Platform · Verification
Verification agents
A simulation result is only as trustworthy as the study that proves it stopped changing. Verification agents re-solve every number under mesh refinement, sample it repeatedly, and cross-check it against independent methods before anyone builds a decision on it.
01The problem
In 1991 the concrete base of the Sleipner A offshore platform sank in a Norwegian fjord during a controlled ballasting test. The cause was not the structure but the computation: a finite-element mesh too coarse to resolve the shear in its cell walls underestimated the critical stresses by roughly 45 to 47 percent, so a wall was built too thin. Published investigations put the loss near 700 million dollars. A single unverified, under-refined solve, trusted as if it had converged, was enough to sink it, which is why no serious team builds a decision on an FEA or CFD number before it has passed a mesh-refinement and convergence study.
02How we solve it
Convergence discipline, run as an agent.
Agentriq packages the convergence and validation playbook that simulation teams already trust into an agent that re-solves, samples, and cross-checks every result before it is reported.
Self-consistency across samples
Draws multiple independent solves of the same problem and treats agreement, not a single lucky run, as the signal that a result is real.
Mesh independence and grid convergence
Re-runs on 3 to 5 progressively refined meshes and computes a Grid Convergence Index (Roache) with Richardson extrapolation to prove the answer sits in the asymptotic range.
Constraint and physical-sanity checks
Asserts hard limits, conservation laws, and sign and bound conditions so any result that violates physics is caught before it propagates.
Cross-validation against methods
Compares each result against a surrogate model and an analytical reduced-order model, flagging any number the independent estimates will not corroborate.
03What it produces
Scatter that collapses to counts.
The same convergence discipline that narrowed code-to-code drag scatter on a shared transonic wing from more than a hundred counts to roughly five turns a raw solver output into a number with a stated error band and a pass or fail verdict.
04The agentic loop
A closed loop around every result.
Agentriq runs verification as a closed agent loop: each candidate result is re-solved, stress-tested against constraints and surrogates, root-caused when it disagrees, and cleared by a judge before it is written back as trusted.
- 01Take in a candidate result together with its solver model, mesh, and boundary conditions.
- 02Verify it: re-solve on 3 to 5 refined meshes, draw multiple samples, compute a Grid Convergence Index, and check hard constraints and surrogate cross-agreement.
- 03When samples disagree or the index misses the asymptotic band, root-cause the discrepancy with Agentic RAG over solver logs, V&V standards, and prior campaigns.
- 04A judge agent weighs the evidence and rules the number verified, flagged, or rejected against ASME VVUQ 20 style acceptance criteria.
- 05Write the verified value, its convergence band, and full provenance back to the shared record so downstream agents can trust it.
05Tooling
Simulation, solvers & frameworks.
Get started