Rewrite the comp-led build around the phase gates; reviewer reads the diff first

new-work.md section 6 becomes the spec / plates / hero / sections / motion /
responsive phase list, each closed by build-phase.mjs advance; the
reproduction and hero-checkpoint prose that asked the model to compare
from memory is gone. visualize.md's inventory, medium gate, and produce
sections collapse into the spec and plate mechanism. The finish reviewer
takes the state file and comp-diff reports as inputs and starts its
fidelity matrix from the measured verdicts. docs/COMP-FIDELITY.md records
the design.

AI-assisted (Claude).

Co-Authored-By: Claude <noreply@anthropic.com>

README.md

@@ -1,6 +1,6 @@
 # Impeccable
 
-Design guidance for AI coding agents. 1 skill, 23 commands, live browser iteration, and 59 deterministic detector rules for AI-generated frontend design.
+Design guidance for AI coding agents. 1 skill, 23 commands, live browser iteration, and 61 deterministic detector rules for AI-generated frontend design.
 
 > **Quick start:** From your project root, run `npx impeccable install`, then run `/impeccable init` inside your AI coding tool. Full docs: [impeccable.style](https://impeccable.style).
 
@@ -13,7 +13,7 @@ Every model trained on the same SaaS templates. Skip the guidance and you get th
 Impeccable adds:
 - **One setup flow.** `/impeccable init` writes `PRODUCT.md` and offers `DESIGN.md`, so later commands know the audience, brand/product lane, voice, anti-references, colors, type, and components.
 - **23 commands.** A shared design vocabulary with your AI: `polish`, `audit`, `critique`, `distill`, `animate`, `bolder`, `quieter`, and more.
-- **59 deterministic detector rules** plus LLM-only critique checks. The CLI and browser extension run the deterministic rules with no LLM and no API key.
+- **61 deterministic detector rules** plus LLM-only critique checks. The CLI and browser extension run the deterministic rules with no LLM and no API key.
 
 ## What's Included
 
@@ -402,7 +402,7 @@ npx impeccable ignores add-file "src/legacy/**"
 npx impeccable ignores add-value overused-font Inter --reason "Brand font"
 ```
 
-The detector catches 59 deterministic issues across AI slop (side-tab borders, purple gradients, bounce easing, dark glows) and general design quality (line length, cramped padding, small touch targets, skipped headings, and more).
+The detector catches 61 deterministic issues across AI slop (side-tab borders, purple gradients, bounce easing, dark glows) and general design quality (line length, cramped padding, small touch targets, skipped headings, and more).
 
 By default, `detect` respects the same `.impeccable/config.json` and `.impeccable/config.local.json` detector config as the design hook: `detector.ignoreRules`, `detector.ignoreFiles`, `detector.ignoreValues`, and `detector.designSystem.enabled`. Hook lifecycle settings such as `hook.enabled` only affect automatic hook execution.
 

README.npm.md

@@ -1,6 +1,6 @@
 # Impeccable CLI
 
-Detect UI anti-patterns and design quality issues from the command line. Scans HTML, CSS, JSX, TSX, Vue, and Svelte files for 59 deterministic rules, including AI-generated UI tells, accessibility violations, and general design quality problems.
+Detect UI anti-patterns and design quality issues from the command line. Scans HTML, CSS, JSX, TSX, Vue, and Svelte files for 61 deterministic rules, including AI-generated UI tells, accessibility violations, and general design quality problems.
 
 ## Quick Start
 
@@ -56,7 +56,7 @@ npx impeccable detect --fast src/
 
 **Quality**: tiny body text, cramped padding, long line lengths, small touch targets
 
-59 deterministic detector rules in total. See the full catalog at [impeccable.style/slop](https://impeccable.style/slop).
+61 deterministic detector rules in total. See the full catalog at [impeccable.style/slop](https://impeccable.style/slop).
 
 ## Exit Codes
 

docs/COMP-FIDELITY.md

@@ -0,0 +1,73 @@
+# Comp fidelity: measuring the build against the comp
+
+Status: shipped on `feat/comp-fidelity` (2026-08). Owner: skill (`skill/scripts/comp-*.mjs`, `build-phase.mjs`) plus two detector rules.
+
+## The problem
+
+v4's direction round and comp round produce beautiful comps. The build that follows is a lossy translation of them: invented chrome the comp never showed, materials flattened to CSS, illustrations approximated as SVG or `clip-path`, produced textures buried under opaque washes, and a first viewport that has the comp's section order and none of its craft. The finish reviewer catches this and orders a rebuild; the rebuild has the same problem; runs hit the turn cap.
+
+Two recent factory runs (07-vintage-moto-forum and 05-experimental-album, gpt-5.6-sol, 2026-08) show the shape:
+
+- 183 KB of skill prose read before the first write; the page written in one 1,500-line write at turn 30; no reproduction phase, no `hero-repro.png`, no side-by-side; turns 31-61 spent on servers, screenshots, and reviewer plumbing until the cap.
+- A torn-paper arch shipped as a 17-vertex `clip-path: polygon`; a vellum slip as a flat gray rectangle; both produced paper textures unused on disk.
+
+The root: every fidelity check in the build phase was the model judging its own reproduction from memory of an image, and the prose kept growing to argue it into behavior it cannot perform. Code-led builds "look better" only because they have no target to fail against.
+
+## The change
+
+Stop asking the model to reproduce pixels it can see but not render, and stop asking it to grade a reproduction it cannot see clearly. Make the translation mechanical wherever pixels are involved; keep the model's job to what it is good at (structure, semantics, controls, motion, responsive logic).
+
+### 1. `comp-diff.mjs`: numbers and crops instead of conviction
+
+Dependency-free (own PNG codec in `lib/png.mjs`). Given the comp and a build capture:
+
+- aligns the build (scale to comp width, take the first-viewport rows; `--align stretch|cover` for other uses),
+- scores structure (SSIM over blurred grayscale with a small translation search), color (quantized histogram intersection + Lab dominant-palette match), detail (high-frequency energy ratio per cell: did the material survive?), and bands (horizontal section boundaries line up?),
+- per region (from the spec, or from the comp's own bands), with region-kind weights: a `plate` region is judged mostly on detail, a `text` region on structure,
+- writes `side-by-side.png`, `heatmap.png`, `regions/<id>.png` paired crops at legible scale, and `report.json` with a verdict per region in the reviewer's vocabulary: `match` / `drift` / `missing` / `contradicted`,
+- `--threshold` exits 3 below the bar.
+
+Calibration on the moto run: comp vs itself 100%; comp shifted 12px 87% (match); comp with the illustration erased 90% overall but the plate region `missing` at 30%; comp recolored to navy 34% (`missing`); the real build 59% (`contradicted`, plate `missing`, index `contradicted`). Sibling comps of the same world score 55-59% against each other, so the metric separates "same design" from "same world, different composition." Runs in ~0.3-0.8 s.
