Grounded Production

Every new Wheel of Heaven artifact — an Article, a wiki entry, a timeline chapter, an illustration brief, an audio-play scene — derives its load-bearing assertions from claim records in the research core before the prose or the brief is written. The artifact is a bounded public projection of a versioned record, not the place where the insight is first invented.

This is the artifact derivation contract, proposed in core RFC 0003 and accepted on 2026-08-15 as ADR 0002. It binds new artifacts and fundamental rewrites. The existing corpus was grandfathered at adoption — and has since been backfilled: ten distillation batches (2026-08-15/16) extracted the corpus’s claim layer into a 69-record catalog (accepted by core ADR 0005 and ADR 0006), and every claim-bearing page now declares its binding. Grandfathering remains relevant only for pages that carry no claim (direct reference, list, and biographical entries).

Why the contract exists#

The project used to run on two constitutions that did not talk to each other.

The publication side — the editorial standards in this section — governs how things are said: voice, terminology, structure, the sourcing floor. The core governs what may be claimed and on what basis: the five orthogonal axes, evidence maps, alternatives, revision triggers.

Before the contract, only one page in the entire corpus was bound to a claim record. In practice that meant an Explainer’s central insight was invented at writing time, disciplined by voice rules alone, and its claim_type badge was a per-page, per-author judgment call with no controlling record behind it.

The contract closes that gap by inverting the order of work. The insight lives in the record; the artifact renders it. The practical payoff is that founder review becomes reusable: instead of reviewing the argument of every article and each of its nine translations, you review the claim record once, and every artifact that later cites it inherits that review.

The test#

An artifact’s insight must be quotable from a claim record, not only from the artifact.

If a load-bearing sentence cannot be traced to a record’s proposition, scope, and public label, either the record is missing (go ground it) or the prose has drifted beyond what the project has actually committed to.

The pipeline#

flowchart LR
    scope["Scope<br/><i>name load-bearing<br/>assertions</i>"]
    ground["Ground<br/><i>resolve or draft<br/>claim records</i>"]
    gate{"new record?"}
    founder["Founder review<br/><i>pause</i>"]
    draft["Draft<br/><i>write prose</i>"]
    edit["Edit"]
    review["Review"]
    bind["Bind<br/><i>declare + reciprocate</i>"]
    publish["Publish pipeline"]

    scope --> ground --> gate
    gate -- yes --> founder --> draft
    gate -- no --> draft
    draft --> edit --> review --> bind --> publish

1. Scope#

Name the artifact’s load-bearing assertions: the claims the artifact would be wrong about if they were wrong. Three to seven, each phrased as a proposition. Incidental background does not count — an Explainer that mentions the Bronze Age collapse in passing is not making a claim about it.

2. Ground#

Resolve each assertion against model/catalog.json in the core.

Hit — read the claim record, its controlling specification in docs/framework/, and its evidence map in docs/evidence/. Note the version. The artifact must stay inside the record’s scope and honour its public_label.

Miss — draft the record before writing prose: the claim record, its controlling specification, a scoped evidence map, source notes for any newly referenced project-held sources, and the catalog entry. Everything is lifecycle: draft. Then stop.

The founder pause. A newly drafted claim record pauses the pipeline. No prose is written against an unreviewed record. This is not a formality — it is where the contract earns its value, because reviewing the record once is what saves reviewing the same argument in every artifact and language that later expresses it.

Agents draft. Only the founder promotes a lifecycle, adds a foundational claim, or identifies referents across traditions.

Ground closes with a grounding report: claims matched (with versions), claims drafted (awaiting review), and assertions demoted to incidental.

3. Draft#

Write under the editorial rules for the target content type — the per-type guides in this section, plus Editorial Passes for claim_type and editorial_pass.

What the contract adds: hedging is set by the record, not by instinct. Canon claims are stated directly; comparative, scientific, and critical layers stay hedged as the bound record specifies. Where a record’s scope.excludes rules out territory — transmission routes, one-to-one identifications, historicity — the prose must not gesture at it either. Where the form has a counter-case slot, the record’s alternatives belong in it.

