# Atelier and Studio · Making it, and being answerable for it · receipt · 13 September 2026

Chapter two of _Creation_. Delivered in the Eve edition on 13 September 2026.
Eleven inherited slides become a fifteen-slide chapter under the
`atelier-studio` id. The brief is
[authoring/atelier-studio-chapter.md](../../authoring/atelier-studio-chapter.md).
Machine-readable state is in [state.json](state.json).

## What was delivered

Fifteen slides: eleven inherited slides rewritten in place and four new — a
cover, two section dividers and a close. Nothing merges away and nothing is
carried. `content/07-creation-and-studio.json` held nothing but this chapter's
eight slides and **retires**; `content/24-creation-deepdives.json` gives up
three of its twelve and keeps nine for the Living Scenes chapters.

`editorial-lifecycle` is also carried by the `review-safety` guide, so that
deck re-rendered with the rewritten slide. No guide lost a slide.

## Evidence

**A real capture**, the first in this track. A new harness,
`tools/presentations/tests/atelier-room-capture.mjs`, renders the shipped
`AtelierIndex` from the pin over its own fixture and asserts three
characterisations off the rendered DOM: the component's own room count (5) and
house-rule count (3), the three house rules verbatim, and the handoff a scene id
the member does not own actually produces — `data-atelier-active-scene="missing"`,
the headline "Draft unavailable.", and **zero** elements carrying a draft title.
Two images are installed under `assets/atelier/`, with `provenance.json` naming
every pinned file read.

Everything else below was executed at the pin on this host.

- **The editorial machine, walked whole.** Twelve states; of the 144 ordered
  pairs exactly **21** are legal. `archived` is the only terminal state, `idea`
  the only state nothing returns to, and no state may transition to itself.
- **Two state names in the architecture diagram are not in the code.**
  `rejected` and `retracted` were both checked against the enum and are absent;
  the code says `changes-requested` and `takedown`.
- **`tryEditorialAdvance` has four outcomes, not the three the feature page
  lists**: allowed, `transition-not-allowed`, `gate-evidence-absent` (required
  and never brought) and `unsatisfied-gates` (brought and not satisfied), with
  `absentGateIds` and `unsatisfiedGateIds` reported separately. A gate the
  checklist never asked for, brought as required and unsatisfied, still refuses;
  the same gate brought optional does not.
- **A replacement is classified before anything is invalidated.** Executed over
  five replacements: identical bytes invalidate **0 of 3** dependents; a
  re-encode, a different photograph, a kind change and an unclassifiable
  replacement each invalidate **3 of 3**. A missing perceptual hash yields
  `unclassifiable` rather than `different-media`, because guessing cheaply would
  waive an accessibility re-review nobody performed.
- **The cascade refuses to guess its own reach.** It takes its dependent list
  from the caller with a `usageCoverage` of `complete` or `partial`, because
  `usedByArtifactIds` has exactly one producer in the repository — `[]` at
  upload, verified by grep at the pin. `surfaceBrokenLinks` follows the same
  rule: executed here, an empty dependent list still raises the rights-expired
  signal under `partial` coverage and is silent only under `complete`.
- **One neighbouring alarm has not caught up.** `buildRightsExpiryAlerts` grades
  severity correctly (block once expired, warn inside the final third, info
  earlier) but its `downstreamCount` still reads `usedByArtifactIds`, so it is
  `0` for every asset the product creates. Recorded on the slide as "read the
  severity; do not read the count as a blast radius."
- **The lineage rule that used to drop every real edge.** Both readers admitted
  an edge only when both endpoints were visible records; every edge the release
  path emits ends at a `job:` node, which is never a record. Executed here over
  the edge shape the release path emits: the old rule admits **0 of 5**, the
  shipped anchored rule admits **3** and still withholds the two that reach out
  to a foreign tenant's output.
- **Derivation is bounded by the class, not the member.** `planBranch` executed
  over six deltas refuses `unknown-parameter`, `out-of-range` and `wrong-type`,
  and a valid branch inherits the parent's `workflowClassId`.
- **Staleness is a tri-state.** Executed over nine scenarios: a moved source
  hash is `stale`; a withdrawn memory entry and a corrected one are `stale` for
  their own reasons; a segment whose inputs nobody indexed is **`unverifiable`**,
  not `fresh` — a different instruction. One source edit over eight locale rows
  returns 8 of 8 stale, all for one reason.
- **Rollback demands the exact blast radius.** Executed over five confirmation
  sets: one id short, one id extra and an empty confirmation all refuse; the
  same ids in another order are accepted; a blank public note refuses.
- **What an emergency will not defer.** `evaluateHotfix` executed over nine
  requests: refused by a blocking authoring finding, a missing executive
  signoff, a correction note under sixteen characters, a missing rights bundle
  or a missing tone review — and **allowed** with no evidence pack, no satisfied
  Isis gate and localization at 0.10, returning a five-step application order.
- **Also executed:** the asset validator's four issue kinds over eight records,
  the bulk-upload manifest's five refusals, the EXIF policy (4 retained keys, 16
  discarded, unknown namespaces discarded), the gallery filter's every-of tag
  facet and persona-null exclusion, the compare grid's layout, mixed-kind and
  16-slot rules with the narration↔music exception, bulk-action tier
  authorization across four tiers with a 300-second undo window, signoff
  refusing a whitespace signer, notification fan-out de-duplicating three
  mentions and two reviewers into three recipients, and template payload
  validation separating missing from invalid.