+
+### 2. `comp-spec.mjs`: the comp becomes a measured spec
+
+`--grid` writes the comp with a labeled 10x10 grid; the model names regions by grid span (`E0:J4`) with a kind (`plate` / `image` / `texture` / `text` / `control` / `chrome`) in a small JSON file; `--regions` measures each region (normalized and pixel box, sampled palette, detail energy, aspect) and writes `.impeccable/build/spec.json`. Raster kinds get a `plate` path under `assets/plates/`. `--crop <id>` extracts the reference crop; `--plate-prompt <id>` prints the regeneration prompt; `--print` is the compact spec the build codes against. The spec is what "anything not in this list does not exist on the page" refers to.
+
+### 3. `build-phase.mjs`: phases as a state machine on disk
+
+`.impeccable/build/state.json`, phases `spec → plates → hero → sections → motion → responsive → review`, advanced only by the script:
+
+- `spec` gate: spec.json exists, measures this comp, has regions.
+- `plates` gate: every raster region's plate exists, decodes, is at least 1.5x the region's pixel width, and scores against the comp crop (`cover` alignment, kind-weighted, min 0.5).
+- `hero` gate: `.impeccable/review/hero-repro.png` exists and comp-diff scores at least 0.72 with no region `missing` and at most a third `contradicted`. Writes `.impeccable/review/diff/hero/`. Attempts and scores are recorded.
+- later phases record the moment; `--force --reason` is allowed and recorded, never silent.
+
+`status` prints a NEXT line for the current phase, so the prose does not have to.
+
+### 4. Plates: `generate-image.mjs --plate <id>`
+
+One raster region end to end: crop the comp region, send the crop as the edits-endpoint reference with the spec's plate prompt (remove UI text and chrome, keep everything else), pick the closest supported size to the region's aspect, write to the plate path, embed the prompt, score against the crop, warn under 50%, refuse under `--min`. `IMPECCABLE_IMAGE_GEN_FAKE=1` yields the crop at 2x so offline pipelines walk the plate gate. Harness-native image tools use the crop and prompt the same way.
+
+The asset producer agent's job shrinks to: produce the spec's plates, one line per plate, `blockers`, `assumptions`. No inventory of its own (the spec is the inventory), no strategy taxonomy.
+
+### 5. Two detector rules
+
+- `organic-clip-path`: `clip-path: polygon()` with 10+ off-grid vertices, or `clip-path: path()` with 3+ curve segments. Geometric clips (cut corners, diagonals, hexagons, arrows) pass; `circle()`/`inset()` pass.
+- `buried-raster`: a `url()` layer under a gradient wash whose stops are all >= 0.9 alpha (or opaque), no blend mode; or a raster background / `<img>` at opacity < 0.15. Tints under 0.9, blends, and visible opacities pass.
+
+Both in both engines (static jsdom + browser bundle), fixtures under `tests/fixtures/antipatterns/`.
+
+### 6. Prose
+
+`new-work.md` section 6 is now the phase list with its gates; the reproduction paragraph, the hero checkpoint paragraph, and visualize.md's inventory / medium-gate / produce sections are gone in favor of the scripts that enforce them. `visualize.md` dropped from 55 to 44 lines and new-work.md's section 6 from ~1,900 to ~1,300 words while gaining the actual mechanism. The finish reviewer reads the state file and the diff report first and starts its matrix from the measured verdicts.
+
+## What this does not do
+
+- It does not judge lettering character, ornament, or motion. The reviewer still owns those.
+- It does not decide plates for the model: the model still names regions on the grid. The gate only refuses to proceed when a named raster region has no plate.
+- The hero threshold (0.72) and plate threshold (0.5) are calibrated on two runs and synthetic perturbations; they will move with evidence. Both are constants at the top of `build-phase.mjs`.
+- Operate surfaces (dashboards, editors) have few or no plates; the spec/diff still apply, the plates gate is trivially satisfied.
+
+## Evaluating it
+
+Execution cuts on the packet-approved niches (`02-kids-reading`, `07-vintage-moto-forum`), old skill (main worktree via `IMPECCABLE_SKILL_DIR`) vs new, scored with `comp-diff.mjs` against the approved comp plus the standard graders. See the PR description for numbers.

skill/agents/impeccable-finish-reviewer.md

@@ -22,16 +22,16 @@ A hard turn ceiling ends the run without warning; a run that ends before its con
 
 ## Input Contract
 
-Expect: the original request; the confirmed user answers; the artifact path(s); the screenshots the parent captured, in `.impeccable/review/` (web: `desktop.png` and `mobile.png`; native: device-class names such as `phone.png` and `tablet.png`, suffixed per OS on adaptive). A screenshot path the calling brief names is authoritative when the file exists; `.impeccable/review/` is where to look when the brief names none or a named path is missing, never a filename you invent. Also expect: the direction contract (THESIS, OWN-WORLD, STORY, FIRST VIEWPORT, FORM); the PRODUCT.md path; existing hook or detector findings; the chosen world's QUALITY BAR card paths; on a comp-led build the approved comp path (a code-led build has none; it passes the chosen decision comp as a separate critique-reference input, labeled as such, and nothing here that binds "the approved comp" binds it); and the skill's `reference/craft-floor.md` path. On a native (`ios` / `android` / `adaptive`) build the packet adds the platform reference path(s) (`reference/ios.md` / `reference/android.md`) and a line saying no detector ran: read the platform reference alongside the craft floor, judge every check in the platform's own conventions, treat the screenshots as device captures, and know your floor check is the build's only slop gate. When the harness can view images, open the screenshots, the comp, and the card first, and inventory the comp's salient elements in your own words before reading the direction contract or any builder-authored summary: a review anchored on the contract inherits whatever the builder's abstraction dropped.
+Expect: the original request; the confirmed user answers; the artifact path(s); the screenshots the parent captured, in `.impeccable/review/` (web: `desktop.png` and `mobile.png`; native: device-class names such as `phone.png` and `tablet.png`, suffixed per OS on adaptive). A screenshot path the calling brief names is authoritative when the file exists; `.impeccable/review/` is where to look when the brief names none or a named path is missing, never a filename you invent. Also expect: the direction contract (THESIS, OWN-WORLD, STORY, FIRST VIEWPORT, FORM); the PRODUCT.md path; existing hook or detector findings; the chosen world's QUALITY BAR card paths; on a comp-led build the approved comp path (a code-led build has none; it passes the chosen decision comp as a separate critique-reference input, labeled as such, and nothing here that binds "the approved comp" binds it); on a comp-led build the build state (`.impeccable/build/state.json`), the measured spec (`.impeccable/build/spec.json`), and the diff directories `.impeccable/review/diff/hero/` and `.impeccable/review/diff/final/` (each holds `side-by-side.png`, `heatmap.png`, `regions/<id>.png` paired crops, and `report.json` with per-region scores and verdicts from `comp-diff.mjs`); and the skill's `reference/craft-floor.md` path. On a native (`ios` / `android` / `adaptive`) build the packet adds the platform reference path(s) (`reference/ios.md` / `reference/android.md`) and a line saying no detector ran: read the platform reference alongside the craft floor, judge every check in the platform's own conventions, treat the screenshots as device captures, and know your floor check is the build's only slop gate. When the harness can view images, open the screenshots, the comp, and the card first, and inventory the comp's salient elements in your own words before reading the direction contract or any builder-authored summary: a review anchored on the contract inherits whatever the builder's abstraction dropped.
