built. phase13 has a spec-level cousin — was the kind recast, did tension move more than 0.08 — but not the descriptor-distance statistic d(first, last) > d(first, middle). That is the ratchet measured in the image rather than in the spec, and §5.3 calls it the one a viewer names as "it kept going somewhere".

13.4 Punctuation (§5.5). Still the only part of the plan wholly unbuilt. Nothing in src/ mentions it. Needs the flash-gate review, which is why it was deferred; §11 said to do it once the numbers were understood, and they now are.

13.5 The no-story control arm. Still never completed. Less urgent than when §11 was written — direction clears the arcless reference by 0.26, which is hard to explain without a story — but it remains the only thing that would attribute the floor rise (0.1697 → 0.2130) to this work rather than to everything else that changed alongside it.

---

14. Composition — the frame has a floor and a ceiling

Not part of the original plan. It comes out of the same place §12 did: the layer stack was being asked to carry the video, and half of what it was stacking was nothing.

The floor. Two thirds of the library is composable — sparse by design, elements ON something. canBackground let almost all of it anchor a section anyway, so a section's bottom layer was regularly a scene painting 2% of the frame with black behind it. Measured across six videos, sampling the middle of every section: 9 of 40 frames were under 20% painted, the darkest at 0.3%. That is a minute and a half of a few bright things on black, and no gate could see it, because every gate on the stack was a limit rather than a floor.

Every section now stands on a GROUND: a canvas painting at least half the frame, cast per section kind so a section's shot cuts change the shot and not the world. When the shot is itself a full canvas it is the ground — two canvases stacked is two pictures fighting. The same measurement now reads mean 92% painted, darkest 63%, nothing under 30%, and phase 12 gates it in the image, not in the spec.

The ceiling. Layering had a rate but no budget, so how full a frame got was a die roll that knew nothing about the song. A stack now has a coverage budget — director appetite (crowd, 1.05 brutalist to 1.7 corrupt), times the section's energy and density, times where the story is — capped at 200%, two frames' worth of material. The ground and the shot are paid for first; the budget governs what may be stacked on top, and the odds of an overlay fall off as the headroom does rather than only at the wall.

What it cost, and what it bought. Same instrument, same six seeds:

                baseline    with grounds
separation      0.10        0.44
direction       0.06        0.21   (arcless reference 0.02 → 0.08)
library         80%         97%    of what the library can express
scale           42%         56%
motion          68%         96%
texture         74%         90%

The ceiling fell (0.2334 → 0.2031) while observed rose, so part of that separation gain is the reference arm coming down: videos that share a ground vocabulary are less unalike even when they share no shots. The honest reading is that the floor rose 0.009, direction tripled, and the ceiling moved toward the floor — the numbers are up, and not all of the rise is signal.

A scene that reads the previous frame cannot be a ground. prev() returns the whole composited frame, including whatever is layered on top of this scene, so a datamosh under a shot is not grounding it, it is eating it. Measured the moment one became a bed: the render stopped reproducing from a seek, and two WebGL contexts diverged by 91/255 against a ceiling of 4 — phase 4 caught both. readsHistory is derived from the shader source rather than declared, and costs the ground pool six of twenty scenes, five of them glitch.

scenes/coverage.js is the measured coverage of every scene, pasted back out of the phase 12 gate, which re-measures and fails on drift over a tenth of a frame. A new scene has no entry and therefore cannot ground anything until someone renders it.

14.1 The measurements moved into the repo

Everything in §14 depends on a number — how much of the frame a scene paints — and the first version of it was a table pasted into a source file by hand. That is the same mistake surface already was: a measured fact written down by a person, correct on the day and wrong after the next shader edit.

src/scenes/metadata.json is now generated, tracked in git, and holds every measured fact about every visualizer: coverage as a shot, coverage as a bed, the variety score, the per-block scores, and the mean structural profile. Nothing in it is typed by a human. surface is derived from it (canvas at 50%, composable below) and declaring surface: in a scene file is now a lint error.

It is refreshed from gallery.html → refresh metadata, which re-renders the library and writes the file through a dev-only endpoint in vite.config.js. The file carries a fingerprint of everything that can move a number in it — the scenes, the shader contract, the identities and palettes they are handed, the descriptor definitions — and phase 12 fails when that stops matching. A stale metadata file does not produce a stale report, it produces wrong videos, so it has to be detectable.

Measuring at the wrong bias measures nothing. Coverage is mostly a function of a scene's parameters, and the first measurement took one number per scene at the busiest section of six songs. Metaballs measures 69% there and painted 6% as an intro's ground. Three things came out of chasing that, in order of how much each was worth:

• A bed is drawn with a moderate hand. extremity pushes parameters toward the ends of their ranges, and half of what decides coverage is a parameter with no bias key at all — Metaballs' threshold — so extremity is the only thing moving it and one end is an empty frame. Grounds sample at extremity * 0.2. The extremes belong to the shot.
• Energy and density are floors for a ground, not reductions. Calming the bed by lowering energy is how to empty it: scenes bias their fill against energy as often as against density.
• Coverage is measured twice, once as a shot and once as a bed, at the same bias and temperament the generator will use. GROUND_BIAS and groundTemperamentFrom live in scenes/surface.js and are imported by both the generator and the measuring pass, so the two cannot drift apart.

