Publication record · Structure · August 2026

How an issue moves from draft to published audit.

The author writes the essay. Before it publishes, a binding accuracy-and-fairness floor is checked. After it publishes, an audit runs — and its record publishes too. The human author makes the decisions and owns the result.

Current operating map Reviewed against the operational state dated August 8, 2026, when the prior process was retired and this one adopted. The prior map, and its preserved failure record, are kept as history below.

The editorial path

Three stages, one binding floor, one published audit.

The essay is the author’s argument, made the way the author wanted to make it. The checks do not certify it — they publish what checking found, so a reader can weigh the argument against its record.

A
The essay Written by the author; AI drafts and critiques
B
The floor pass Binding accuracy-and-fairness rows, one redline
C
The audit Autonomous, published, no edit authority

Default order Essay → floor pass → the author publishes → the audit runs and publishes its record. An issue’s audit is live before the next issue ships.

“Autonomous” means the audit runs and publishes without the author’s touch — not that it has authority over the essay. The audit proposes; the author decides. Its critique runs on at least two independent frontier models, and every audit page states exactly which models ran it, how, and what independence discount that implies. A piece that audits someone else’s claims can carry a high-stakes designation, which moves the audit — or a named part of it — before publication.

The floor — before publication

No false or fraudulent claims, operationalized

Every issue is checked against binding rows before it publishes: exact figures; verbatim quotes verified against primary sources; fairness to named people; receipt-or-cut for unquantified-frequency and unevidenced negative-existence claims; calibration that respects statistical qualifiers; opinion identified wherever a reader arrives; coherence, with corrections that extend the record rather than rewrite it; no empty hedging; honest disclosure of AI’s real role; and at least one real-world anchor.

A failing row blocks publication.
The author picks the remedy — receipt or cut — never whether to remedy.
An override is possible, but only explicitly, with ownership of the risk, on the record.

The audit — after publication

Three lists, never a verdict

The audit page reports objections found and what happened to each, changes proposed or made, and what could not be checked. Blocked is “null,” never zero. There are no survival counts, no badges, and no “passed review” language — the essay’s own argument for why.

Findings are produced with no edit authority, then confirmed or refuted in fresh context; only confirmed findings become proposed changes.
Machine-checkable things first — quotes, figures, attributions, links, dates, units — with reproducible evidence; judgment-class critique after, labeled as judgment.
The audit page is checked against its own raw returns before it publishes, because the first one misreported itself in ways that flattered the audit.

The human-final boundary

The AI process can draft, critique, check, stage, and publish audit records. The human author makes the essay’s decisions: what publishes, final wording, whether a floor failure is overridden on the record, and every public act — publish, send, upload, schedule. A floor-class defect found after publication stages a correction for the author; the audit never edits a published essay. Tool access never expands that authority.

What the system is for

Constraints, protections, and limits.

The point is to keep prose momentum, confidence, and process display from outrunning the record.

It constrains

What may happen, and when

Floor rows are binding; a failure blocks publication absent an explicit, recorded override.
No pass in the process may upgrade certainty or strip a load-bearing hedge.
Corrections extend the record; nothing is silently rewritten.
Tracked claims carry next-check dates; a lapsed claim may not be cited in new prose until its check runs.
An issue’s audit is live before the next issue ships.
Publish, send, upload, and schedule are the author’s acts, always.

It tries to prevent

Predictable forms of drift

Staged critique being mistaken for independent scrutiny.
“Survived review” standing in for the specifics of what was checked.
Confidence language outrunning the evidence behind it.
A published essay changing quietly after the fact.
The process display itself becoming a certification badge.

It cannot guarantee

Truth, independence, or completion

A shared model blind spot can survive every check.
A cited source can still be wrong, incomplete, or inaccessible.
A recorded trace is only as complete as what was captured.
No reader-comprehension instrument exists — a planned real-reader test was dropped on 2026-07-25 without ever running.
Human cold-read defects — register, cognitive load, surface drift — are a named standing blind class.

Open a stage

What each stage must produce—and where it can stop the work.

Each record separates the work, required trace, constraint, decision right, failure route, and remaining risk.

A The essayThe author’s argument, developed directly with a frontier model — deliberately unprocessed.

Work

The author develops the essay in their own conversations with a frontier model, varied over time. No required steps, no gates, no artifacts — the essay is the argument the author wanted to make, the way they wanted to make it.

Required trace

The final draft text. Optionally, the exported conversation, if the author wants a public origin record — it publishes only after a disclosure pass and the author’s approval of the redacted version. Optionally, a high-stakes designation, which moves the audit (or a named subset) before publication; pieces that audit someone’s claims default to it.

Constraint

None here, by design. Every check binds later — the floor before publication, the audit after. The stage is honest about what it is: the author developing an argument, with AI as the drafting and critique tool.

Decision right

The question, the argument, and the wording are the author’s.

Failure route

None; everything routes through the floor.

Remaining risk

An argument developed with one model can inherit that model’s defaults and blind spots. The floor and the audit are the counterweight, not a cure — the audit’s cross-model critique exists for exactly this reason.

B The floor passBinding accuracy-and-fairness rows checked against the finished text; one redline back to the author.

Work

The process checks every floor row against the finished essay — figures and quotes against the primary sources themselves, clause by clause, not sentence by sentence — and returns one artifact: a redline against the author’s text, floor rows only.