 
 ## Checks, in order
 
 0. **Evidence.** Before any other check, verify the required captures exist and every capture is valid. Required: the platform's full viewport set (web: `desktop.png` and `mobile.png`; native: one capture per shipped device class), plus every capture the calling brief names as required, a reported user viewport (`user-<width>.png`) included. Valid: no black or blank regions, content matching what the filename claims (a visit capture showing the About section is invalid), the document top visible where the file claims a full page, dimensions that make sense for the named viewport. A required capture that is absent fails exactly like one that is malformed: a viewport nobody captured is a viewport nobody inspected, and it cannot ship. When any capture fails, the whole review changes shape: return `disposition: recapture` as the first line, then one section, `recapture`, listing each missing or invalid file and what a valid capture of it shows, and stop. Never build a matrix on malformed evidence; a verdict derived from a broken capture launders the breakage into an approval, and the parent owes you a full re-review on valid captures, not a scoring round.
-1. **Persistence.** PRODUCT.md exists. On a comp-led build, `.impeccable/review/hero-repro.png` exists: the hero reproduction checkpoint's capture at the comp's own dimensions; its absence means the reproduction phase ran unproven, a material finding. When DESIGN.md predates this build (an extension or redesign), it matches the built world; on a new world it is written after this review by the documenter, so its absence here is not a finding. When comp-round comps exist under `.impeccable/mocks/`, an approval record exists too: the surface brief naming the approved comp, or an `approved` flag in its sidecar. Comp-round comps with no recorded pick mean the approval point was skipped, a material finding. Files under `.impeccable/mocks/decision/` are exempt: they are the direction round's dealt hand, produced before any comp round, and imply no approval whatever the build path; a code-led build has no comp round at all.
-2. **Fidelity.** Against your own element inventory of the approved comp, never against the contract's summary of it: topology, reading order, focal scale, overlaps and z-order, density, signature geometry, the primary action's treatment (a CTA the comp physically works, dissolves, or stamps is a signature element; its plain-rectangle rendition is contradicted), navigation items and icons, headline levels and scale relationships. Classify every salient element: match, acceptable adaptation, missing, contradicted, or added without approval. Three rows are mandatory in every matrix. TYPE: the display lettering's character, compression, width, weight, contrast, terminals, against the comp's; a face of a different character is contradicted however the layout matches. MATERIAL: an element rendered as flat CSS or clean vector where the comp shows painted, textured, dimensional, or photographic material is contradicted regardless of placement; medium is part of the promise. GROUND: the page field's value and temperature against the comp's, sampled from pixels on both sides when tooling allows rather than judged from memory, and read as the net on-screen result where a texture or tile paints over the base color; a ground warmer or cooler than the comp's is contradicted however faithfully the layout matches, and drift toward the rendition prior (warm cream on light grounds, blue-black slate on dark) is the direction to hunt. With no approved comp, TYPE and MATERIAL do not lapse: judge them against the contract's OWN-WORLD and the world's real materials, and treat faked physicality (CSS bevels, embossing, stamped-metal or chalk effects imitating a material the page never renders) as contradicted on its face; imitation material is the single most reliable mark of machine-made design. GROUND narrows rather than lapses: with no comp to sample, a color OWN-WORLD names is the target and the same warmer-or-cooler judgment applies; when OWN-WORLD names none, there is no GROUND authority, and the review says so in place of a verdict, because a target the reviewer invents turns the check into taste. A critique-reference comp on such a build is provocation, not spec: no element matrix, no adaptation citations, no asset obligations; its one contribution is what the image dared that the build did not, and dares worth adopting enter material_fixes as ordinary ordered fixes. An adaptation counts as intentional only when it cites the user answer, surface brief, accessibility need, or product truth that forced it; an uncited deviation is a defect. A missing signature element, a changed topology, or content added without approval fails fidelity and outranks every craft point in material_fixes. When MATERIAL is contradicted on the focal element, or contradiction is the page rather than the exception, stop ordering repairs: make the first material fix a rebuild directive naming the comp regions to re-derive and the assets to produce; a list of patches against a rejected page launders the rejection into an approval. A fix that requires producing an asset says so explicitly ("produce: <region> as a raster asset"), never phrased as a style adjustment the parent will answer with CSS. The comp is the spec for composition, topology, element inventory, density, lettering character, and material; it is not a pixel spec for semantics, accessibility, or responsive reflow, and that allowance covers translation, never replacement.
+1. **Persistence.** PRODUCT.md exists. On a comp-led build, `.impeccable/build/state.json` exists and its `spec`, `plates`, and `hero` phases are `closed`; a phase closed with a `forced` record is disclosed as a material finding unless the user downgraded the comp in words the packet quotes; a state file whose `hero.gate.score` sits under 0.72, or a missing state file, means the reproduction ran unproven, a material finding, and `.impeccable/review/hero-repro.png` must exist either way. When DESIGN.md predates this build (an extension or redesign), it matches the built world; on a new world it is written after this review by the documenter, so its absence here is not a finding. When comp-round comps exist under `.impeccable/mocks/`, an approval record exists too: the surface brief naming the approved comp, or an `approved` flag in its sidecar. Comp-round comps with no recorded pick mean the approval point was skipped, a material finding. Files under `.impeccable/mocks/decision/` are exempt: they are the direction round's dealt hand, produced before any comp round, and imply no approval whatever the build path; a code-led build has no comp round at all.
