Publication record · Structure · July 2026
Rival answers—called landings in the working process—are developed and pressured before prose begins. Sources are checked as claims are written. Changes route backward when they alter meaning. The human author makes the decisions and owns the result.
Current operating map Reviewed against the operational state dated July 20, 2026. Current required steps and active pilots are labeled below.
The editorial path
The stage codes preserve the real operating order. The phase names make that order readable without pretending each pass is an independent inspector.
Default order 0 → 1 → 1C → 1D → 1B → 1A → 2 → 2A → 5 → 6 → 7 → 7A → 8A → 8B → 8 → 9 → 10 → optional 10A/10B → 11 → 12 → 13 → 14
“Outside” means outside the internal run, not outside AI. The author manually runs Steps 1A, 2A, 8A, 8B, and 9 in a separate frontier-model interface, then decides what to accept. These are not human expert reviews and do not guarantee independent cognition. Optional branches include author-directed Council input (4B), a bounded perspective scan (5A), public records (10A/10B), and derivatives (15); none replaces a required gate.
For a meaningfully contested question, the process adds a sealed, context-free default-prior probe and a landing-grade research dossier. It normally develops three rival answers, with two to four allowed.
The record must say why the lite lane applies. It uses two rival landings and omits the probe and full contested dossier.
The AI editorial process can research, draft, critique, edit, and prepare packages. The human author chooses the question and landing, controls protected-item removal and final wording, adjudicates outside input, and alone publishes or authorizes public actions. The Referee labels risk and recommends; it does not delete claims. Tool access never expands that authority.
What the system is for
The point is to keep prose momentum, confidence, and process display from outrunning the record.
Open a phase
Each record separates the work, required trace, constraint, decision right, failure route, and remaining risk. The exact stage codes match the current operating map.
Recover the latest completed artifact and next gate. Check the weekly problem brief or record why it is not used, recurring claim constraints, the issue’s increment over prior work, and at least one plausible outside anchor. Before Builder, record what a first-time reader must be able to place, what may feel missing or socially costly, and which internal terms need translation.
A question brief, lane record, declared starting belief and how it might steer the work, primary audience, landing-neutral reader success test, cross-issue increment, and candidate outside anchor. A contested issue also requires a sealed probe.
No landing, arc, or draft is chosen here. The reader success test must fit every possible answer, and the probe may not steer or validate the issue.
The human author selects the question. The lane and reason are recorded. The AI process may surface options and risks but cannot convert a candidate answer into the assignment.
No real increment or plausible anchor means recast or hold. A material question reframe must be recorded and may require a fresh probe before research continues.
State recovery and probe sealing are manual. A stale artifact can be resumed, and a sealed file can be exposed if containment fails.
Before any arc, record the human stakes and author-protected material that must remain legible when needed. Steelman major live positions on prior-triggering questions, prepare the contested-issue dossier, then build rival answers and run six questions about what each is willing to be wrong about.
Two to four genuine candidates, each with its sharpest defensible thesis, load-bearing premises, burden of proof, kill evidence, concessions, and protected essentials. Record the selected answer and direction, findings, dissent, decisive evidence or missing evidence, probe match when applicable, word counts, and who coded the result.
Candidate premises must be vetted before scoring. No sacrificial alternatives, no bland “survival” through empty hedging, and no symmetric or unsettled answer without either verifiable parity evidence or the specific missing evidence that would decide it.
Findings report; they do not rank. The human author selects the landing and may choose one outside the slate if that authored decision is recorded. Only after selection is a contested probe unsealed and compared.
Weak steelmanning, unvetted premises, reversible burdens, or an answer that survives only by saying little route to more research, a new slate, or a hold. A later direction change returns to selection.
The same model can generate and criticize the rivals, and the process can still manufacture apparent competition. The landing materials aim for about 15,000 words total and 4,000 for the selector; the journal remains a hygiene record, not a statistical detector.
Build a plain-language narrative arc from the selected landing and submit it to a manually run outside critique. Builder then drafts from the approved arc, current style rules, and original sources, followed by a mandatory outside full-draft adversarial review.
The selected-landing arc, outside arc return and disposition, Builder draft, live edit log, source package, seven-field source checks, citation-trap log, and outside full-draft return with adjudication.
For every source-citing sentence, open the source itself and verify figure, unit and denominator, design, population and setting, date or version, claim type, and magnitude language. The dossier is not the source.
The AI process stages the outside prompts. The human author runs them manually and treats the returns as input. A material thesis or architecture change requires an authored selection before work proceeds.
An arc that changes direction returns to landing selection. A source correction is a Builder repair unless it changes what the claim can support; then it routes to Referee.
The outside run can share the Builder’s underlying model and assumptions. Paywalls, version drift, wrong-but-prominent search results, and incomplete source access can still defeat the check.
The Core Adversarial Battery attacks assumptions, mechanism, evidence, alternatives, and scope. Referee turns the findings into claim-level constraints. Editor then improves readability, and Reference Link settles the outside support.
