A RESEARCH PROTOCOL FOR PUBLICATION AUTHORITY
Make AI-assisted claims independently challengeable.
Connect the evidence, the exact candidate, human authorization and the act of publication. Make their bindings open to inspection.
PAC-2026 research candidate · arXiv v2
Which version was
actually authorized?
A convincing claim can arrive with evidence, a valid signature and a human approval — yet each can refer to a different state.
AIJIM makes the publication transition explicit. The record connects what was checked, what a person authorized, what a reader sees and what was actually published.
Follow the bindingsA stronger status needs a valid witness
bound to the exact state.
No silent epistemic promotion
One record.
Explicit obligations.
What happens when the evidence, the candidate or the required run changes? Inspect four illustrative cases.
Single-use permit can be minted · not consumed
Readiness assumes every premise, beyond the four drawn bindings
Admitted exact context; complete, fresh verification record; exact active authorization; ready surface; accepted lifecycle plan; unused nonce; substantive PASS and GATE_PASS for every E1, R1, M1, A1, S1 and L1 obligation. Permit consumption is a separate publication act.
One exact candidate. All readiness premises assumed.
This illustration assumes an admitted context, a fresh complete record, exact active human authorization, a ready surface, an accepted lifecycle plan, an unused nonce, and substantive and gate passes for all six obligations. A single-use permit can then be minted. Publication is a separate act that consumes it.
Illustrative scenarios derived from the specification and diagnostic artifacts. Displayed human authorization and readiness are assumptions of the example. Selecting a scenario does not sign, authorize or publish anything.
A valid witness establishes the required obligation for this target.
A localized counterexample shows a required obligation is violated.
The required obligation cannot be established from the available inputs.
Three distinct results. No average, no confidence score. Both FAIL and CANNOT_VERIFY block promotion.
What is bound, and where can it be challenged?+
Six publication obligations
Evidence; runs and artifacts; measurement disclosure; authorization; surface correspondence; lifecycle continuity. Each obligation needs its own target-bound result.
A fresh, complete all-pass record derives a permit. The real consumption path binds that single-use permit to one publication record.
Three observer locations
Producer: inspect the objects and bindings. Surface: inspect what the recipient can see. Audit: reconstruct and challenge the portable record.
These are observation locations, not three scores or successive pipeline stages.
Inspect the SF-4 archiveAccountability you can inspect.
Within a declared profile.
AIJIM specifies how publication records remain challengeable. A valid signature, configured signer trust, an analytical result, human authorization and a fresh publication act establish different things. Accountability Closure is relative to the fixed profile.
Protocol conformance does not establish factual truth or automatically certify compliance. PAC-2026 is a research specification candidate, not an external standard.
Inspect the argument and its limitsTest the boundary.
Explain the difference.
Three targeted, author-led diagnostics show which protections the constructed comparison profile supplied only after an explicit repair.
Diagnostic 01
Admit the state before evaluating it.
Can a malformed measurement state enter the publication record?
SF-4 refused four malformed state representations before measurement evaluation. The constructed comparison profile committed them.
An explicit state-domain guard restored agreement on all nine selected cases.
Each mark is one count in a total, not an identified case moving between rows.
How the guard works
Schema admission comes before substantive measurement evaluation.
These four representations form one malformed-state fault class. Schema refusal is not a substantive M1 FAIL.
Three author-led, machine-assisted diagnostics compare unchanged SF-4 with a constructed in-toto/OPA/SQLite publication profile. Counts describe targeted cases, not population accuracy. This is a code-aware comparison, not an external independent evaluation; parity after each repair does not establish an omnibus regression result.
Inspect the evidence collectionWhat can change
without losing the
statistical qualification?
For a fixed hypothesis universe, evidence and level, exact edit regions identify which changes to a claim selection preserve the specified statistical criterion.
A minimum retained core qualifies every selection in the specified edit region. For U = H, the all-core frontier uses quadratic arithmetic work after sorting. The construction builds on established e-closure methods; it does not test the calibration of the e-values.
Analytical validity does not transfer publication authority. A changed candidate still needs the appropriate authorization.
Explore the mathematical evidenceStrongest check: S = {0, 1}B = 2/9 ≤ α = 1/2
FDPS(R) = 1 · mean e = 9/2. All nonempty S satisfy the inequality.
{2} is another minimum core. Selections outside this interval can qualify separately.
What is being checked?
For each selection R, B(R) is the largest ratio FDPS(R) / ēS across all nonempty subsets S of H. Qualification requires B(R) ≤ α. Here FDPS(R) = |R ∩ S| / (|R| ∨ 1).
The inherited guarantee concerns expected false discovery proportion under valid e-value assumptions. The fixture checks exact arithmetic; it does not establish calibration, realised error, or a truth probability for an individual claim.
| Selection R | B(R) | α = 1/2 | α = 1/20 |
|---|---|---|---|
| ∅ | 0 | Qualifies | Qualifies |
| {0} | 2/9 | Qualifies | No |
| {1} | 1 | No | No |
| {2} | 2/9 | Qualifies | No |
| {0, 1} | 1/2 | Qualifies | No |
| {0, 2} | 3/17 | Qualifies | No |
| {1, 2} | 1/2 | Qualifies | No |
| {0, 1, 2} | 1/3 | Qualifies | No |
728problems checked
4,010all-h comparisons
Finite checks on the declared grid, not calibration. Timing results are mixed; no general end-to-end speed advantage is claimed.
Read it.
Run it.
Challenge it.
Start with the published evidence and its reader guide. Follow each artifact’s own version and execution instructions.
PAC-2026 · SF-4 · 0.4.0-sf.4Executable evidence collection
Public collection 1.0.0. Claim-edit regions, publication protections, reference dependencies and an offline reader guide.
The evaluated SF-4 baseline
Versioned specification and original evidence package 0.1.0. Preserve this identity when reproducing the study.
The separate correction candidate
Evidence package 0.2.0. A distinct version; it does not silently replace the evaluated SF-4 baseline.
Portable integrity, mathematics and archived reanalysis are separate from native replay. Native diagnostics use the documented macOS / Apple Silicon profile and pinned dependencies.
Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records
Read the public preprintVersion 2 · revised 21 October 2026 · preprint, not peer-reviewed. Includes the claim-edit regions and native protection diagnostics. Executable materials are available in the separate evidence collection 1.0.0; the evaluated specification remains SF-4.
Cite the public preprint
@misc{olivitiltack2026challengeable,
title = {Making AI-Assisted Claims Independently Challengeable: Publication Authority and a Protocol for Falsifiable Publication Records},
author = {Olivi Tiltack, Torsten and Dong, Yifei and Yu, Kun and Wang, Xu and Liu, Wei and Zhou, Jianlong and Liu, Ren Ping and Chen, Fang},
year = {2026},
eprint = {2609.17631},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2609.17631v2},
note = {Version 2; preprint, not peer-reviewed}
}
THE RESEARCHER
Torsten Olivi Tiltack
First author · PhD Candidate at UTS
My research connects evidence-first systems, publication authority and independently challengeable AI-assisted claims. This specification paper is part of that work, developed with the coauthors credited above.
On TechJournal, I bring the research, hands-on engineering and films together to make the ideas tangible.