The models make the work
Contestants receive frozen test conditions and produce the artifacts being evaluated. Their identities remain hidden during a blind Duel.
Human judgment
stays human.
Proofrun is an independent Finland-based editorial project built to make AI testing more inspectable: show the artifact, preserve the conditions, disclose the assistance, and keep the conclusion smaller than the evidence.
The person behind Proofrun
RRoni
Human editor
Finland
Human editor
Finland
One project. One accountable editor.
Proofrun is currently created and operated by Roni, an independent builder and reviewer in Finland. There is no hidden research team: one person chooses the editorial questions, inspects the evidence, makes the requirement decisions, assigns the craft scores, records uncertainty, and approves what becomes public.
AI is a close collaborator in the software, research workflow, evidence organization, and drafting. It does not become an unnamed human panel, and it does not own the final verdict. Roni remains responsible for every published judgment and correction.
Small is a limitation worth naming—and a reason to keep every claim inspectable.
Who does what01 02 03 04
The human editor judges it
The editor inspects the artifacts, decides requirement outcomes, confirms craft scores, records confidence and rationale, and locks the verdict before reveal.
AI assists the investigation
AI systems help build Proofrun and may analyze preserved evidence or draft editorial copy. Assistance is recorded on the affected report; findings and prose require human acceptance.
The record constrains everyone
Prompts, settings, attempts, failures, evidence, costs, evaluation state, identities, and limitations cross a guarded publication boundary and remain attached to the claim.
One-reviewer limitation
A named editorial view, not a consensus score.
A single reviewer can be consistent, attentive, and accountable, but cannot provide inter-rater agreement. Proofrun therefore publishes the frozen requirements, weighted criteria, evidence, confidence, rationale, and limitations needed to examine or disagree with the judgment.
Scores describe the artifact in that run under that rubric. They are not universal measurements of a model, and they should not be mistaken for a laboratory consensus. Independent audits and additional reviewers may become valuable later; Proofrun will not imply that they already exist.
Editorial independence
No paid winner. No invisible exception.
Proofrun does not sell placement in verdicts. Any sponsorship, supplied credits, employment, investment, or other material relationship relevant to a test must be disclosed beside that result. A relationship cannot rewrite a locked evaluation after identities are revealed.
Model providers do not receive advance approval over conclusions. Access method, resolved provider, reported generation cost, retries, and protocol exceptions are preserved when the publication format supports them.
Trust comes from visible boundaries, not a claim of perfect neutrality.
Corrections and privacy
Correct openly. Collect little.
Historical evidence is not silently rewritten. Material corrections should appear as explicit amendments, while the original hashed record remains attributable. A failed run may be published, but a rerun is a new record rather than a replacement.
The journal currently has no reader accounts, comments, or newsletter form. Interactive web artifacts run on a separate isolated service and are linked through hash-bound receipts; the public journal does not execute contestant HTML.
Questions and correction requests can be sent to @proofrunai ↗.
The compact promise
Show enough to disagree well.
Proofrun is not a laboratory claim of universal model quality. It is a growing journal of bounded, reproducible editorial observations. The aim is to let a reader see the work, understand the procedure, notice the uncertainty, and reach a different judgment without having to trust a hidden process.