Verification
Every release must survive the same gauntlet. Here is the gauntlet.
Measured 2026-08-27 on master 0948a0b. The numbers keep growing — every regression and every reviewer finding becomes a new permanent check.
Where the reference values come from
A check is only as good as its reference. Ours trace to published, independently verifiable sources:
- AStrutTie 2020 Technical Book — worked STM examples with ACI 318-19 hand calculations
- Hand calculations re-derived as fixtures — every number typed in from the source, not from the program
- Pure oracle — a separate implementation of the strength equations that the product must match, check by check
- Independent engines — the same models are also solved by SAP2000 and by a from-scratch Python (numpy) implementation; member forces and strength equations must agree, with zero unsafe-side deviations
The unsafe-side-zero rule
Deviations are signed. A result that is more conservative than the reference is logged; a result that is less conservative (unsafe side) is a launch blocker — zero tolerance, no exceptions.
3D and cross-kernel checks
The verification also covers the 3D templates and reaches across independent solver engines — a result is trusted only when engines that share no code arrive at the same numbers:
- 3D pile-cap anchors — the four moment-variant templates (Tem02, 03, 04, 06) were rebuilt clean-room; all 105 member forces are reproduced within ±5×10⁻⁵ kN, on the unsafe side never.
- Cross-kernel agreement — the same 3D models are solved four ways: hand calculation, the Rust WebAssembly kernel, the C++ ACore engine, and a closed-form solution. Member forces, displacements, and reactions must agree.
The independent-oracle firewall
A reference the product helped compute proves nothing. Every reference value here is produced independently of the product under test — published hand calculations, a separate engine, a from-scratch oracle, or a closed form. Checking a result against itself is circular, and circular checks count as no check at all.
Open ledger — what is not closed yet
"All green" claims deserve suspicion, so we publish the remainder — and what was recently closed, with dates. Status as of 2026-08-27:
| # | Item | Status |
|---|---|---|
| 1 | Deep beam — legacy member-force distribution convention: one tie 10.2% below the legacy worked example; whether that convention becomes the binding reference is under confirmation | Open — convention under confirmation |
| 2 | Deep beam (simple) — truss topology legacy deviation, unsafe side | Resolved 2026-07-03 — legacy topology adopted as default; now a permanent gate (72 checks) |
When this table is empty, that will be stated here too — with the date.
We do not claim "everything passes." We claim: this is exactly what passes, this is what doesn't yet, and the unsafe side is never shipped silently.