The evidence and language contract

Binding on every signal, citation and surface in the Blanket Builder and the signal inventory. Set by the operator 2026-08-20. Enforced mechanically by viewer/verify_signal_inventory.cjs.

1. The language contract — why we never write "fact"

"science is never done and we never say FACT, we say confirmed observation, and standing durable measure with observed evidence." — the operator, 2026-08-20

This is not decoration. A word like fact closes a question that the scientific method holds open by design; it invites a reader to stop checking. The purpose of these surfaces is the opposite — to make checking easy, and to restore rather than degrade a reader's grip on how knowing actually works.

Permitted vocabulary

Forbidden on these surfaces: fact, proven, proves, definitive, settled science, obviously, certainly, undeniable, the truth is. The gate greps for these.

Hedging is preserved, never stripped. If a source writes "may", "is thought to", "we propose", those words survive into our quote and into our summary. Removing an author's hedge is a form of fabrication.

2. The provenance contract — click-through to the original, every time

The operator's requirement:

"If we cite, link to the quote, link the quote to the source, make the source a copy in our repo and UI and the source copy MUST fully link to the original source."

Every signal therefore carries a four-link chain, and the UI renders all four:

CLAIM  ──▶  VERBATIM QUOTE  ──▶  LOCAL SOURCE COPY  ──▶  ORIGINAL (DOI / publisher URL)
 what we      <=40 words,          mirrored in-repo        the thing itself, so a
 assert       with a locator       where licence allows    reader verifies us

Required fields per signal (schema uni.flagellum.signal-inventory/1.0.0):

• verbatimQuote + quoteLocator — the exact words, and where in the document they sit
• apaReference — full APA 7, the citation style a journal, a paper and a non-fiction science book all carry
• doi / stableUrl — the persistent identifier
• urlFetched + retrievedOn — what was actually retrieved, and when
• localCopy — path to the mirrored source in this repo, or null with the reason
• licence — as stated by the source page
• fetchStatus — FETCHED_AND_READ · ABSTRACT_ONLY · PAYWALLED · FETCH_FAILED
• sourceType — PEER_REVIEWED · REVIEW · TEXTBOOK · PREPRINT · DATABASE · PROFESSIONAL_BODY · BLOG_OR_POPULAR · FRINGE_OR_CONTESTED

Mirroring rule — copy what we may, link what we may not

• Open licence (CC-BY, CC0, public domain, PMC open subset): mirror the full text into docs/sources/ and record the licence. The local copy header links back to the original.
• All rights reserved / paywalled: we mirror only the metadata and the short quotes we rely on (fair dealing for criticism and review), and the entry is marked localCopy: null, reason: "licence does not permit mirroring". We never pretend to hold what we do not hold.
• Every mirrored file carries a provenance header: source URL, DOI, retrieval date, licence, sha256 of the retrieved bytes.

OPEN GAP, recorded 2026-08-21 — mirroring is NOT yet done

localCopy is populated on ZERO of the 159 signals. The operator's requirement — "make the source a copy in our repo and UI and the source copy MUST fully link to the original source" — is NOT yet satisfied. The fetch fleet returned metadata and quotes but did not mirror full texts.

Worse, and corrected the moment it was found: the Blanket Builder drawer was rendering "not mirrored — licence does not permit" as a fallback on every unmirrored row. That asserted a licence ruling nobody had made, on 159 rows. It now reads "NOT MIRRORED — mirroring has not yet been attempted for this row", which is the truth. The one-line lesson: a fallback string is a claim.

To close the gap: for each row whose openAccess states an open licence (CC-BY, CC0, PMC open subset), retrieve and mirror the full text into docs/sources/ with a provenance header, and set localCopy. For the rest, set localCopy: null with the licence text that was actually read, not an assumed one.

The anti-fabrication rule

A signal may only exist if a source was actually retrieved and read. Not remembered, not inferred, not "well-known". An entry with no verbatim quote is not an entry. A failed fetch is recorded as a failed fetch — an honest negative row is worth more than a confident guess.

Never guess a DOI, year, volume, or page. Absent → null plus "not stated on the fetched page".

3. What we claim, and how we say we are testing it

Every signal and every blanket in the builder answers five questions on its face, because a claim without these is an assertion, not science:

1. What we claim — stated in permitted vocabulary, at the strength the evidence licenses.
2. Why we claim it — the quote, the citation, the click-through.
3. How we are testing it — the gate, the run, the artifact; or honestly, NOT YET TESTED.
4. Controls — the negative control, the null model, the comparison that could have come out the other way. A claim with no control is labelled NO CONTROL — reported only.
5. Adversarial check — who tried to break it, and what survived. Includes contradicting sources, which are carried side by side and never reconciled by us.

4. Contradictions are evidence, not noise

Where two sources disagree, both are shown, both are cited, neither is silently dropped, and the disagreement is labelled CONTESTED. Resolving a contradiction requires new evidence, not an editorial decision. (This estate already carries live examples: Ito-vs-Wadhwa dwell direction; the Johnson/Singh attribution conflict.)

Fringe and edge material is welcome and labelled FRINGE_OR_CONTESTED, and never appears without the mainstream response beside it. Excluding contested work hides the argument; printing it unlabelled misrepresents its standing. We do neither.

5. Why this exists

"we have no pride, no ego, no hubris, we are the EFE curiosity turnt up and eager to share with the world how to EFE and stay grounded and produce durable science that others can use to further their observational learning." — the operator

A published failure is worth the same as a published success, because both let someone else move. The provenance chain exists so that a stranger can check us without asking us — and find us honest, or find us wrong, on the evidence.