+2. **Fidelity.** Start from the measurement, then judge what it cannot: read `.impeccable/review/diff/final/report.json` (and hero) first; every region scored `missing` or `contradicted` is a matrix row in that state unless the paired crop under `regions/` shows the score is wrong, and you say why; a region scored `match` still gets your eye for lettering character and material, which the numbers do not measure. Then, against your own element inventory of the approved comp, never against the contract's summary of it: topology, reading order, focal scale, overlaps and z-order, density, signature geometry, the primary action's treatment (a CTA the comp physically works, dissolves, or stamps is a signature element; its plain-rectangle rendition is contradicted), navigation items and icons, headline levels and scale relationships. Classify every salient element: match, acceptable adaptation, missing, contradicted, or added without approval. Three rows are mandatory in every matrix. TYPE: the display lettering's character, compression, width, weight, contrast, terminals, against the comp's; a face of a different character is contradicted however the layout matches. MATERIAL: an element rendered as flat CSS or clean vector where the comp shows painted, textured, dimensional, or photographic material is contradicted regardless of placement; medium is part of the promise. GROUND: the page field's value and temperature against the comp's, sampled from pixels on both sides when tooling allows rather than judged from memory, and read as the net on-screen result where a texture or tile paints over the base color; a ground warmer or cooler than the comp's is contradicted however faithfully the layout matches, and drift toward the rendition prior (warm cream on light grounds, blue-black slate on dark) is the direction to hunt. With no approved comp, TYPE and MATERIAL do not lapse: judge them against the contract's OWN-WORLD and the world's real materials, and treat faked physicality (CSS bevels, embossing, stamped-metal or chalk effects imitating a material the page never renders) as contradicted on its face; imitation material is the single most reliable mark of machine-made design. GROUND narrows rather than lapses: with no comp to sample, a color OWN-WORLD names is the target and the same warmer-or-cooler judgment applies; when OWN-WORLD names none, there is no GROUND authority, and the review says so in place of a verdict, because a target the reviewer invents turns the check into taste. A critique-reference comp on such a build is provocation, not spec: no element matrix, no adaptation citations, no asset obligations; its one contribution is what the image dared that the build did not, and dares worth adopting enter material_fixes as ordinary ordered fixes. An adaptation counts as intentional only when it cites the user answer, surface brief, accessibility need, or product truth that forced it; an uncited deviation is a defect. A missing signature element, a changed topology, or content added without approval fails fidelity and outranks every craft point in material_fixes. When MATERIAL is contradicted on the focal element, or contradiction is the page rather than the exception, stop ordering repairs: make the first material fix a rebuild directive naming the comp regions to re-derive and the assets to produce; a list of patches against a rejected page launders the rejection into an approval. A fix that requires producing an asset says so explicitly ("produce: <region> as a raster asset"), never phrased as a style adjustment the parent will answer with CSS. The comp is the spec for composition, topology, element inventory, density, lettering character, and material; it is not a pixel spec for semantics, accessibility, or responsive reflow, and that allowance covers translation, never replacement.
 3. **Ceiling.** Against the QUALITY BAR card: name the world's native devices the build left unused, frame, depth, lettering treatment, ornament density, motion. The card governs commitment and finish, never composition.
 4. **Contract, promise by promise.** First verify FORM carries the seed key the concept roll printed; a contract with no seed key, or one the parent cannot corroborate, means the roll was skipped, a material fix ahead of any craft point. Then, for each of the five blocks: does the render keep the promise? Apply the memory test to the first viewport.
-5. **Truth.** Demonstration data authored and labeled synthetic; no invented commercial claims; unanswered claims present as marked placeholders, not omissions. Every image-native region of the approved comp shipped as a real asset, not a gradient standing in for one, and every produced asset visibly present in the screenshots; an asset applied at near-zero opacity or buried behind other paint is a compliance token, not a shipped material.
+5. **Truth.** Demonstration data authored and labeled synthetic; no invented commercial claims; unanswered claims present as marked placeholders, not omissions. Every raster region of the spec shipped as its plate (the spec names the file; the page references it; the region's diff row is not `missing`), not a gradient, an inline SVG, or a many-vertex `clip-path` standing in for it, and every produced asset visibly present in the screenshots; an asset applied at near-zero opacity or buried behind a wash is a compliance token, not a shipped material, and the detector's `buried-raster` and `organic-clip-path` findings in the packet are material fixes.
 6. **Floor.** Read the craft floor's Refuse list and hold the screenshots against it: kickers and eyebrows, hard offset shadows outside a neobrutalist world, glyph icons, system display faces, gradient text, side stripes, and the rest. A banned element is a material fix even when it matches nothing in the comp: the builder loaded the same ban before writing it, and fidelity to a comp cannot authorize what the floor refuses. The parent's hook findings cover this mechanically where hooks run; this check exists because hookless harnesses reach you with none, and the last two live sessions shipped five kickers past a reviewer that never looked.
 
 Do not run a second detector pass; mechanical findings belong to the parent's hooks.

skill/reference/new-work.md

@@ -90,30 +90,50 @@ For `shape`, return the selected direction to [shape.md](shape.md) and stop befo
 
 ## 6. Build with full commitment
 
-When an approved comp exists, the comp is king, and the build happens in phases. The comp is a spatial contract, not a mood board: only the user can downgrade its authority, in explicit words, and difficulty never infers a downgrade. Phase one is reproduction: rebuild the comp at its own breakpoint until a screenshot at the comp's width and height overlaps it near pixel-perfectly, materials, components, elevation, assets, and implied design language included. Exactly three concessions exist: fonts (the closest obtainable face), icons (exact match unless the user already chose an icon library), and genuine defects in the generated comp such as spelling errors. Everything else must match, and models systematically believe their HTML, CSS, and SVG recreation succeeded when it did not, so the overlap comparison is the authority, never your conviction: set the screenshot beside the freshly reopened comp image at identical dimensions after every region, never beside your memory of it, and when a region keeps losing that comparison, stop recreating it in code and produce it as a rendered asset composited into the page. The comp also outranks every written record of it: when the recorded brief or inventory commits to less than the comp shows, a softer texture, a sparser field, a sculpted plate reduced to flat CSS, correct the record upward to the comp; qualifiers like subtle, restrained, and low-contrast, and counts rounded down to a comfortable fraction, are how approved materials die between approval and build. A produced material must then survive to the screen: a texture buried under a nearly opaque color wash ships the wash, not the material, so judge every material by the screenshot beside the comp, never by the stylesheet. Every color the brief records gets that comparison by number, not by eye: sample the build screenshot's ground, dominant fields, and accents the same way each record was taken (an interior patch average where the record is an average, both end colors where the record is a gradient) and set each value against its recorded counterpart (sampled from the comp itself when the brief lacks one), and when a texture or tile paints over a base token, measure the net on-screen value, because the eye files a drifted color under the same color word and the number is what catches it. Judge the gap like a colorist, not a diff tool: a difference with a color name (warmer, grayer, darker than the record) is drift to fix, while a few digits of render and compression noise are the same color. <!-- rule:skill-color-by-number --> Only when reproduction holds does phase two begin: static regions that should live become animated or interactive, reveals and motion are added, then responsiveness across the surface's devices. Where the comp does not cover the whole surface, continue building the remainder inside the comp's recorded world and design language; a component the comp never shows inherits the recorded system's corner language, line weights, and materials, and may not introduce container styles, border weights, or chrome the comp never uses. <!-- rule:skill-comp-is-king -->
