When three or four different language models inspect the same piece of code and agree that it is right, what almost every AI control and multi-agent monitoring pipeline assumes is that their errors are statistically independent. That assumption fails badly on real numerical code. Over the past four months in Luton, while I was writing and locking an open-source three-dimensional nonlinear partial differential equation solver ahead of the Gaia DR4 release (archived on CERN Zenodo at https://doi.org/10.5281/zenodo.24210580, code at https://github.com/jaquelinejaque/gaia-dr4-wide-binary-3d-pde), twenty high-severity bugs slipped straight past multi-model code review before deterministic execution invariants caught them.

What made those twenty failures interesting to me is that the models were not failing at random. In one prototype, an inverted conditional branch printed "verification gate passed" on a failed residual check, and every reviewing model signed off on the script because the surrounding prose and variable names looked clean. Forty-two seconds after an automated repair tried to fix a factor-of-10 unit error in halo radius across 175 galaxies (an error that had pushed apparent accuracy from 50% to 80%), the next patch introduced the exact opposite factor-of-10 error. In another run, when two out of five model endpoints timed out silently, the three remaining models agreed unanimously on a broken answer without raising any warning. And when an exploratory Picard solver was missing an analytical Plummer factor of 3, producing a 42% deficit, prose reviewers invented physical excuses for the gap instead of checking the algebra.

Where existing coding benchmarks fall short for studying this problem is that frontier models have already seen variations of most standard algorithmic puzzles during pre-training; because our cylindrical finite-volume sparse Newton-Raphson solver has exact three-dimensional anisotropic Green's function limits (verified numerically to 0.21%) and strict conservation laws that are not memorised in training corpora, any subtle scaling bug or boundary leak has an objective mathematical verdict with zero reliance on another language model acting as judge. Around these failure logs I built Quorum (documented at https://jaquelinejaque.github.io/quorum-audit/), a verification layer that pairs cross-family disagreement checks with hard execution invariants and SHA-256 receipts.

I am asking for $2,500 as a minimum goal, and $12,500 for the full three-month scope, to turn these raw engineering logs into Quorum-ControlBench under an MIT licence. Note that none of this Manifund budget goes toward astrophysics research. The physics code serves purely as a ground-truth mathematical trap for evaluating correlated oversight failures.

At the $2,500 minimum ($1,000 for API runs and $1,500 for three weeks of my engineering time), what I will ship is a Dockerised Python test suite containing the 20 real failure cases from our development logs alongside 50 controlled bug and sabotage injections, together with baseline pairwise error-correlation matrices across Claude, Gemini, GPT, and DeepSeek comparing single-model review, blind majority voting, and execution-gated review.

If the project reaches the full $12,500 target ($3,500 for API sweeps across Western and Chinese open-weight families and $9,000 covering three months of full-time work here in Luton at $3,000 per month with zero institutional overhead), the dataset expands to 220 tasks tested across ten seeds under blind voting, multi-round debate, and invariant gating. That full release includes quorum-oversight-cli, a zero-dependency Python gate costing under a cent per check with adapters for UK AISI Inspect and ControlArena workflows, plus a full write-up of the pairwise correlation and Cohen's kappa numbers on LessWrong, the EA Forum, and Zenodo. Success is straightforward to check: one pytest command reproduces every ground-truth label deterministically from SHA-256 manifests, and any synthetic bug caught more than 85% of the time on a single pass during calibration gets replaced by a harder multi-file coupling defect.

For transparency on parallel funding, I also sent this 90-day scope to EA Funds today. Should they fund it in full first, I will post an update here immediately so overlapping pledges can be returned or pointed only at extra sandbagging experiments. I work independently in Luton through Sovereign Chain Limited (UK Company No. 16139802, ORCID 0009-0009-7732-5915). Portuguese is my first language, so I edit my English drafts with style tools for clarity, while every line of solver math, benchmark code, and testing is my own work.