Required trace

The redline, and behind it a per-claim receipts record: what was checked, against which source, with what result. A load-bearing source that could not be reached is a named, author-decidable row — never a silent pass.

Constraint

Floor failures block publication. Discretionary improvement suggestions are off by default — the floor is a floor, not an editor.

Decision right

The author rules on every row — receipt or cut — and may override a row only explicitly, with ownership of the risk, on the record. The author publishes; the process never does.

Failure route

A failing row is fixed, cut, or overridden on the record. There is no fourth path.

Remaining risk

A clause can be faithful to its anchor and still carry an implication the anchor does not — that class ships and returns as audit findings and corrections. The floor’s execution mode (mechanical verification plus a single-model AI pass) is stated on the audit page.

C The auditRuns against the published essay; publishes its record to this site; proposes, never edits.

Work

Machine-checkable verification first — quotes, figures, attributions, links, dates, units and denominators — with reproducible evidence. Then judgment-class critique, labeled as judgment: a standing adversarial battery, fairness-in-context, and a steelmanned counter-reading, on at least two independent frontier models.

Required trace

Every check emits an inspectable record of what was checked, how, and what could have changed the claim. “Ran, passed” never appears. Before the audit page publishes, it is mechanically checked against its own raw returns — every objection traced to a source return, every count recounted.

Constraint

Kills, changes, and blind spots — never verdicts. No survival counts, badges, or validation language. Per-claim calibration (claim type, evidentiary confidence, what would change the conclusion) is required reporting, not a verdict. Every page carries its execution-mode line: models, transport, independence discount.

Decision right

Audit pages and the tracked-claim ledger publish under the audit’s own authority. Three things always escalate to the author: any edit to published essay text, any claim that would bind the author’s personal conduct or identity, and anything the author has flagged.

Failure route

A floor-class defect found in a published essay stages a correction under the standing corrections regime and notifies the author. A wrong staged correction still costs the author time, so the audit’s false-positive rate is tracked on its own page.

Remaining risk

The audit’s zero-hit results inherit its refute-by-default bias, and its largest gaps are named on each page — including, on Issue 23, a fully specified disconfirming experiment that was not run. Timing is a disclosed trade: judgment-class defects the old pre-publication gates would have blocked now ship first and return as corrections.

What this replaced

The prior process, retired — and its record kept.

From March to August 2026, issues ran a 21-step operating map: stage codes 0–14, rival-landing slates, sealed default-prior probes, contested and lite lanes, a Battery→Referee→Editor sequence, and five manually run outside passes. On August 8, 2026 the author retired it: its time cost had outrun its value, and its length had itself begun to read as a credential. The retirement and what survives it are recorded in the public constitution’s amendment set (B1–B8).

What survived

Carried into the current stages

The accuracy-and-fairness floor, as binding rows.
The adversarial battery’s questions, now inside the audit.
Detection separated from mutation; findings verified in fresh context.
Human-final publication and the corrections regime.
Traces for every check — “ran, passed” never appears.

What retired

Gone as standing process

Landing slates, sealed probes, contested/lite lanes.
The Battery→Referee→Editor→Reference Link required sequence.
The five manually run outside passes — replaced by API cross-model critique with a disclosed execution mode.
Pre-publication meaning-risk and plain-English gates; their defect class now ships and returns as corrections, stated as a trade.

Dropped without running

Named, not hidden

The five-question real-reader minimum test was dropped on 2026-07-25; it never ran.
No reader-comprehension instrument currently exists; reader evidence remains a post-publication signal, not a gate.
Older-issue record backfill is held, not active.

A recorded failure, not a silent repair

During Issue 21 setup, under the prior process, the contested-lane sealed probe was exposed to research agents and therefore voided. The issue ran without it, with a narrower comparison against the author’s declared starting belief. The exception was recorded rather than repaired silently. The probe instrument was retired with the process on 2026-08-08; this record stands as part of the site’s history, and the probe methods page is preserved with its supersession notice.

Conditions that stop the work.

The process does not award a “survived review” badge. It names what blocks, and what happens when something fails:

A floor row fails: publication blocks until the claim is fixed, cut, or explicitly overridden on the record.
A floor-class defect surfaces in a published essay: a correction is staged and the author notified; the record extends, the history stays visible.
A tracked claim’s scheduled check lapses: the claim may not be cited in new prose until the check runs.
The previous issue’s audit is not live: the next issue does not ship.
An exported development conversation is unreviewed: it does not publish.

Where independence exists—and where it does not.

The floor pass is mechanical verification plus a single-model AI pass — useful friction, not independent cognition. The audit’s critique runs on at least two independent frontier models via API, and every audit page states its execution mode: which models, over which transport, and what discount that implies for independence. Model diversity reduces shared blind spots; it does not eliminate them.

Signal & Noise does not claim human subject-matter-expert approval, and no reader-comprehension instrument currently exists. That is why the audit reports specifics — what was raised, what was checked against sources, logic or external evidence, what changed, and what remains unresolved — and never a verdict. A report like that can be faked too; faking specifics means publishing claims a reader can check.

Practical implication: more weight belongs on source contact, published specifics, corrections, and reader challenge—not on the number of passes.