Blocks

You are given one region of an accepted layout and you return 3–4 block anatomies, each a JSON structure the owner can choose between — or one refusal naming the product decision that is missing. Like the layout stage this is not a compiler, and for the same reason: which anatomy is right is a product decision, and a single answer would pretend it had already been made.

Law

An anatomy names the parts, how many times the block rests, every state it draws, and who owns its data. It never names a class.

A state the region can enter and the anatomy does not draw is a defect, not a detail for later. The states are enumerated before anything is designed, because an anatomy built for the populated case has to be redesigned rather than extended when the empty case arrives.

Inputs

Seven, and no more.

Input 5 exists because the contract cannot answer it. optional: true states presence and nothing more: pending, failed and empty all reach the same flag. Separating them is read from the page and block source, never assumed from the registry.

Not read at this stage: class arrays, theme variables, locale copy, lints.

Reading an accepted region

1. Require an accepted layout hash. A proposed layout is not a starting point.
2. Enumerate the states first. Populated, empty, pending, failed, partial, forbidden — which can this region actually enter? Read it from source; do not infer it from optional.
3. Look up the parts by why, not by shape, and check every leaf and composite name against the vocabulary.
4. Choose the axes the anatomies will differ on. Identical axis sets are one anatomy.
5. Decide who owns the data — the block fetching it and the parent passing it in are different products, not two spellings of one.
6. Refuse rather than invent. Whether an empty region is a real outcome is the owner’s call, not a default.

Anatomy axes

repeats without a restingCount is refused: a repeating block with no resting count has no shape to review.

Per-part verdict

Every part resolves to exactly one of three, against the contract and the vocabulary:

A citation that cannot be checked against the vocabulary is an invented name, whatever its verdict says.

Anatomy laws

Fourteen laws every anatomy must satisfy. An anatomy that breaks one is not a weaker option — it is not an anatomy.

Rules

1. An anatomy carries no class, no token, no colour.
2. The state set is enumerated before the anatomy is designed, and the anatomy draws all of it.
3. repeats carries a restingCount.
4. Every part cites a name that exists, or declares a new one with its why.
5. No two anatomies in a batch share the whole axis set.
6. 3–4 anatomies when the region admits more than one; fewer with the reason stated, never padded.
7. A missing product decision is returned to the owner.
8. The JSON is canonical, and its hash is what approval attaches to.
9. Feedback opens a new round; an accepted anatomy is never edited in place.

Refusal

Refusal is an output. It is used when:

• the request does not say whether an empty region is a real outcome;
• who owns the data cannot be determined from the region or the source;
• the resting count is unstated and the region gives no basis for one;
• a part requires a component that does not exist, which makes it a contract or component change rather than an anatomy choice.

refusal: returned-to-owner
missing: <the decision nobody has made>
blocked: <which parts cannot be resolved without it>

Output

The output is JSON, and its authority is schema.json beside this record. envelope holds what varies between runs — including the accepted layoutHash this region came from — and the hash covers an anatomy only.

{
  "schema": 1,
  "envelope": {
    "session": "coding-drill-result/2026-08-18",
    "round": 1,
    "project": "starci-academy",
    "region": "criteria",
    "layoutHash": "f5534ef5e7fbe30c385108fb95702a64ac66d905414e0f7105873d67822be54c"
  },
  "anatomies": [
    {
      "id": "a",
      "axes": {"dataOwner": "parent", "repetition": "repeats", "weight": "populated", "composition": "label-value"},
      "citesPrecedent": "none",
      "states": ["populated", "empty", "pending"],
      "restingCount": 4,
      "parts": [
        {
          "name": "criterion-row",
          "cites": {"kind": "entry", "verdict": "generalize", "from": "flashcard-result-fact-row", "to": "fact-row", "callSites": 1},
          "whyMatch": "a name read against one stored value, repeated as rows on a shared baseline"
        }
      ],
      "reason": "why this anatomy is worth the owner's attention"
    }
  ]
}

Validate before writing and before hashing:

node <trust>/scripts/validate-artifact.mjs --schema <trust>/brainstorms/blocks/schema.json --data <batch.json> --hash

Beyond the shape, the validator refuses a class token anywhere in the batch, two anatomies sharing an axis set, a repeats anatomy with no restingCount, and a batch where no anatomy cites none.

Worked example

Region. “The criteria list on the coding drill result page: each criterion with its score.”

The states are read from source first: the list is populated when the drill scored criteria, empty when it scored none, and pending while the result loads. Three states, so every anatomy draws three.

Anatomy a reuses the accepted shape with the entry generalised — one call site, so the rename is one edit. Anatomy b keeps the same shape but treats the empty state as the one carrying the region, because a drill that scored nothing is a real outcome and a silent list reads as a load that never finished. Anatomy c is the departure: no repetition at all, the criteria as one grouped visual with a caption, citing profile-breakdown.

Because c cites none, the batch has its departure. Without it, a and b differ on one axis and the owner is shown one anatomy twice.

Scope

This stage decides what a block is made of and which states it owes. It does not decide where the region sits — that was the layout — and it does not decide a class, which is the law’s business. The fourteen laws an anatomy must satisfy are stated above as BLOCK-n codes, so an anatomy is checked against a citable code rather than against a reader’s memory of the legacy tree.