Selection state
Chooses what may runProblem registry + source audits
Exact statements, lanes, priority, recognition, current status, verification contract, and external acceptance route.
A proof engine
that remembers.
Proof Factory is an AI-assisted research system for making useful mathematical contributions, no matter how small or large, while pursuing hard finite problems without losing the evidence, failures, or context between passes.
01 / END-TO-END ARCHITECTURE
One evidence loop
Every pass begins from current source material and durable state. It ends by changing that state only when the evidence survives its gates.
Primary sourcesProblem registryPrior-art registerResearch memoryLab events
- 01Scout + intake
Find a real opening
Audit a current, recognized source; check comments, claims, active work, acceptance path, and whether a compact certificate is possible.
Output · exact target - 02Baseline
Map what is known
Read primary literature, capture established facts, datasets, live methods, exclusions, novelty risks, and verification tools.
Output · sourced research map - 03Strategy
Choose a discriminator
Compare witness, impossibility certificate, structural reduction, alternative formalism, and adjacent-field transfer by decision value per cost.
Output · prediction + stop rule - 04Research + lab
Run the bounded test
Reason in a research pass; move long deterministic work to a resource-capped lab with immutable inputs, checkpoints, manifests, and logs.
Output · artifact + receipt - 05Adversarial verification
Try to break it
Replay cleanly, check scope and quantifiers, use an independent implementation or formal kernel, and repeat the novelty search.
Output · verified scope - 06Decision
Learn, redirect, or release
Record the exact delta. Continue, hold, redirect, or close the route. Candidate work still waits for contribution, skeptic, and human gates.
Output · next state
↳
Feedback, not amnesia. Facts, failures, route scores, artifacts, and the next first action return to the next epoch. A failed search narrows only its recorded scope.
02 / RESEARCH ORCHESTRATION
How a pass thinks
Model agreement is never treated as validation. Reconnaissance diversifies the search; the principal chooses and audits one concrete test.
Admitting evidence
Canonical route brief
Statement, source status, research map, tactical incumbent, challenger, roadmap, prior art, and any completed lab event.
Terra reconnaissance
Source discriminatorExact status, smallest executable test, outside acceptance path
Prior-art challengerOverlap, missing premise, genuinely different route
Experiment verifierControls, failure modes, stop conditions, independent check
Roles are admitted only when their evidence can change the route.Sol principal
Select one bounded discriminator
- Audit delegate claims
- Predeclare success and failure
- Reject duplicated mechanisms
- Update the five-route portfolio
Execution split
≤ 2 minutesRun inside the research pass with hashes, seed, limits, logs, and measured output.
> 2 minutesSubmit a shell-free lab job with pilot, immutable inputs, checkpoints, resource caps, and durable completion events.
Review return
Prediction → observation
Record surprise, reusable assets, constraints learned, failure signature, bottleneck update, and the cheapest next discriminator.
03 / KNOWLEDGE + STATE
What the engine remembers
Different stores answer different questions. Mutable projections make the next pass efficient; append-only records and hashed artifacts preserve what actually happened.
Working memory
Lets the next pass resumeResearch map + tactical memory
Established facts, scoped exclusions, open leads, strategies, route fingerprints, prediction/observation learning, and next-session checkpoint.
Anti-rediscovery layer
Stops renamed repetitionPrior art + cross-problem brain
Nearest historical mechanisms, required material delta, source URLs, shared concepts, reusable methods, and explicit transfer hypotheses.
Immutable evidence
Records what happenedAttempts + experiment receipts
Claims, citations, hashes, commands, seeds, limits, logs, manifests, certificates, checker results, failures, and human adjudications.
Provenance + projection
Makes the trail inspectablePer-problem Git repos + public ledger
Readable checkpoints, ordinary-sized artifacts, AI/tool disclosure, publication packets, limits, and the public distinction between attempt and accepted result.
04 / PERSONAS
Six roles, separate powers
The names describe responsibilities, not claims of personhood. Separation keeps discovery, verification, and publication from collapsing into one optimistic voice.
Finds legitimate openings
Checks live source status, upstream work, recognition, tractability, certificate shape, and a real external recipient.
Cannot lower the intake barChallenge before committing
Source, prior-art, and experiment specialists produce compact memos with falsifiers and stop conditions.
Memos are leads, not evidenceOwns the research decision
Selects one route, audits every inherited claim, executes bounded reasoning, and leaves structured state for the next epoch.
Cannot self-promote a resultRuns deterministic compute
Executes argv without a shell, under CPU, memory, time, workspace, checkpoint, and artifact-growth controls.
Queueing compute is not evidenceStarts from the artifact
Checks the statement, scope, certificate, independent replay, and novelty without inheriting the researcher's conclusion.
Model agreement does not countAccepts responsibility
Charlie reviews the bounded release packet and explicitly approves publication. Outside experts or maintainers determine external acceptance.
Approval is not peer review05 / CLAIM CONTROL
The publication firewall
A correct computation can still be unoriginal, irrelevant, or too narrowly scoped. Each gate asks a different question and fails closed.
- 01Contribution gate
Is there a meaningful delta, recognized contribution class, reproducible novelty search, independent validation, and real acceptance path?
- 02Isolated skeptic
Does the claim survive statement, scope, certificate, adversarial, and literature checks without relying on the research transcript?
- 03Charlie approval
Is the bounded packet honest, useful, appropriately disclosed, and ready to put his name behind?
- 04Mechanical release
Publish the artifacts, manifest, citations, limitations, AI disclosure, validation plan, and public classification.
- 05External acceptance
Repository merge, expert confirmation, venue review, or another sourced outside event. Self-publication alone scores zero.
AttemptCandidateComputationally verifiedIndependently reviewedPublished / accepted
Inspect the machinery.
All engine code, public problem repositories, research records, limitations, and tool disclosures are inspectable. Commits establish provenance; certificates, checkers, and outside review establish mathematical confidence.
View source and repositories →