### What this chapter's evidence tops out at

Every module here is pure: it returns a verdict and performs nothing, and
timestamps are arguments rather than clocks. So a passing verdict is a statement
about the inputs it was handed, never that a publication occurred. The capture
is the shipped component over its own fixture at a pinned commit — the room
presentation data is fixture-backed at this pin by design — so it is evidence
about what the surface promises and refuses, not that any engine produced
anything.

## Checks

| Check         | Result                                                                |
| ------------- | --------------------------------------------------------------------- |
| Build         | 1,695 slides, 714 sources; freshness check clean                       |
| Python suite  | 490 tests pass                                                         |
| Node suite    | 508 tests pass                                                         |
| Browser sweep | 15 slides × 1440, 1280 with audio chrome, 390 and print — 0 findings   |
| PDF           | 15 pages, exact authored visual text; the chapter has no table         |
| Capture       | 3 characterisations, 2 images, provenance recorded                     |
| Narration     | 15 tracks, 28.4 min; library-wide `--check` verifies 1,695 tracks      |
| Bindings      | 0 reviewed units cite this chapter; 0 stale across the library         |

One sweep fix: `atelier-studio-worked` ran 6 px over at 1280 with audio chrome
as a four-step case study and became a three-step one.

One fix the sweep could not see. Read against the rendered slide, all four
capture callout pins sat on the content their legends name — one over the
"Draft unavailable." heading, one over a room numeral, one over a `SPEC` label,
one over the first house rule. Overflow checks do not catch an overlay covering
text, so the four coordinates were re-measured against empty regions of the
capture and the slide re-swept. The deck, the PDF and the reader both before and
after are otherwise identical.

## Tests updated

Two, each with the reason beside it:

- The tracked-chapter count in `test_build.py` went 1,285 → 1,300.
- The rollback guard gained `atelier-studio` in `REWRITTEN_CHAPTER_GUIDES`.

One catalog change the tests required: `atelier-studio` now declares `isis` as
its prerequisite, because the reading-route test requires every chapter after
the first in a track to depend on something earlier in that track.

## Treatment deviations

- **`creative-media-classes` and `atelier-studio-failure`: table →
  `compare-panel`.** Neither needed exact values side by side; the executed
  numbers live in the notes.
- **`output-lineage`: flow → `step-journey`.** It is a record gaining
  obligations, and each step leaves one.
- **`asset-library`: prose columns → `decision-tree`.** The replacement
  classifier is a decision with four named outcomes.
- **`editorial-lifecycle`: flow → `state-machine`.**
- **`collaboration-and-templates`: prose columns → `stat-panel`.**
- **`atelier-house-rules`: flow → `capture-callouts`**, over the new harness.
- **`atelier-studio-worked`: flow → `case-study`**, of three steps rather than
  four, for the overflow reason above.

## What this chapter does not establish

No generation was produced, no provider was called, no asset was stored and
nothing was published. The one live seam in the Atelier is the image cast, which
returns `provider_not_configured` without a credential, and this chapter did not
call it.
