A documented, citable methods case study in AI-assisted mathematical research.
Matthew J. Goss Jr. | Quantiterate | Preprint v1.0, published November 1, 2026
What should an AI mathematician remember about an argument that fails?
Many AI-for-mathematics systems are evaluated by the theorem they ultimately prove. A months-long investigation has a harder problem: it must retain what has already been tried, distinguish a genuine mathematical obstruction from a defective review, preserve the exact status of conditional arguments, and decide whether a new approach actually changes the next proof obligation.
This study documents how Quantiterate MathLab evolved during 2026 to address that challenge. The archive records a human-directed adversarial RH War Room in March, four-role MathLab persistence and the Fibonacci Jury design specification in June, corpus synthesis and proof-ledger planning in the summer, and obstruction governance and source-integrity audits in September and October. The manuscript locates these in a broader development history that also includes Crystal Farmer, Strand Loop, proof-writer designs, observer channels, and research-state checkpoints. It does not claim that Quantiterate invented multi-agent mathematical research or deployed every component described in historical specifications.
The research contribution
The paper's unifying idea is an evidence-justified research-state transition: a claim should be promoted, held, quarantined, abandoned or reopened only when the evidence needed for that particular change has been identified and checked. An algebraic identity, a local code replay, an AI critique, a completed mathematical proof and an independently refereed result support different decisions. These distinctions are not interchangeable confidence scores.
An obstruction ledger makes failure actionable rather than merely archival. The system's candidate methods are compared against known blocked implications, and the work stops when a new construction reaches an unchanged obligation without producing new evidence. Conversely, a reviewer objection caused by missing source text should trigger evidence repair, not an assertion that the mathematics itself is false.
Two technical case studies
Riemann Hypothesis, R88–R89. A sequence of Abel recoupling and meet-kernel calculations produced valid scoped algebraic identities and a positive semidefinite meet-form structure. They did not establish the required signed arithmetic cancellation. Comparison with earlier R80–R86 obstructions led to a scoped NO GAIN decision: the tested branch did not remove the load-bearing obstacle. This is a research-routing conclusion, not a theorem that all related approaches fail, and not a proof of RH.
Navier–Stokes, War Room evidence delivery. Source reconstruction and deterministic tests revealed that some supposedly missing cited statements were present, but the review pipeline had omitted or mislocated critical context. Correcting that fault changes the status of specific review objections; it does not settle the underlying PDE argument or establish Navier–Stokes global regularity. The paper separates source availability, reviewer context depth and mathematical validity.
What this paper does not claim
It does not establish a Millennium Prize result, claim that AI editorial ratings are independent human peer review, report a completed randomized comparative study, or claim that the original Fibonacci Jury protocol's proposed 19–20× confidence amplification has been empirically achieved. Some components are historical specifications or manual processes, not verified production deployments.
Why publish a negative result?
Stopping a mathematically redundant route can conserve resources and prevent future researchers from making the same inference. Correcting a false missing-source accusation can preserve the integrity of adversarial verification without inflating a mathematical claim. Both are consequential and testable decisions in long-running research, even when the headline theorem remains unresolved.
Access and correspondence
Preprint: Zenodo, version 1.0 — DOI 10.5281/zenodo.23867304.
Evidence companion: a source-linked and rights-screened public index is in preparation; restricted mathematical source packets will be supplied to qualified referees when appropriate.
Editorial status: documentary preprint, not an ACM-accepted paper. Formal journal submission will be recorded when a journal receipt exists.
Corrections: mathematical counterexamples, source discrepancies and provenance errors can be reported through the existing Quantiterate research correspondence channel. Corrections to a published preprint will be recorded in new dated versions, not silently overwritten.