Engineering
How the models are built, tested, and kept current
Written for engineers and operators evaluating us. Every claim in these pieces traces to something you can open — a spec, the Concordance harness, the facts registry, or our public telemetry.
2026-08-23
The Test That Asked for More Precision Than a Double Has
A property test failed a conservation law in our tax models. The arithmetic was correct to within less than one representable number. The assertion was demanding better than that — and on measurement, one assertion of its kind was reddening about 6% of builds while four siblings carried the same defect unfired.
2026-08-23
We Shipped the Same Units Bug Twice
Two models overstated a headline number by roughly 12×. Two independent implementations agreed on both wrong answers, and the test harness stayed green. What that says about what verification actually proves — and the gate we added once we understood the class.
2026-08-16
Nine Models Took an Open-Book Math Test
FinPlanBench hands a language model the entire published spec of a deterministic financial model — formulas included — plus one set of inputs, and grades the exact outputs. First published scores: two models above 89% at harness tolerance, one at 98.7% within-a-dollar that is almost never exact, and two that fail no matter how generous the bar gets.
2026-08-16
Two Implementations or It Doesn't Ship
How we Concordance-test 55 financial models: an independent Python implementation rebuilt from the spec alone, 13,750 deterministic cases diffed at one part in a billion, and a CI gate that stops any release where the two disagree.
2026-08-16
Your Fintech Copilot Has No Answer Key
LLM eval stacks grade tone, safety, and retrieval. Nothing in them knows whether the refinance break-even your copilot just quoted is numerically right. Wiring in a computational answer key takes an afternoon.
2026-08-16
TY2026 Retirement and Estate Numbers, With Receipts
Every IRS and SSA constant our models consume for tax year 2026 — the value, the period it applies to, the primary source, and the date a human last checked it. The registry is the source of truth; this post is its narration.
Product and integration guides live under /writing; the method itself is defined at /proof.