Rendered result across six videos, sampling the middle of every section: mean 87% painted, darkest 30%, against 51%/0.3% before any of this.

Eleven scenes can ground a section — five organic, five geometric, one structural. There is no minimal, flow or glitch canvas in the library that paints half the frame without reading prev(), so every quiet section in every video stands on one of five organic beds. That is the largest remaining hole and it is library work, not generator work: the fix is writing dense, self-contained canvases in the three thin families.

14.2 The red line is gone

MIN_VARIETY — 0.1, drawn across the gallery, "below this a scene is the same picture wherever it appears" — was right when a section was one scene and is wrong now. A section is a ground, a shot on it and sometimes a pass over that, so what a viewer sees is a combination, and a scene that is reliably itself is a good ingredient in one. Nineteen scenes were failing a bar for being consistent.

What replaced it is a level up, and it is the reason the profiles are in the metadata: the generator weights every layering choice by structural distance — how unalike two scenes measured, on the same descriptor the variety harness compares videos with. Family labels and the render disagree often enough to matter; two geometric scenes can be 0.31 apart and a flow and an organic scene 0.04, and stacking the second pair is one picture at double density. Measured over 436 stacked pairs: mean distance 0.268, zero near-twins.

14.3 The other end of the frame

The floor had a ceiling missing. Measured across twelve videos, sampling the middle of every section: a median 24% of every frame was clipped to pure white, and whole sections rendered at 100%. The videos were washed out, and no gate could see it because every guard on the frame asked whether there was enough in it.

Three causes, all introduced or exposed by §14, in the order they were found:

The shot was screened over its ground. Screen is a lightening operator — correct for a few bright elements over a bed, and over a filled canvas it drives everything toward white. Replaced with a lumakey blend (passes.js): the shot's own brightness is its alpha, so it replaces the ground where it paints and leaves it where it does not, keeping its own colour instead of adding it to the bed's. Median clipping 24% → 11%.

The feedback loop was an accumulator. cur + hist * decay * amount settles a still image at 1/(1 - decay*amount) times its drawn brightness — 2.3x at the settings the generator hands out. Survivable when a frame was a few bright things on black; fatal the moment every section stood on a filled ground. Turning feedback off took a blown frame from 100% to 34% mean luminance, which is the whole diagnosis in one number. It now divides by the same gain, at 0.6 rather than 1.0: full normalisation took the lift out along with the blowout — the median frame fell from 95% painted to 71% and fourteen sections dropped through the black-frame floor.

Nothing rolled off. The grade clipped. A highlight shoulder now compresses everything above 0.75 toward but never to 1. Clipped white is not brightness, it is missing information: every difference inside it has been deleted.

Rendered, after all three — 74 sections, twelve videos:

                 before        after
painted          51% mean      88% mean, darkest 43%   (was darkest 0.3%)
clipped white    24% median    0% median, worst 21%    (was worst 100%)
luminance        77% median    42% median, worst 80%

The gate is now two-ended: composition · a rendered section is neither black nor blown out.

A ground has to survive the song's identity. Chasing the dark end turned up the same class of error as §14.1: a ground measured at 61% painted 2% in one particular video, because that song's ink treatment is hollow — outlines, no fill. Two changes came out of it. The world is now drawn solid and the song's hand is kept for the subject (groundPersonalityFrom), and eligibility takes the worst identity into account as well as the mean: a ground fills the frame on average and never vanishes. Holding the worst case to the full 50% would leave four castable beds in the library, which is a worse video than an occasionally dim intro.

Seven scenes can ground a section — five geometric, two organic. Every quiet section of every video stands on one of two beds. Restated from §14.1 because it got worse, not better: this is the library's largest hole.

14.4 Blazing on purpose

Screening the shot over its ground and letting the two brightnesses sum is a real effect — a drop that goes to paper for eight bars reads as the song peaking. The mistake in §14.3 was not the effect, it was that the effect was the default: every section did it, so a median quarter of every frame in every video was clipped, and nothing about that said "peak".

It is a decision now. Each director declares a blaze appetite — brutalist 0.05 (mass does not glow), corrupt 0.5 (overload is the subject) — and a section still has to earn one: loud, and late in the story. Quiet kinds never blaze, because a breakdown that goes white is not a choice, it is a bug with a rationale. The default composite is the lumakey, which keeps the shot's own colour.

The ceiling gate matches the intent rather than banning brightness. Per section, only a hard cap — past about half the frame at full white there is no picture left to read. Across the population, no more than a fifth of sections may be hot at all, which is the number that actually distinguishes a director choosing to peak from a pipeline with no headroom. Measured now: 1 of 74 sections blazing, worst 30% clipped.