This version: commit 20d5409, .
This is the standard I hold this publication to: the claim I want to be able to make, and the tests that decide whether I may make it. Nothing on this page says a test is met. Which tests are met, and by how much, is published in the quality report and in Methods. The page is rendered from the file the room works from, at the commit shown above, and it changes only by my ruling. —Craig
The claim being operationalized
I intend 365 Changes to be best-in-class among citation-grounded consumer health content: every published claim reader-traceable to identified, independent, quality-graded sources; the verification process’s error rate measured and published; and confirmed errors corrected on a stated clock. —Craig
Three properties make this operational: every clause is checkable by a named instrument; the claim is dated and renewable (it decays and must be re-earned each cycle); and a threshold breach suspends it until restored.
The class, and what best-in-class means
Class: citation-grounded consumer health/nutrition publications at scale (hundreds of pieces, general audience).
| Practice | Best current practitioner | Ours |
|---|---|---|
| Evidence graded per claim | Examine.com | C1 (grades) + F3 (badges) |
| Named human review | Healthline/WebMD (“medically reviewed”) | a verification report per piece plus the editor’s approval, disclosed in F1 (methods) |
| Declared systematic methodology | Cochrane plain-language summaries | F1 (methods) |
| Public corrections history | Wikipedia (revision/talk) | E3 (ledger) → F4 (errata) |
| Claim-level reader traceability | nobody at scale | A3 (mirror) + F2 (resolution) |
| Published, measured verification error rate | nobody | D1 (resample) |
Best-in-class here means parity on every practiced dimension and leadership on at least one that nobody yet practices. The last two rows are the frontier. This keeps the phrase a comparison a hostile reader can check, not a self-referential slogan.
Scope — the reader surfaces this standard binds
Companion bodies (site + email render), excerpts (listings, email, logged-out teaser), Further Reading sections, reference cards/pages, and prompts. Prompts are in scope — ruled 2026-08-02 (ratification 3): the 500 prompts enter the standard through a one-time claim audit, since each prompt chains to its response and companion.
The dimensions
Format: predicate → instrument → threshold [status as of 2026-08-02].
A (attribution)
- A1 (coverage) — every claim-bearing sentence in a published companion resolves to ≥1 RID whose chain bottoms out in held source material judged sufficient to support the claim (not a fixed holdings bar: an abstract may suffice, depending on what the abstract, the knowledge file and the claim say) →
research-verified-v2plus a complete verification report per piece +library_health(missing-report, bare-disposition checks) → 100% of 500, zero bare dispositions. - A2 (chains) — companions cite k/d only; k-files cite c only; every inline RID resolves → existing
library_healthsections → zero violations. - A3 (mirror) — repo ↔ reference-DB ↔ republished-post sync: every republished companion’s closure is in the DB, no drift →
reference_push --allidempotence + a diff mode (small tool gap) → zero drift;republished≥ v2 on 500/500.
B (identity)
- B1 (twins) — one record per work: no two c-RIDs share a normalized DOI/PMID/PMCID/ISBN → new
library_healthsection over machine-readable keys → zero after merges. - B2 (independence) — no k-file bullet’s multi-citation backing collapses to fewer distinct works; corroboration language (“multiple studies,” “converging evidence”) requires ≥2 distinct works → new check layered on B1 → zero.
- B3 (keys) — every citation carries a canonical identifier (precedence DOI > PMID > ISBN > normalized URL); webpage sources carry an archived copy (link-rot is the webpage’s retraction) → schema heading +
library_health→ 100%.
C (grading)
- C1 (grades) — every citation carries an evidence-quality record: study design (RCT / meta-analysis / observational / mechanistic / animal / review / book / webpage — finer than
Type) and retraction/contested status, with an automated retraction sweep of all PMIDs/DOIs (PubMed retraction notices + Crossref/Retraction Watch) → new required heading + quarterly sweep → 100% coverage; zero retracted sources cited anywhere. - C2 (tiers) — no HIGH-severity directive rests solely on mechanistic/animal/single-small-study evidence without in-text disclosure → severity tags (fc-reports already carry the concept) × C1 tiers → zero undisclosed mismatches.
- C3 (theses) — every k-file backing ≥10 companions carries a recorded epistemic-status review (settled / converging / contested), explicitly covering the inherited Hungry Brain vs. Obesity Code paradigm tension → a reference doc + per-k-file status → complete for the top tier, then maintained.
D (reliability) — the frontier dimension
- D1 (resample) — blind re-verification: per cycle, sample 20 companions (~600 claims; 95% CI ≈ ±1.6 points near a 5% rate), a fresh verifier, no access to the prior report, diff dispositions. “Material disagreement” = a disposition flip (backed↔cut) or both passes finding the same claim unsupportable → ratified 2026-08-02: ≤2% material on load-bearing claims, ≤5% overall; first run sets the baseline, thresholds re-affirmed against it. Measured 2026-08-09 — run one, the baseline (20 companions, 256 assayable claims, the blind pass attested): the ratified material-disagreement rate is not yet computable and is published as such, not substituted — it diffs against prior per-claim records, and v2 coverage at run time was 32/500 (6.4%), one eligible companion in the draw; it becomes computable as the re-verification wave lands. Cycle-2 planning re-decides the stratified verifier-draw question against that day’s coverage number (deferred until that day) — an interim instrument priced today would be priced against a stale picture. What run one baselines: assayed-false 2.3% ±1.9 (nearest in kind to the ≤5% overall threshold, inside it); load-bearing assayed-false 6.6% ±6.2 — outside the ≤2% threshold in kind, stated head-on; run one cannot fail and cycle two reads against this, at a restated coverage figure; overall defect (false + unreproducible) 53.5% — of 256 assayable claims, 6 are wrong and 131 are unsourced; reader-traceability gap 40% of true claims unlinked; chain integrity clean, zero unresolved RIDs. The instrument this standard called its most load-bearing now exists and has run.
