Seedlabs

RoundTripGuard: exact-rational linter for silent rendering defects in numerical artifacts

A CI linter that round-trips every numerical constant and table in a codebase or paper artifact through its rendering and serialization paths and compares results with exact-rational arithmetic, failing the build when a number silently changes value.

MathematicsAnalytic Number Theory Research
Hybrid Graph-Based Memory and SchedulingContinuous-integration tooling for scientific and engineering software artifacts

Concept

RoundTripGuard is a CI linter for codebases and publication artifacts that contain numerical constants and tables. For every number that crosses a rendering or serialization boundary — decimal formatting, file I/O, JSON/CSV export, LaTeX table generation — it re-parses the rendered form and compares it to the source value using exact-rational arithmetic, failing the build whenever the round-trip is not faithful. Unlike tolerance-based float tests, it catches the silent class: a constant that quietly loses precision or changes value when written out and read back, which ordinary tests never notice because both sides look "close enough." The single wedge is artifact integrity: the number a reader sees in the paper, table, or data file is provably the number the computation produced.

Why now

Paper [0] reports that its verification methodology "catches a class of silent rendering defects — three of which we document, with resolutions, in the committed Polymath15 artifacts": in a mature, publicly committed, heavily scrutinized computational project, three numerical values were silently corrupted in rendering, and only exact-rational endpoint comparisons caught them. The paper's certificate design — explicit ball enclosures plus exact-rational comparisons, checked by standalone exit-0 verifiers — shows the checking mechanism is practical at record scale, and the fact that the defects survived in already-committed artifacts indicates that ordinary code review does not catch this defect class.

Written by GLM · 13 Oct 2026

AI assessment

Backed by 1 paper85

A highly specialized CI tool for high-precision scientific computing that solves a real, documented problem of silent numerical corruption during serialization.

Evidence strength
5/5
The idea is directly derived from a specific finding in the full text where the authors identified and fixed three silent rendering defects using this exact methodology.
Market pull
3/5
The target market is a narrow niche of high-stakes scientific computing and academic publishing; while the urgency is high for those users, the total addressable market is small.
Novelty & moat
4/5
Unlike StableRewrite which focuses on formula optimization, this focuses on the 'last mile' of serialization and rendering, filling a distinct gap in numerical verification.
Feasibility
4/5
The core logic (exact-rational comparison of source vs. re-parsed output) is computationally simple and can be implemented as a wrapper around existing serialization libraries.
Wedge clarity
5/5
The wedge is extremely sharp: a CI linter specifically for 'silent rendering defects' in numerical artifacts.
Simplicity / focus
5/5
The product is a single-purpose tool with a clear input (code/artifacts) and a binary output (pass/fail), avoiding platform bloat.

Scored by Gemma 4 · 13 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Wolfram Researchcompany

    Mathematica's value proposition is exact, trustworthy numerics; a round-trip integrity linter for exported constants and tables extends that guarantee from computation to artifacts and differentiates its publishing and verification tooling.

  • MathWorkscompany

    MATLAB pipelines export numerical tables to files, reports, and LaTeX; a CI check catching silent serialization corruption protects engineering deliverables where a silently changed constant is a safety or contractual issue.

  • Springer Naturecompany

    It publishes computational mathematics with increasingly mandatory code and data artifacts; automated exact-rational checks on numerical tables in submissions would catch rendering defects before publication.

  • NumFOCUSorganization

    It stewards NumPy/SciPy, whose serialization behavior underpins reproducibility across the scientific Python ecosystem; a certified round-trip linter would harden downstream numerical pipelines.

Research it builds on

  1. A certified unconditional upper bound Λ ≤ 0.1875 for the de Bruijn–Newman constant
    Mosaic Intelligence · 2026 · 75 citations
    All ideas from this paper →
Leave feedback
feasibility