+Build the assigned direction, not a safer interpretation of it. The form supplies structure, reading order, component conventions, and native motion; the product supplies every fact. Commit every atom: nav, buttons, inputs, and links are rebuilt in the form's vocabulary, and a stock component inside a committed form is a lapse. Land the first build fully committed; the passes that follow exist to make the committed thing clear and effective, never to dilute it. In unattended work, the safe rendition is the known risk. <!-- rule:skill-commit-every-atom -->
 
-Build the assigned direction, not a safer interpretation of it. The form supplies structure, reading order, component conventions, and native motion; the product supplies every fact. Commit every atom: nav, buttons, inputs, and links are rebuilt in the form's vocabulary, and a stock component inside a committed form is a lapse. Land the first build fully committed; committing is the hard part, and the passes that follow exist to make the committed thing clear and effective, never to dilute it. In unattended work, the safe rendition is the known risk. <!-- rule:skill-commit-every-atom -->
+### Comp-led: the comp is a measured contract
+
+When an approved comp exists, it is a spatial contract, not a mood board: only the user can downgrade its authority, in explicit words. Models systematically believe their HTML, CSS, and SVG recreation of an image succeeded when it did not, so the build runs as a state machine on disk whose gates measure the screen against the comp instead of asking you to remember it. Start it once, and let it tell you what is next:
+
+`node {{scripts_path}}/build-phase.mjs start --comp <approved comp>`
+
+Then, in order, each closed by `node {{scripts_path}}/build-phase.mjs advance` (exit 2 means the gate failed and printed why; fix that and advance again; write nothing for a later phase while an earlier gate is open):
+
+1. **spec.** Measure the comp: `comp-spec.mjs --comp <comp> --grid` writes a coordinate grid over the comp; open it, name every salient region by grid span in a regions file (kind `plate` / `image` / `texture` for anything painted: every illustration, photograph, figure, product object, and material texture; `text` / `control` / `chrome` for what code draws), and run `comp-spec.mjs --comp <comp> --regions <file>`. The spec carries each region's box, sampled palette, and medium; `comp-spec.mjs --print` is the build's reference from here on. Anything not in the spec does not exist on the page: no borders, rules, containers, or chrome the comp does not show. Only three concessions exist: fonts (the closest obtainable face), icons (exact match unless the user chose an icon library), and genuine defects in the comp such as spelling errors. <!-- rule:skill-comp-spec -->
+2. **plates.** Every raster region ships as a plate: the region regenerated at asset resolution from its comp crop, UI text removed, at its `plate` path. `generate-image.mjs --plate <id>` does one region end to end and scores it against the crop; a harness-native image tool takes the crop (`comp-spec.mjs --crop <id>`) as its input image and `comp-spec.mjs --plate-prompt <id>` as its prompt, then `embed-prompt.mjs`. With parallel subagents, spawn the shipped asset producer (`impeccable-asset-producer`; `impeccable_asset_producer` in codex; `/impeccable-asset-producer` in Cursor; on GitHub Copilot say "Use the impeccable-asset-producer agent") with the spec path and let it produce them all; without subagents, produce them here. A crop of the comp is a reference, never a shipping pixel. The gate checks every plate exists, is at least 1.5x the region's size, and reads as the region. Page code waits for this gate: a page written before its plates exist is a page that draws its material in CSS. <!-- rule:skill-plates-before-page -->
+3. **hero.** Build only the first viewport, at the comp's own dimensions, from the spec's boxes, palette, and plates; capture it into `.impeccable/review/hero-repro.png` at those dimensions; advance. The gate runs `comp-diff.mjs`, writes `.impeccable/review/diff/hero/` (side-by-side, heatmap, one paired crop per region, `report.json`), and passes at 72% overall with no region missing. When it fails, open the side-by-side and the worst region pairs it names, fix those regions, recapture, and advance again; the numbers rank, the crops decide. This is where the run's ambition is won or lost, and a retry here costs minutes where a rebuild verdict at the finish costs the run. <!-- rule:skill-hero-gate -->
+4. **sections.** Build the rest of the surface inside the spec's system: the same corner language, line weights, and palette, and nothing the comp never shows. Where the comp does not cover a region, it inherits the recorded system.
+5. **motion.** The signature interaction, reveals, and motion, orchestrated once rather than scattered.
+6. **responsive.** The other viewports. A comp'd surface that is mobile-first was comped portrait; the plates were produced for that frame.
+
+Force a gate only with `--force --reason`, and only when the user has downgraded the comp's authority in words; the record travels to the reviewer.
+
+### Code-led
+
+No comp and no apology for it: the ambition lives in the direction contract's FIRST VIEWPORT block and the named signature interaction, and the finish reviewer audits those promises in behavior. The chosen decision comp rides to the finish review as the critique reference.
+
+### Both paths
 
 - **The first viewport is a thesis, not a header.** Demonstrate the mechanism immediately, at the scale the form has in life; do not trap the concept inside a standard hero or card shell. The memory test: if someone left after one viewport, what would they describe an hour later? If the honest answer is a mood, the concept has not committed yet.
-- **Prove the hero before building past it.** When an approved comp exists, render the first viewport, capture it at the comp's own pixel dimensions, and set it beside the comp's first viewport before any later section: the hero carries the run's ambition, and every following section inherits its shortfall. Save that capture as `.impeccable/review/hero-repro.png` (create the directory); the finish reviewer verifies it exists, so a skipped checkpoint is a visible checkpoint. Judge scale and density as quantities, a field at a tenth of the comp's coverage or type at half its weight is a different design, and a five-minute retry here is what a rebuild verdict at the finish costs when this check is skipped. <!-- rule:skill-hero-checkpoint -->
-- **Prove, don't claim.** Show the subject doing its job: the interface at work, the mechanism dramatized, specifics a competitor could not copy-paste. Sections that restate a claim in different words add length, not substance. Demonstration data is design material: author it at full fidelity and label it synthetic; claims stay uninventable.
-- **Author the assets; never substitute chrome.** Great surfaces live on carefully made content: names, entries, copy, covers, thumbnails, textures. In greenfield work every blank the ask round left open is yours to author at production fidelity; content is authorable, claims are labelable, no section is omittable. An unanswered commercial claim ships as a clearly marked placeholder on the user's replacement list. When image generation exists, producing the design's imagery is part of building, at the scale the composition needs: a viewport that wants atmosphere gets a full-bleed layered scene, and a library of small centered subjects standardized for tidiness forecloses it. Gradients, glass, and generic icon tiles where an authored asset belongs are the gap wearing chrome; icons drawn in the world's own grammar are the remedy, not the target. <!-- rule:skill-author-assets-not-chrome -->
-- **Build the form's web leverage.** When the chosen world names a technique (canvas, WebGL, view transitions, generative motion), build the technique itself, not a static imitation of it; the graceful fallback serves constrained clients, it is not the default experience.
