Binding real canon

The shipped example is a generic invention, and the honest-status table says so: binding real canon is a deliberate, human decision. This page walks that decision once, end to end, so the first real character you bind is a workflow and not an archaeology dig through the example.

The input is whatever your studio already has: a reference sheet, a costume plate, a lore paragraph with the character’s non-negotiables. The output is a faction + character contract pair the gate can enforce on pixels.

1. Split the sheet into claims

Read the reference material and write down every visual fact a reviewer would reject the art for missing. One fact per line. Concrete nouns, no vibes:

• grey-ash tabard worn over the torso
• white triple-bar sigil on the tabard
• ash-grey / bone-white / blood-red palette
• no gold trim anywhere (house rule)
• bone-spike bracer on the right forearm

If a line cannot be checked by looking at one picture, it is not a claim — it is lore. Lore stays in your canon docs; only depictable facts become atoms.

2. Decide what is faction and what is character

Anything true of every member goes in the faction contract; the character contract extends it and adds the specifics. Inheritance is fail-closed: a character may raise a faction requirement, never relax or silently drop one — so put the non-negotiables at the faction level and let characters only tighten.

From the list above: the palette and the no-gold rule are faction facts. The tabard, the sigil and the bracer belong to the character.

3. Give each claim its checker

Every atom carries a check_type — which gate tier verifies it:

the claim is about	check_type	why
a colour scheme	palette	deterministic, no model, cheapest of all
a thing being present	siglip2	cheap presence screen, good first pass
composition, wearing, relations	vqa	the compositional tier; costs more, sees more

One install note before you rely on the model tiers: palette works from a base install, but siglip2 and vqa need packages no pip extra declares — t2v-metrics for the VQA family, ai-eyes-mcp for the screen. pcraft doctor reports both under “model tier”, and a gate missing them says SKIPPED rather than silently passing — so wire the atoms now and trust the census to tell you when they can actually score.

Severity is the andon cord: required atoms block the bind, optional atoms warn. Give depends_on edges where a claim is meaningless without its parent — the sigil depends on the tabard, because there is no point scoring a sigil on a garment the gate just decided is absent. (A dangling or cyclic edge is refused at pcraft validate with a named code, so wire them freely and let the loader check your work.)

must_not entries are checked as absence on the pixels — the no-gold rule goes here, not into a negative prompt.

4. Scaffold, then edit

pcraft new emits a loadable skeleton for each half of the pair, so authoring starts from a valid file instead of a copied example. Scaffold the faction, then the character that extends it, then replace the starter atoms with your claims from steps 1–3.

The shipped reference pair (faction:ashen-pact-cloud + char:ashen-reaver-cloud) is the worked answer sheet: keep it open beside your own pair. Note it really is a pair — an identity plate’s method is inherited like everything else, so the faction half carries the choices every member shares.

5. Validate, gate, bind

pcraft validate                       # resolve + compile the DAG; refusals name the line
pcraft gate render-01.png             # score one candidate against the contract
pcraft bind                           # run the loop; a bound receipt is your provenance

Read the exit code, not the prose: 2 means a required atom failed, 3 means the human band, 4 means the gate could not run — and could-not-check is never checked-clean.

6. Calibrate before you trust the bands

The shipped threshold table says of itself that it is a generic seed. Before real canon binds at volume, label a holdout (~50–100 sprites per check type: present / absent / borderline per atom) and run the calibration workflow:

pcraft calibrate holdout.jsonl --out sprite.cal.v2.json
pcraft regrade --table sprite.cal.v2.json --records-dir records/

calibrate fits bands from your labels and emits a new table — it never silently adopts one. regrade answers the question that matters before adopting it: what would this table have decided about everything already bound? Receipts stamp both the table version and a hash of its values, so a retune can never silently rewrite history.

What this page deliberately does not cover

Wiring the identity sub-gate (it stays unwired until a holdout justifies its thresholds), prompt-engineering the synthesizer (the contract is the prompt’s source of truth), and generating the pictures themselves — this page is about making the checking real, because that is the half a pipeline forgets.