Earned Trust: Verifiable Disclosure as a Credential-Substitute for AI-Assisted Work

Abstract

Institutional credentials function as pre-paid trust. People without them have had no comparable way to earn a hearing. This paper proposes earned trust through verifiable disclosure: where a credential buys trust in advance, transparency can earn it afterward. The instrument is the AIast disclosure — a compact counted mark ([LI-]AIast(n)) placed once at the author’s choice — byline, Earned Trust Record block, or heading the AI-use disclosure — with a controlled-vocabulary breakdown of which AI services were used and what each did (function tags: i ideation · d drafting · c critique · v verification · s source retrieval), a workflow statement, and an evidence commitment: working records retained and producible on request. An optional, field-relative Lost Innovator (LI) designation identifies authors working without the credentials their field conventionally treats as a license to author. The paper situates the instrument in the 2025–2026 disclosure landscape (STM's activity classification; Weaver's AID Framework; the Global Reporting Standard consultation), defines when the mark applies (a substantive-influence threshold with a peer-need-to-know test), specifies the affirmative negative declaration AIast0, catalogs the standard's failure modes, and states its limits. The paper is itself disclosed under the standard it proposes. Changes in version 1.8: This version adds specification depth from successive further separate adversarial review rounds, every accepted finding verified against primary sources and incorrect findings declined on the record. Additions: a conformance section (3.4) stating the load-bearing requirements in MUST/SHOULD/MAY vocabulary, with minimal conforming and non-conforming specimens, an evidence-failure rule (a claim whose evidence cannot be produced becomes unsubstantiated, not automatically false), and a manuscript-level multi-author rule; a notation rule making service codes and full names both conforming; the count restated over disclosed entries — user-facing services plus self-hosted models above the threshold; a registry admission test and process (who runs the inference; permanent codes; registration, not endorsement; never blocking an author); an evidence-package specification in Section 3.3 — the four-part retained record, a five-year retention floor yielding to stricter requirements, production scoped to the challenge, redaction narrowed to two marked categories at the author's judgment with its price stated plainly, and optional sealing of the record by published cryptographic fingerprint — with a corresponding altered-records failure mode in Section 8; a distinction in Section 7 between perceived trust, warranted trust, and correctness, engaging the experimental finding that disclosure can lower immediate perceived trust (Schilke & Reimann, 2025); and four further comparators acknowledged in Section 2 (GAIDeT; Resnik & Hosseini; AMEE Guide No. 192; AIR), none of which carries an evidence obligation. The workflow statement now requires explicit human responsibility. Disclosure: ChatGPT's function tags add d, recording that Section 3.4's first draft arose from its review exchange and was adopted with revisions by the author; the service count is unchanged at five. Corrections: empirical claims qualified; coercive phrasing softened; US spelling standardized; separate AI critique distinguished from independent human review; and references completed with verified locators. Version 1.9 — changes from version 1.8 (the ETR edition). This version adds the Earned Trust Record (ETR): a final page carrying the work's session record — dates, shifts, services, roles, and responsibility — encoded into a QR code that contains the record itself. Scan it from print, tap it on screen, or read it at earnedtrust.org/reader. The ETR mark appears on the cover's publication block. No changes to the standard's normative text. Record integrity is verifiable; authorship disclosures are provided by the author. THE SEALED RECORD (added August 2026) This version of the deposit adds the work's sealed record. Two files now accompany the paper, and here is exactly what each one is, what it does, and how anyone can check it. FILE ONE — THE PUBLIC MANIFEST (a small text file, readable by anyone).When this work was prepared for sealing, the author gathered its retained working files — every edition v1.6.2 through v1.9, both volumes of the work record, the deposit note, the adversarial-review rulings, the record link, and the build materials — twenty files — into one folder, and on 2026-08-04 that folder was sealed: every file was fingerprinted using the SHA-256 algorithm. A SHA-256 fingerprint is a sixty-four-character code computed mathematically from a file's entire contents. If even one letter of a file changes, its fingerprint changes completely, and constructing an altered file that produces the same fingerprint is computationally infeasible