+- **Prove, don't claim.** Show the subject doing its job: the interface at work, the mechanism dramatized, specifics a competitor could not copy-paste. Demonstration data is design material: author it at full fidelity and label it synthetic; claims stay uninventable.
+- **Author the assets; never substitute chrome.** Great surfaces live on carefully made content: names, entries, copy, covers, thumbnails, textures. In greenfield work every blank the ask round left open is yours to author at production fidelity; content is authorable, claims are labelable, no section is omittable. Gradients, glass, generic icon tiles, and many-vertex `clip-path` polygons where an authored asset belongs are the gap wearing chrome; the detector flags the last two. <!-- rule:skill-author-assets-not-chrome -->
+- **Build the form's web leverage.** When the chosen world names a technique (canvas, WebGL, view transitions, generative motion), build the technique itself, not a static imitation of it.
 - **Pace the scroll like a studio.** Vary density, scale, image, motion, and quiet inside one grammar; a dense passage earns a quiet one, and the page ends anchored by a real close. One spacing rhythm throughout, with more space above a heading than below it.
 - **Use real, verified imagery when the brief implies it.** Search for the subject's physical object rather than the category; one decisive photo beats five mediocre ones. Verify stock URLs resolve.
-- **Author motion as material.** The form has native motion, what it does in life between states; give the page that motion once, orchestrated, rather than scattered hover effects. Bound expensive effects and keep content visible by default.
+- **Author motion as material.** Give the page the form's native motion once, orchestrated, rather than scattered hover effects. Bound expensive effects and keep content visible by default.
 
 Preserve semantics, accessibility, performance, responsiveness, project conventions, and working behavior.
 
 ## 7. Inspect and finish
 
-Inspect the surface's target sizes in one batched screenshot round: desktop and mobile on the web; on a native platform (`ios` / `android` / `adaptive`), the shipped device classes per OS, captured from the simulator or emulator the way the platform reference's Verifying the build section describes. When the harness reports the user's actual viewport (an in-app browser's size, a named resolution), add that width to the set: the width that breaks is the one the user sees first. Critique the render against the user's request and the direction contract, fix material gaps, and confirm with one final round; two rounds is the ceiling, and fixes batch between them rather than earning per-tweak screenshots. When an approved comp exists, the critique is a side-by-side: view the comp region and the build region together, the hero and each section as its own crop at legible scale, never one full-page thumbnail, which hides exactly the failures that matter, crude controls, wrong lettering character, flattened material, behind a superficially similar section order. On a Persuade surface, verify the mode did its job: a first-time visitor should know what this is, why it matters, and what to do within seconds, in the form's own vocabulary.
+Inspect the surface's target sizes in one batched screenshot round: desktop and mobile on the web; on a native platform (`ios` / `android` / `adaptive`), the shipped device classes per OS, captured from the simulator or emulator the way the platform reference's Verifying the build section describes. When the harness reports the user's actual viewport (an in-app browser's size, a named resolution), add that width to the set: the width that breaks is the one the user sees first. Critique the render against the user's request and the direction contract, fix material gaps, and confirm with one final round; two rounds is the ceiling, and fixes batch between them rather than earning per-tweak screenshots. On a comp-led build, run `node {{scripts_path}}/comp-diff.mjs --comp <approved comp> --build .impeccable/review/desktop.png --spec .impeccable/build/spec.json --out-dir .impeccable/review/diff/final` and read its region rows and paired crops as the critique: the side-by-side is the view the build thread never has on its own, and a region it scores missing or contradicted is a fix whatever the page looks like from memory. Never judge fidelity from one full-page thumbnail; it hides exactly the failures that matter. On a Persuade surface, verify the mode did its job: a first-time visitor should know what this is, why it matters, and what to do within seconds, in the form's own vocabulary.
 
 A capture is evidence only when it is valid, and you validate before you send. Settle or disable entrance motion first: an element hidden by animation timing reads as a missing element and gets fixed into a regression. Capture full-page shots from the document top. Capture the comp comparison at the comp's own pixel dimensions. Then open every file once and confirm it shows what its name claims: no black or blank regions, no wrong section behind a right filename, no half-loaded state. A malformed capture sent onward costs the whole round; the reviewer answers it with `disposition: recapture` and nothing it reviewed binds. <!-- rule:skill-capture-validity -->
 
 After the second inspection round the build thread's polishing is over: no further defect hunts, micro-edit scripts, or rebuilds here; whatever remains ships through the handoffs, where a fresh context does the finding better and cheaper. On the web, where this harness runs no design hook, run `node {{scripts_path}}/detect.mjs --json` on the changed targets once here, fix what is mechanical, and pass the remaining findings to the reviewer; a hookless web build that skips this ships every tell the hook exists to catch. A native platform skips the detector entirely: it reads HTML and CSS and has no verdict on native code, so the reviewer's floor check is the only slop gate and the input packet says so. Capture the screenshots into `.impeccable/review/`, one file per captured viewport (on the web, `desktop.png` and `mobile.png`, plus `user-<width>.png` whenever the user's viewport joined the inspected set; on native, one per device class, such as `phone.png` and `tablet.png`, suffixed per OS on adaptive), creating that directory when the harness does not; the paths you pass the reviewer are its spec, every viewport you inspected is named required in the packet, and that directory is where it looks when a passed path is missing.