4. Edit#

Adjudicate the artifact against each bound record clause by clause (scope respected, label honoured, alternatives represented, excluded territory untouched), then against the editorial standards.

When artifact and record disagree, the record controls. Either fix the prose or propose a record revision back at stage 2 — never silently reinterpret the claim in prose.

5. Review#

Fresh eyes, working from the artifact and the records only — not from the writer’s or editor’s reasoning. Re-derive the binding: would a reader reconstruct this claim, at its label and scope, from the artifact alone? Verify citations at their recorded access levels. Attempt the strongest refutation. Return a verdict: pass, mechanical fixes, or substantive issues for human judgment.

6. Bind#

Declare the binding in the artifact’s [extra]:

core_claim_ids = ["woh-claim-0004"]
core_versions = { woh-claim-0004 = "0.1.0" }

Then reciprocate: add the artifact to each record’s public_derivatives, and set claim_type to match the record’s public_label. Re-run the core validator with the sibling repos present; it must report publication_integration=checked with no errors.

Field details: Frontmatter → Core claim binding.

7. Publish#

Hand off to the normal publish path for the content type — Articles to the article publish pipeline, timeline chapters to fan-out, everything else to its section flow. Translations inherit their source page’s binding and need no grounding of their own.

Who does what#

The role charters live in the core at docs/methodology/production-roles.md (accepted, version 0.1.0) and are deliberately harness-agnostic: Claude Code, Codex, and human contributors all follow the same charter, and the orchestration wiring points back at it rather than restating it.

Two rules cut across all five. Stage boundaries are real — a writer who discovers a missing claim does not mint one inline; the work returns to the Researcher. And founder gates are not pipeline stages — no downstream stage starts against unreviewed output.

Running it#

In Claude Code, the woh-produce skill wires the charters to the harness and runs the stages in order, including the founder pause. It fronts the type-specific publish pipelines rather than replacing them.

Other agent harnesses perform the same stages manually, reading the charters directly from the core. The stage outputs — grounding report, editorial report, verdict — are plain markdown handoffs; preserve them and the founder gates exactly. The contract is satisfiable entirely by hand.

Scope of adoption#

The backfill of grandfathered pages completed in 2026-08: ten distillation batches applied the reverse-pilot pattern below at corpus scale — extract the record from the artifact, then bind — leaving unbound only pages with no claim to carry.

Worked example: the reverse pilot#

The contract was validated backwards before it was adopted forwards.

The Explainer The World Behind the Odyssey was written the old way — prose-first — and its central reading was then extracted into woh-claim-0004, “Greek theomachy as Council–Serpentine conflict memory” (interpretation / comparative / speculative, depending on claims 0001 and 0003). The extraction produced the full apparatus: a claim record, a controlling specification, an evidence map whose matrix names the relationship each comparison actually supports, and a source note for the in-house Theogony translation. The article now declares the binding.

The exercise was worth running for what it exposed — all of which grounding-first would have caught at stage 2:

• An unrecorded dependency. The reading’s real foundation — the Council–Serpentine conflict account — turned out to exist only as corpus prose, not as a claim record. It is now a documented gap with a revision trigger.
• A label with no controlling record. The article’s speculative badge was a sound editorial call, but nothing outside the article backed it.
• Falsifiability terms trapped in prose. The conditions under which the reading should be surrendered lived in one article’s closing section; they are now revision_triggers on the record, where every future artifact inherits them.

The visual layer#

Illustrations, hero images, and audio-play scenery ground the same way, but against depiction notes rather than claim records — a separate document class accepted as ADR 0003. An image cannot hedge, so the record hedges for it: each note separates what sources visually report from what the project interpolates, what is free for art direction, and what must not be shown.

The class is accepted; its pilot notes and validator support are still being implemented. See Research Core → Depiction notes.