under current knowledge — for any person, any computer, or any artificial intelligence. The manifest is the complete list of those sealed files: each file's name, its exact size in bytes, and its fingerprint. It proves what existed on sealing day and makes any later alteration detectable. FILE TWO — THE LOCKED BOX (earned-trust-standard-vault-2026-08-04.7z).This is the sealed folder itself: the files named in the manifest, plus the manifest, stored in a single archive encrypted with AES-256 — the same class of encryption used in banking. Anyone may download this archive. Anyone may compute its fingerprint and confirm it matches the one published here: SHA-256 of the locked box: 2bc13593ba51373bc1f10ab004b64ec713a5eeb45a7aeb574c68a7b3fdb5979d The names of the files inside the archive are visible to anyone who opens it — the list is the window. The contents of the files are what the encryption seals. The archive's contents can be opened only with the author's key, which exists solely in the author's custody and is not held by any company, service, or artificial intelligence. HOW TO REACH THE WORK'S RECORD.The vaulted logs are sealed by the author. The work's record card — the index of how this work was made — is reached through the QR code printed in this edition. Scanning that code opens the record card in the Earned Trust Reader, where the index can be read in full and where any file later produced by the author can be verified against the manifest's fingerprints on the reader's own computer. HOW TO REQUEST THE SEALED EVIDENCE.Editors, reviewers, and readers with a genuine question may petition the author for the sealed working files. Email the author directly: William@earnedtrust.org. The author produces requested files individually, at the author's discretion; each produced file can then be checked against the manifest published here, proving it is byte-for-byte the file that was sealed on sealing day. Questions about the Earned Trust process itself: William@earnedtrust.org. WHAT THIS PROVES AND WHAT IT DOES NOT.The fingerprints prove that the sealed files have existed, unchanged, since the date of this deposit. A seal proves listed files unchanged; it never by itself proves the record was complete when sealed. Stating that limit is part of the method. EDITION v1.10 (August 2026) — THE SIGNATURE RESTATED.This edition also corrects the work's signature under the author's ruling that the person carries no marks — the work does. The author's name stands alone; the LI-AIast5 disclosure mark stands in the seal row beside the Earned Trust Record mark. The QR code in this edition carries the work's corrected record card, including the author's petition contact. Content is unchanged beyond the front matter, the disclosure line, and the record page. This work was made and documented under the Earned Trust (AIast) standard for verifiable disclosure of AI-assisted work, DOI: 10.5281/zenodo.20719927. The record is public. The receipts are held. Both are checkable. WHAT VERSION 1.11 ADDED (August 2026) Two changes were made to the standard's text, both dated 2026-08-06, and both are recorded in the document's own version history. FIRST — THE TRANSPARENCY DILEMMA, CONFRONTED (Section 7). Published experiments have shown that disclosing AI use can lower a reader's trust rather than raise it — thirteen experiments, with the effect running through perceived legitimacy (Schilke & Reimann, 2025, Organizational Behavior and Human Decision Processes). Version 1.11 stopped citing that finding in passing and confronted it at full strength, in a section written after the full text of the study was read at its source. The section states the objection as a critic would: if disclosure signals deviation from the norm of human work, a richer disclosure is a louder confession, and this standard should erode trust rather than earn it. It then answers from the experiments themselves: every disclosure they tested was a bare, unverifiable one-line label with no record behind it — verifiability, the property this standard is built around, appeared nowhere in the design; the trust penalty fell as legitimacy rose, and the study's own authors named alignment with professional standards as the intervention most worth testing; and the actor exposed by a third party was trusted least of all, so candor remained the cheaper bet on the objection's own evidence. The section closes with the standard's falsifiable prediction and the three-arm experiment that would test it: if verifiable disclosure carries no smaller penalty than a bare label, the credential-substitute thesis fails at the reception layer. A standard about verifiable claims should make one. SECOND — THE RETENTION-DISCHARGE RULE (Sections 3.3 and 3.4). The retained record is the author's, not the service's. Most AI services do not promise five years of producible history, and several delete or overwrite session data on a far shorter cycle. Version 1.11 therefore made explicit that the retention duty is discharged at the