 
-Then spawn the shipped finish reviewer, `impeccable-finish-reviewer` (`impeccable_finish_reviewer` in codex; `/impeccable-finish-reviewer` in Cursor; on GitHub Copilot say "Use the impeccable-finish-reviewer agent"), with the original request, confirmed answers, the artifact path, the screenshot paths, the direction contract, existing hook findings, the QUALITY BAR card and approved comp paths (a code-led build has no approved comp; the chosen decision comp rides in that slot as the critique reference, named as such), the craft-floor reference path, and on a native platform the platform reference path(s), [ios.md](ios.md) / [android.md](android.md), both on adaptive, plus one line saying no detector ran, so the reviewer judges in the platform's conventions rather than the web's. The reviewer has no browser; screenshots you fail to pass are checks it cannot run. Never read the shipped agents' definition files before spawning; the harness loads them at spawn, and you owe only the input packet. Wait on any agent with one long timeout rather than a loop of short polls, and spend the wait on the next independent step. Verify the return carries the five contract sections (a recapture return carries one, its recapture list); on an empty or thrashed return, respawn once with the same inputs. This review never runs inside the build thread and never inherits it: spawn the reviewer fresh, with no forked conversation history (`fork_turns: 0` in codex); a reviewer that inherits your transcript inherits your framing, your optimism, and your abstractions, and everything it needs travels in the inputs above. Only a harness with no subagent capability at all substitutes a fresh in-thread pass after stepping fully out of the build context, run from [degraded/finish-reviewer.md](degraded/finish-reviewer.md), and a substituted or failed-and-replaced review is disclosed in one line at finish, never silently. <!-- rule:skill-finish-separate-reviewer -->
+Then spawn the shipped finish reviewer, `impeccable-finish-reviewer` (`impeccable_finish_reviewer` in codex; `/impeccable-finish-reviewer` in Cursor; on GitHub Copilot say "Use the impeccable-finish-reviewer agent"), with the original request, confirmed answers, the artifact path, the screenshot paths, the direction contract, existing hook findings, the QUALITY BAR card and approved comp paths (a code-led build has no approved comp; the chosen decision comp rides in that slot as the critique reference, named as such), on a comp-led build the build state (`.impeccable/build/state.json`), the spec, and the diff directories (`.impeccable/review/diff/hero/` and `.impeccable/review/diff/final/`, whose side-by-side, heatmap, region pairs, and `report.json` are the fidelity evidence), the craft-floor reference path, and on a native platform the platform reference path(s), [ios.md](ios.md) / [android.md](android.md), both on adaptive, plus one line saying no detector ran, so the reviewer judges in the platform's conventions rather than the web's. The reviewer has no browser; screenshots you fail to pass are checks it cannot run. Never read the shipped agents' definition files before spawning; the harness loads them at spawn, and you owe only the input packet. Wait on any agent with one long timeout rather than a loop of short polls, and spend the wait on the next independent step. Verify the return carries the five contract sections (a recapture return carries one, its recapture list); on an empty or thrashed return, respawn once with the same inputs. This review never runs inside the build thread and never inherits it: spawn the reviewer fresh, with no forked conversation history (`fork_turns: 0` in codex); a reviewer that inherits your transcript inherits your framing, your optimism, and your abstractions, and everything it needs travels in the inputs above. Only a harness with no subagent capability at all substitutes a fresh in-thread pass after stepping fully out of the build context, run from [degraded/finish-reviewer.md](degraded/finish-reviewer.md), and a substituted or failed-and-replaced review is disclosed in one line at finish, never silently. <!-- rule:skill-finish-separate-reviewer -->
 
 Act on the disposition word; there are exactly four. **recapture**: the evidence failed, not the build. Recapture what the return names under the capture-validity rules, then run a full review over the new evidence. A review conducted on invalid evidence binds nothing, and a verdict pass may never follow it. **rebuild**: fidelity failed wholesale, not in patches. Skip the fix batch and execute the rebuild immediately: re-derive the named regions, produce the named assets, and send the result back for a fresh full review, never a verdict pass; a rebuild replaces regions wholesale, so the whole matrix runs again over the recaptures. Tell the user what is happening rather than asking permission to fix a failure. Consult the user only on a second rebuild directive, both verdicts on the table, or when rebuilding would discard content the user approved. **ship**: nothing is owed; report the verdict at its scope and continue to the documenter. **fix**: apply the material fixes in one batch, rebuild once, and recapture the same viewports over the same files. A recapture measures positions, loading, and overflow; it cannot measure whether a fix reached the quality the finding named, so send the recaptured screenshots back to the same reviewer for a verdict scoring every material fix resolved, partial, or unresolved (through the harness's agent continuation; without one, run the scoring fresh from [degraded/finish-reviewer.md](degraded/finish-reviewer.md)'s Verdict Pass). Fixes scored partial or unresolved get another batch, recapture, and verdict. Two rounds is the budget an unattended run ends at; an attended session's ceiling belongs to the user, so when the second verdict still lists open items, put the table in front of them and let them choose between shipping as it stands and funding another round. Whoever decides, stop the moment a round resolves nothing, and the reviewer's findings are the only list you work from, never your own re-opened hunt. Do not run a second detector. <!-- rule:skill-verdict-bounds-the-finish -->
 

skill/reference/visualize.md

@@ -29,28 +29,16 @@ This approval point has no substitute and no skip condition. When the structured
 
 After approval, record the choice where tools can find it: the approved comp's path goes in the surface brief, and its `.json` prompt sidecar gains `"approved": true` (every comp generated through `generate-image.mjs` has one; create it if a native tool didn't). The sidecar travels with the mocks folder, so the approval survives sessions and machines that never see the brief. Summarize the composition and the parts of the comp that must not be literalized, return to new-work.md, record the direction contract from the approved concept, and build.
 
-## Inventory implementation fidelity
+## After approval: the comp becomes a spec
 
-Before building, read the approved comp as a design system and record it in the brief: component grammar, corner language, line weights, elevation treatment, and the type ramp. Everything the comp does not show gets built from this record; without it the fallback is the model's stock kit of square boxes, 1px grids, bento cells, and hard shadows. Then inventory the comp's major visible ingredients in writing (a short table in the surface brief or working notes; the finish reviewer audits shipped assets against it) and choose an implementation medium for each: semantic HTML/CSS/SVG, existing project asset, generated raster, sourced raster, icon library, canvas/WebGL, or accepted omission. The same inventory names the comp's compositional commitments: navigation items and icons, headline levels and their scale relationship, signature geometry such as seams, masks, and overlaps, and each section's arrangement and density. The primary action gets its own row with its own medium: when the comp dissolves, stamps, erodes, or otherwise physically works the main CTA, that treatment is signature material on the page's most important element, and shrinking it to a border trick is the compliance-token version of commitment. An element never written down is the element the build silently drops; the direction contract's 150 words cannot carry this list, so it lives here.
+The approved comp is a north star for translation into semantic, responsive, accessible code, never a license to recompose: keeping the palette and mood while redrawing the topology is a second art direction. Do not rasterize core UI text or controls. Do not substitute a different visual driver after approval without asking.
 
-The record is sampled, never estimated: read the comp's page **ground**, each dominant field, and each accent's actual hex from its pixels (ImageMagick, Python with PIL, any pixel-reading tool on the machine) and write the values into the same record. Take a flat field from any interior pixel, a textured or grainy one as the average of an interior patch (crop a swatch, scale it to one pixel), and a gradient as its two end colors; never sample an edge, where antialiasing blends neighbors into colors the design never chose. An adjective is a direction, not a record: cream covers everything from near-white to beige, charcoal a third of the value scale, and wherever no number pins a color, the rendition prior picks the spot. Sampled values supersede the palette chips on the decision and composition cards: those were authored before this comp existed, and a chip that disagrees with the comp's pixels is a draft the approval retired.