- D2 (escape) — verification-class errors found after the
proofreadstamp (incidents, sampling, reader reports), per companion per quarter → the incident record → provisional: ≤2%, trending down. The proofreading desk’s catches are not escapes — that’s gate two working; what passes both gates is the product’s error rate. - D3 (conversion) — every veracity incident closes with a new automated check or a recorded decision not to build one → the room’s record → 100%.
E (correction)
- E0 (the reversal protocol — ruled into the standard 2026-08-09, from a worked instance): a claim’s withdrawal is not complete until the corpus has been swept for its live restatements, and the sweep is part of the act of reversing — performed by whoever executes the reversal, recorded in the d-file itself: scope searched, surfaces found, disposition of each (fixed / hedged-already-sound / deliberately-left, e.g. historical exports). The failure this closes was never detection or discipline — one reversal was executed perfectly on the companion in front of the decider and nowhere else, because nobody looked past the file in hand; the withdrawn claim then sat five days in a prompt, the most-read surface, found by accident. The sweep keys on the claim, never on retired wording alone — a wording-keyed check convicts lookalikes (the recorded false positive: a glycemic “false hunger signals” that is a different claim). Companions fixed under a sweep get the killed claim removed, not a re-verification they have not earned. A withdrawn claim is never queue material: drift can wait for a wave; a claim the human weighed and killed cannot.
- E0.1 (the two-hands protocol — entered 2026-09-03, after its second run in one evening): a cross-layer repair goes wrong → safe → right, across two desks, and the middle state is a lawful place to stand. When a reader-facing surface carries a claim whose truth lives in another layer (a k-file, a source record, a citation), the desk at the page often cannot make the page right — only safe: hedged or de-asserted so it misleads nobody while the layer that owns the truth is repaired by the desk that owns it. The safe state is explicit, journaled, and held deliberately — never silent (the forbidden middle is silent-where-it-was-wrong, already in the taxonomy); the repair completes only when the owning desk lands the truth and the page is re-pointed at it, and the chain rule then carries the page through re-proof and republish. Both runs held cleanly: wrong at the page, safe at the page within the hour, right only after the owning desk’s layer landed. This is how cross-layer repairs go here — protocol, not improvisation.
- E1 (latency) — error-confirmed → live: corpus-layer fixes (k/c records via the backbone API, no republish) ≤48h; body fixes (republish path) ≤7 days — ratified 2026-08-02 → timestamps in journals/logs +
republishedstamps. - E2 (staleness) — no companion stays published on stale verification (k-file modified after its v2) beyond one batch cycle (14 days); zero >30 days →
ReverifyStalenesscounts + ages. - E3 (ledger) — complete internal corrections record (feeds F4 (errata)) → journals/board discipline → every confirmed error logged.
F (transparency)
- F1 (methods) — a published methodology page/series describing the RID system, verification, grading, corrections policy — with the anti-trust-washing clause: zero claims about process the system doesn’t actually perform, audited against practice at every renewal.
- F2 (resolution) — zero unknown-RID renders in production (the reference-backbone plugin logs them), zero dead
/reference/{rid}/pages → plugin logs + crawl → zero. - F3 (badges) — evidence grades reader-visible on reference cards/pages once C1 (grades) exists → plugin render → 100% of reference pages.
- F4 (errata) — the public corrections log, E3 (ledger) made visible → live and current.
The Methods series amendment (ratified 2026-08-02): F1 and F4 move from phase 3 (frontier) to phase 0 as evolving surfaces — the public series starts with the editor’s personal statement and carries the plan, progress, and metrics as they happen. Series invariants: every published number traces to an instrument run; every process claim is currently-true or explicitly marked planned; posts are dated and never silently edited (a correction to a post is a visible correction note); each post is a projection of a repo-canonical artifact, redacted only for secrets/credentials/infrastructure. The repo remains the system of record; the series is a render.
G (process)
- G1 (guides) — guides and laws pass
guide_health: referenced paths exist, quoted commands parse, named fields matchFrontmatterSchema, duplicated law is byte-identical → zero findings. - G2 (frontmatter) — tool-mediated everywhere → a validation sweep across all types → zero.
- G3 (tests) — suites green; no test touches the real corpus, journals, or board (the FrontmatterToolTest landmine class) → zero.
- G4 (health) —
library_healthfully clean at steady state, its section count growing only via D3 (conversion) → zero across all sections.
Verdict semantics — when the claim may be made
- Level 1 — “Sound.” A (attribution), B (identity), G (process) green across the full corpus; E2 (staleness) holding. Every claim traceable, corpus identity-clean, machinery trustworthy.
- Level 2 — “Measured.” D1 (resample) baselined and its thresholds ratified, then two consecutive cycles inside thresholds; E1 (latency) met; C1 (grades) at 100% with a clean retraction sweep.
- Level 3 — “Best-in-class.” C2 (tiers) + C3 (theses) complete, F1–F4 live, the parity table all-green, at least one frontier row live in public. The claim is dated (“as of …”), renewed each cycle, and auto-suspends on any threshold breach until restored — strict, ruled 2026-08-02. Suspension honesty is what makes the claim safe to print.
Verdicts are the audit desk’s to declare; the numbers come from instruments, not judgment — that is the point of everything above.
