A spreadsheet never tells you it is wrong. Total a column, forget the last row, and Excel hands back a number that is clean, confident, and wrong. Then you email it to your board. I built a system that makes financial models confess, and brings proof with every accusation.
Built solo over three days for the micro1 Frontier Engineering Challenge 2026, a field of roughly 2,400 engineers. The brief asked for a baseline solution and a meaningfully better one, judged on correctness, reproducibility, and engineering judgment rather than demo polish. See the challenge.
the problem
The defects that matter most are invisible on the face of a sheet. A formula edited after the file was last calculated, so the cell displays one number and computes another. An annual churn rate charged in a single month. A closing balance that quietly omits one of its own movement rows. Nothing turns red, so the only defence is a senior person reading every formula by hand, roughly 38 minutes per model, and they still miss things.
what I built
A staged pipeline in which every claim the model makes is checked by something deterministic. Eleven checks recompute the workbook independently and decide what is arithmetically true. The model may then propose any cell no rule anticipated, but it has to arrive carrying the formula it believes belongs there, and that formula gets written into a copy of the workbook and executed. If it computes the value the cell already gives, the proposal is discarded however confident the prose was. The output is an audit memo, a proposed-fix workbook that explains its own diff, and a judgment ledger so a decision made in March is not re-litigated in April.
my role
Sole author of the detector layer, the recomputation engine, the agent pipeline, the evaluation harness, and the web interface. I built the baseline deliberately strong, because measuring yourself against a straw man tells you nothing.
what it does
- Independent recomputation of every formula, catching cells that display one number and compute another
- Eleven deterministic detectors, plus an assumption ledger resolving period basis
- Open-ended model review gated by execution, so a proposal must state its formula and survive being run
- Triage layer that dismisses correct-but-unusual cells using row labels and owner notes
- Proposed-fix workbook written alongside the original, with an `_Audit changes` sheet explaining its own diff
- Judgment ledger keyed to row label and formula shape, so resolved calls carry forward across runs
- Audit memo with a Requires your judgment section and an always-present What I did not check section
results
- Detection F1 0.927 at recall 1.000, a 103% improvement over a strong single-prompt baseline
- 0.0% unsupported-claim rate, because no finding can exist without a recomputed value behind it
- Caught all four defect classes the baseline detected at exactly 0%, the ones that can only be found by executing the sheet
- Cut repeat-run findings from 114 to 60, with 54 of 54 judgment calls never asked twice
- Every number reproducible from committed evidence in about twenty seconds, with no API key
built with
Python, openpyxl, formulas, Anthropic API, pytest, uv, Next.js, TypeScript, pnpm