Research journeys · Record 003
Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records
How can AI-assisted claims remain connected to evidence, analysis, human authorisation, the presented surface and later corrections? This public preprint develops and examines a protocol candidate for that question.
- Status
- Preprint · not peer reviewed
- Version
- v1 · 15 September 2026
- Subjects
- cs.AI · cs.CR · cs.CY
The question behind it
A publication should remain independently challengeable.
The preprint starts from a problem: AI-assisted claims can appear authoritative when evidence, analysis, human authorisation, presentation and correction history refer to different states.
The proposed permission to publish (Publication Authority) applies to one exact checked state: it cannot be transferred and can only be used once. The preprint sets out its rules as a machine-readable protocol candidate called the Publication-Accountability Calculus (PAC-2026).
The rules and accompanying models can be checked by machine. That is what ‘executable publication semantics’ means here, not a running end-to-end system.
Understand the approachSix bound obligations
The fourth preserved version of the checking rules is called SF-4 (semantic freeze) in the preprint. It joins six obligations: passing one checkpoint cannot compensate for missing evidence at another. This list summarises the source, not an established standard.
- 01
Evidence
Evidence is bound to the state being examined.
- 02
Runs & artefacts
Runs and artefacts remain part of the check.
- 03
Measurement
Measurement and its disclosure are included.
- 04
Authorisation
Authorisation remains a distinct, human-bound check.
- 05
Surface correspondence
The presented surface must correspond to the accepted record.
- 06
Lifecycle continuity
Correction and lifecycle information remain connected.
What the preprint reports
Bounded internal evidence, not a universal result.
The abstract reports results for finite formal models, deliberately constructed adversarial cases and earlier implementations. The figures below are taken from the public preprint; they were not rerun or independently recalculated for this record.
Explore the reported internal results
- finite formal models (not AI language models)
- 10
- reachable states without violations of the checked safety rules
- 110,764
- unsafe configurations producing the expected violation or countermodel
- 76
- scored expectations on previously unseen test instances
- 183
- archived observations in the same-host package run
- 240
‘Safe’ refers to the rules checked within these models. In the instance-blind test, the case classes were known but the individual test instances were not known in advance. The abstract says the results support internal consistency, bounded safety, fault detection and limited feasibility of construction.
Where to go from here
The specification follows the reference model.
The preprint is its own public research state. It remains distinct from the published BIS reference-model paper, software artefacts and any later independent or field-oriented validation.
- Explore the BIS reference model
The published Springer contribution with its own evidence and status.
- Research on TechJournal
Further research journeys and their bounded sources.
Read the original
The public preprint.
The exact version v1, dated 15 September 2026, is linked here. The preprint has not been peer reviewed.
- 01
Preprint · not peer reviewed
Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records
arXiv:2609.17631v1 · cs.AI, cs.CR, cs.CY · all eight authors affiliated with UTS
Yifei Dong is corresponding author · submitted 15 September 2026 · DOI 10.48550/arXiv.2609.17631
Read the preprint on arXiv
Cite this work
Version v1 is available as BibTeX and RIS with its DOI, arXiv identifier and preprint status.
Take another look
A preprint is a visible interim state.
Later versions, corrections and a personal reflection will have a visible place here. Until then, the current state remains explicitly v1 and not peer reviewed.
- Which objections would change the specification or its boundaries?
- What evidence would be needed for independent reproduction?
- Which validation remains outside this preprint?
Public arXiv metadata, abstract, version and status linked. Reported internal results kept distinct from general efficacy or truth claims.
The work remains challengeable.
Further research records show what each source supports — and what it leaves open.
Explore the sources