Method statement
How verification works
The promise is narrow and checkable: the AI never does arithmetic. This page shows the machinery that enforces it — and what it does not promise.
What happens to your problem
Route
One AI call reads your problem. Its only allowed action is filling in one solver's input form — it cannot answer in prose, and it cannot skip the form. If nothing fits, it must say so.
Compute
A deterministic solver — plain engineering code, the same every time — does all the math. Units are converted and dimension-checked at the boundary; the solver works in SI internally.
Explain
A second AI call writes the explanation from the solver's output only. It never sees your original problem text, so it has no numbers to do arithmetic on — it can only describe the computed steps.
Audit
A deterministic checker extracts every number from the explanation and requires each one to appear in the computation's own number set. If a number cannot be matched, the explanation is regenerated once — and if it still fails, the solution is visibly flagged and the step trace is what you trust.
Where the answers are checked against books
Every solver ships with reference problems from published sources — FE review texts, standard handbooks, classic course texts. Each records where its answer comes from and how it was verified, and every one runs as an automated test on every code change. Three rules are absolute:
- Reference answers are never fabricated.
- Reference answers are never generated by an AI.
- A solver is never used to check itself — answers come from published sources or independent hand computation, recorded in the test file.
Specimen — an actual reference record from the test suite
- problem
- US single catchment: C=0.3, i=2 in/hr, A=5 ac
- source
- Rational formula (NCEES FE Reference Handbook); exact unit conversion — Q = C·i·A with ac·in/hr → cfs carrying the 1.00833 factor
- verified by
- hand-recomputed
- derivation
- i = 2·0.0254/3600 = 1.41111e-5 m/s; A = 20234.282 m²; Q = 0.3·1.41111e-5·20234.282 = 0.0856583 m³/s = 3.025 cfs (folklore shortcut gives 3.0).
- must produce
- peakFlow = 3.025 cfscompositeC = 0.3
Counted, not claimed
These numbers are computed from the codebase when this page is built — they cannot drift from reality.
54
deterministic solvers
212
reference problems with checked answers
58
error fixtures — inputs that must refuse
9
disciplines covered
Units and conventions
- All solvers compute in SI internally; your units are converted and dimension-checked at the boundary, and answers display in SI or US customary — your choice.
- “lb” is read as pound-force in load contexts (the civil convention), and the solution says so whenever that reading was applied.
- Numbers are rounded only for display, with significant figures stated. The computation itself is full-precision floating point.
- Iterative solutions (normal depth, friction factors, rates of return) report their convergence, and warn when a result sits near a sensitive edge.
What this does not promise
Routing is AI, and can miss
Choosing which solver fits your problem is the one job the AI keeps. It can pick a lesser solver or misread a poorly stated problem — but the chosen solver is always named on the solution, its inputs are shown, and its steps are deterministic. A wrong route is visible, never a silently wrong number.
Refusal is a feature
Unsupported topic, missing data, and physically infeasible inputs are explicit outcomes with explicit messages. The system never pads a gap with a guess. Unsupported topics are logged and drive what gets built next.
Design codes have editions
Solvers that touch design standards (sight distance, soil classification) name the standard they implement. Your course may use a different edition — check the cited convention on the solution.
For studying, not for stamping
CivilSolve is a study tool. It is not a substitute for engineering judgment, review, or a licensed engineer's seal.
Who makes this
CivilSolve is a Runestone Labs product. The principle this page describes — deterministic code does the computing, the AI is confined to reading the problem and narrating the result, and an audit checks every numeral — is the same principle behind our work on governing AI agents.