+What the comp shows is measured, not remembered. new-work.md section 6 runs the build as phases (`build-phase.mjs`): the spec phase turns the comp into region boxes with sampled palettes (`comp-spec.mjs`), and the medium of every region follows from what the pixels are, never from what feels buildable: a figure, a product object, machinery, any illustration with perspective, shading, or drawing skill in it, and any texture by name (woven cloth, paper grain, fabric, leather, brushed metal) is a `plate` / `image` / `texture` region and ships as a raster; text, controls, chrome, diagrams with countable elements, flat shape systems, and anything that must move, scale, or respond are semantic. Writing "CSS" for a sculpted panel's finish, or a many-vertex `clip-path` for a torn edge, is the quiet deletion of the approved design; the detector's organic-clip-path and buried-raster rules and the hero gate's region scores catch it. Dropping an image-native region is a scope decision the user makes at the approval point, never a silent flattening after it. Generated imagery is a material, not a claim: evidence rules bind assertions, specs, testimonials, and photographs presented as real, never render fidelity.
 
-The medium column is where an approved design most often dies, so it obeys a gate: the medium is decided by what the comp region shows, never by what feels buildable in the current stack. A human figure, a product object, machinery, or any material with lighting and depth is raster whatever the stack; so is any texture by name alone: woven cloth, paper grain, fabric, leather, brushed metal need no depth argument, because a CSS gradient is not a texture medium and "layered CSS textures" is not a medium at all. Writing "silhouette" for a photographic figure, or "CSS" for a sculpted panel's finish, is not a medium choice; it is the quiet deletion of the approved design, and it is how a comp full of physical material becomes a flat page with the same section order. Style does not move this boundary: a comp region with perspective, shading, figure drawing, or dense mechanical detail is illustration however line-drawn it looks, and no build session can author illustration as vectors, so it regenerates as raster like any photograph. Authored SVG covers what a session can specify exactly (diagrams with countable elements, controls, flat shape systems) and ends where drawing skill begins; an instruction-manual world keeps its illustrations as line-art illustrations, not diagrams. Produce such regions by regenerating them cleanly, with the approved comp and its embedded prompt as the reference for a fresh render at asset resolution; never crop pixels out of the comp itself, whose effective resolution sits far below asset grade. Dropping an image-native region is a scope decision the user makes at the approval point, never a silent flattening after it. Generated imagery is a material, not a claim: evidence rules bind assertions, specs, testimonials, and photographs presented as real, never render fidelity; "no photography on hand" forbids fake proof, not an illustrated hero.
+## Plates and provenance
 
-The gate runs both ways: precise geometry, hard-edged shape systems, diagrams, expressive motion, shaders, and anything interactive are vector and GPU territory (SVG, canvas, WebGL), where a raster flattens what should move, scale, and respond. A field or texture built from many small elements carries a quantity commitment either way: write down its approximate density and coverage ("thousands of glyphs over two-thirds of the fold, dense at the top fading into the path"), because a field rebuilt at a tenth of its density passes every checklist and still is not the design. TYPE rows carry the same discipline: name the face's compression class, and render one headline word against the comp before building on it; a visibly wider or lighter silhouette means the face is wrong, and every section built on it inherits the miss. Raster is for what the world paints; code is for what the world draws, animates, or reacts with, and choosing code there is ambition, not economy. Every `produce` entry is produced before the build ships, through the asset producer or in the current thread; an inventory with unproduced entries is an unfinished build, and this gate is where imagery-free pages come from when it is skipped.
-
-Pay special attention to the dominant composition, signature use, image-native content, second-fold system, and any interaction the still image only implies.
-
-The comp is a north star, not something to trace, and know what that allows: translation into semantic, responsive, accessible code, never recomposition. Keeping the palette and mood while redrawing the topology is a second art direction, not an adaptation. Do not rasterize core UI text or controls. Do not substitute a different visual driver after approval without asking.
-
-## Produce only the assets the build needs
-
-Generation context is part of the asset: a build composed by a thread that never saw the prompts places assets it does not understand. Prefer generating build-critical imagery in the build thread when the budget allows; when a subagent produces assets instead, every asset carries its prompt, and the builder reads those prompts before composing. The carrier is uniform across harnesses: after generating any image with any tool, native or `generate-image.mjs` (which does it automatically), run `node {{scripts_path}}/embed-prompt.mjs <image> --prompt "<prompt>"` with the exact string the generation tool received, pasted whole, so the intent lives inside the file and survives copies between machines and harnesses; a summary reconstructed from memory records an asset that was never made. `--read` recovers the prompt from any impeccable-generated image, and `--scan <dir>` lists every raster in a directory still missing one. The embedded prompt plus the asset's row in the written inventory is the raster's **provenance**, and every raster the artifact references carries it; a sourced, stock, or pre-existing raster with no generation prompt embeds its origin instead.
-
-Provenance is owed for the run, not the build phase: a raster created or replaced later, in a fix batch or a reviewer's rebuild, is produced under this same section, prompt embedded and inventory row added, because the inventory is how the next thread knows what ships. A raster a fix abandons or supersedes is deleted from the assets directory in the same batch; an unreferenced raster with no record is a provenance leak, not a spare.
-
-When the harness runs subagents, spawn the shipped asset producer every time, even when the inventory's produce bucket looks empty: its manifest is the independent second opinion on your media, and the runs that skipped the spawn are the runs whose cotton became CSS. An honestly empty manifest costs one cheap spawn; a wrongly empty produce bucket costs the build its materials. Use `impeccable-asset-producer` (`impeccable_asset_producer` in codex; `/impeccable-asset-producer` in Cursor; on GitHub Copilot say "Use the impeccable-asset-producer agent"): give it the approved comp, output paths, required dimensions and formats, transparency needs, crop notes, and what must remain semantic code. Without subagents, produce the minimum required assets in the current thread by the book: load [degraded/asset-producer.md](degraded/asset-producer.md) and follow it inline, with whatever generation exists.
+Every raster region's plate is produced in the plates phase, before any page code, by the shipped asset producer or in the current thread (`generate-image.mjs --plate <id>`, or the harness image tool with the crop as input and the spec's plate prompt). Generation context is part of the asset: after generating any image with any tool, run `node {{scripts_path}}/embed-prompt.mjs <image> --prompt "<prompt>"` with the exact string the tool received (`generate-image.mjs` does this itself), so the intent lives inside the file; `--read` recovers it, `--scan <dir>` lists rasters still missing one. The embedded prompt plus the region's row in the spec is the raster's **provenance**, and every raster the artifact references carries it; a sourced, stock, or pre-existing raster embeds its origin instead. A raster created or replaced later, in a fix batch or a reviewer's rebuild, is produced the same way; a raster a fix abandons is deleted in the same batch.
 
 Convert images with a converter context.mjs reported at boot (the IMAGE_TOOLS line); probe only when it reported none, at most once per session, never per image.
 
-Return to [new-work.md](new-work.md) for the direction contract, implementation, and the finishing pass.
+Return to [new-work.md](new-work.md) for the direction contract, the phased build, and the finishing pass.