An inspectable Battery memo, any recurring-claim recommendation, Referee labels and constraints, authored departures from warnings, meaningful compression notes, and a Reference Link record.
Attackability is not invalidity. Editor may not upgrade confidence, strip a hedge, weaken a warrant, erase a caveat, homogenize voice, or polish away useful anchored friction.
Referee may label a claim keep, qualify, downgrade, fact-check, or overreach, and may recommend removal or no-publish/rework. The human author decides retention, removal, wording, and publication.
If readable language cannot preserve a Referee constraint, return to Referee. Step 7A re-runs Source Fidelity only on source-citing sentences added or changed since Builder.
Internal roles can share blind spots. A conformance check cannot detect a defect in the artifact it treats as authority, so internal fidelity, primary-source checks, and outside review remain different functions.
Run the outside plain-English pass, then the outside meaning-risk and compression pass. Adjudicate both, run the issue checklist and the title, summary line, and section-header skeleton check, then run and adjudicate the final outside Narrative Integrity Review.
The 8A and 8B packets, returns, and dispositions; checklist result; protected-item status; skeleton result; final review return; and Step 10 decision on every accepted change.
No new claims, source changes, confidence upgrades, caveat loss, protected-item rewriting, or silent substantive reorder. Meaning-risk comes before compression, and useful repetition may stay.
The author manually runs the outside passes and adjudicates their returns. These are AI-model reviews, not subject-matter-expert approval or a real-audience comprehension test.
Any change that touches a claim, source, landing, caveat, protected item, or package surface returns to the affected earlier gate. Open outside review or unadjudicated changes block packaging.
A persuasive late rewrite can still smuggle in drift. AI cannot fail to understand in the way a real non-specialist reader can, so real-reader evidence remains a separate post-publication test.
When warranted, prepare a process summary, Extended Development Record, or World Behind issue slice. Then freeze the final text, title, dek, and footer; prepare the HTML, manifests, thumbnail, source package, and any approved derivative; and run package QA.
Always: frozen publication authority, package and asset manifests, verified links, crop and feed checks, and a package-QA result with no required adjudication open. When a public process record is built: its scope, redaction review, and approval.
A process summary is not an EDR. An EDR is scoped, public-safe, verbatim author/AI conversation in order. Packaging inherits verified sources and rechecks only facts that may have changed with time.
Conversation is private substrate by default. Public use requires scope selection, redaction and privacy review, relevant claim and link review, safe framing, and explicit human approval.
An asset or package element that adds a substantive claim returns to Referee. A mismatched final file, broken link, unresolved review, or unapproved public record blocks release.
Redaction, link, and manifest checks remain manual. Older records are incomplete, and a standing accessibility package check is not yet built.
The human author publishes or explicitly authorizes publication. The process verifies the public page and RSS, archives the final URL, updates project state and recurring claims when needed, records the landing journal, then prepares approved distribution and reader follow-up.
Public URL and feed verification, archived finals, updated state, any needed claim update, the landing record, approved distribution copy, any reader disagreement or correction signals received, and any post-publication tightening log.
Publish, schedule, send, and automation activation remain human-final. A post-publication edit is a new event: extend the record, preserve earlier turns, and keep the canonical slug unless the author decides otherwise.
The author approves public wording and actions. Reader signals can trigger correction, clarification, or later work; they do not become automatic editorial authority.
A mismatch between local and public text, title, image, audio, or feed is recorded and repaired as a public change. Optional audio or video stays subordinate to the written issue and gets its own QA trace.
Public-state verification and distribution are manual. Reader response may be sparse or unrepresentative, and a correction record cannot undo the reach of an earlier error.
Current status
A current process page should not turn young experiments into permanent machinery or describe adoption as proof that a check works.
During Issue 21 setup, the contested-lane probe was exposed to research agents and therefore voided. The issue is running without it, with a narrower comparison against the author’s declared starting belief. That exception is recorded; it does not remove the standing probe rule. The process now combines an explicit prohibition on opening sealed probe files with keeping them unavailable while research agents run.
Not standing process: proposed gate mergers and register-first refactors; conformance to the full plain-science guide; controlled-English drafting; AI-simulated expert or audience approval; and a complete accessibility packaging check. These are rejected, deferred, or still open—not current credentials.
The process does not award a “survived review” badge. It names where the work has to move when a condition remains:
Researcher, Builder, Battery, Referee, and Editor are roles inside an AI-assisted system. Role boundaries and separate prompts can create useful friction; they do not create independent cognition.
The five manually run outside passes inspect artifacts from outside the internal sequence, but they are still AI-model reviews. Signal & Noise does not currently claim human subject-matter-expert approval. The post-publication reader test is the only current step designed to collect direct audience-comprehension evidence, and it is still a pilot.
That is why the process keeps three functions separate: opening the primary source, checking conformance to an internal record, and asking a separately run model to challenge the artifact. Each can see failures the others cannot; none proves the work true.
Practical implication: more weight belongs on source contact, explicit decisions, route-backs, public correction, and reader challenge—not on the number of passes.