Auditable scientific publishing

Research method

Auditable scientific publishing in the AI era

This project is published as an auditable research object: the paper, the claims underneath it, their dependencies and status, the evidence attached to them, the calculations, the version history and the machine-readable records needed to inspect the whole chain.

AI systems are beginning to make genuine contributions to research-level mathematics. The implication for scientific publishing is not that AI-assisted work deserves extra trust. It is that sophisticated arguments can now be generated, explored and revised much faster, while verification remains expensive. The publication itself should therefore expose more of the structure needed to check the argument.

That is the model used here. Human judgement sets the research direction, assumptions and interpretation. AI systems contribute synthesis, mathematical exploration, consistency testing, literature mapping and rapid iteration. The author remains responsible for every published claim. AI participation is part of the research method; it is not evidence that the physics is correct.

The standard is auditability

The relevant question is not whether a sentence was produced by a human or an AI. The scientific question is whether the claim can be identified, attributed, traced to its premises and evidence, reproduced where appropriate, and challenged without reconstructing the entire project from prose.

One claim, one controlled record

Every substantive scientific claim on this site resolves to a controlled record. The record states the exact technical claim, its public wording, its status, assumptions and dependencies, its source in the paper, the external foundations it uses, and the limits or open calculations that remain.

Public prose can simplify a claim, but it cannot silently broaden it or change its status. A claim may be Derived while depending on Selected or Proposed premises; the dependency chain therefore remains part of its scientific status. This makes criticism addressable: each record lists its upstream premises and its downstream claims, so a reader can challenge one premise and see which conclusions inherit that challenge.

Evidence, calculations and provenance

The reference record is claim-aware. A source is not listed merely because it is relevant to the general subject: each entry states why it is cited and what part of the argument it supports. The project manuscript remains the source of the new construction; external references supply the established results, measurements and comparison context on which particular claims depend.

The numerical chain is published separately as a calculation record, and the paper is released in PDF, native HTML and source form. Re-running the arithmetic verifies the numerical chain; it does not by itself verify the physical selection of the channel. That distinction is part of the published record.

The same boundaries for humans and machines

Human and machine readers should inherit the same scientific boundaries. The claim register, the dependency graph and the reference data are published in machine-readable form alongside the human-facing pages; research-object.json is the canonical machine entry point, and llms.txt states the calibration rules an automated reader should preserve when summarising the work.

The goal is semantic parity, not machine certification: the machine representation exposes the same claims, statuses, dependencies, open calculations and falsifiers that a careful human reader sees. Operational prompts, model orchestration and internal transcripts are not part of the scientific record; the public record is the argument, its assumptions, calculations, sources, versions and corrections.

Why this matters now

Research-level mathematical reasoning by AI is no longer hypothetical. In May 2026, an OpenAI reasoning model autonomously disproved a longstanding Erdős conjecture in discrete geometry, with the proof checked by external mathematicians and reported in Nature. In June, the Leiden Declaration on Artificial Intelligence and Mathematics, endorsed by the International Mathematical Union, called for transparent tool disclosure, precise references, open-science practice and work that makes review easier.

These developments point in the same direction: stronger generative capability should be matched by stronger auditability. This site is one working implementation of that principle. Context dated August 2026; reviewed at each substantive release.

What the format changes

This format does not certify that the physics is correct, and it does not replace expert review. It changes the object being reviewed. A reviewer can move from a headline statement to its controlled claim, inspect the dependency chain, challenge a selected or proposed premise, reproduce the numerical calculation, inspect the sources used for a specific step, and see how the wording and status changed between versions.

The aim is to make review more granular, faster to navigate and easier to automate, without transferring scientific responsibility to the automation, and to make individual claims easy to challenge directly. A challenge to one premise should reveal what changes, what survives and where the disagreement actually sits.

A reusable publication model

Nothing about this publication architecture is specific to cosmology. A different research programme could replace the physics while retaining the same pattern: a formal paper, stable claim identifiers, explicit status and dependencies, source provenance, reproducible calculations, versioned corrections, stated falsifiers and machine-readable exports.

cosmological-constant.org is therefore also a working demonstration of one possible publishing model for the AI era. Its value as a publication experiment does not depend on this particular cosmological mechanism being correct: if a claim fails, the format should make it easier to identify where, propagate the consequence through the dependency graph, and preserve the results that remain intact. The useful parts should be copied, tested and improved.