Methods · Reviewed 13 July 2026
How evidence earns authority.
Integrated architecture
The fast lane proposes. The truth lane decides what the record may say.
Compute, telemetry and evidence are bound into the same governed sequence.
- 01Identity
Brief, geometry, candidate and workload digests.
- 02Fast lane
Analytical models, optimisation and advisory AI.
- 03Governor
Budget, applicability, policy and promotion ticket.
- 04Truth lane
Native solver bytes, fields, traces and convergence.
- 05Telemetry
Bounded operational observations and stage receipts.
- 06Evidence
Authority class, uncertainty, boundary and handoff.
Operating principles
Nine rules carry across the suite.
They are designed to survive a change of solver, model family or scientific domain.
Separate maturity scales
Product certification, evidence fidelity and component readiness answer different questions. Complete labels replace unexplained level numbers.
Preserve identity
OpenUSD can carry composition and provenance, while native STL, mesh and field bytes remain authoritative for each numerical engine.
Escalate deliberately
Cheap models screen. Promotion requires an explicit ticket, compute decision and retained evidence contract.
Fail exact geometry closed
Source, transformed geometry, mesh recipe, patches and fields must match before an exact result can promote.
Report numerical uncertainty
Residuals alone are insufficient. Convergence, conservation and quantity-of-interest stability remain visible.
Compare learned models
Novel families compete against strong baselines on leakage-safe holdouts, calibration, domain shift and cost.
Stress software honestly
Campaign tests prove ingestion, abstention and reproducibility—not scientific validity.
Bound telemetry
Low-cardinality labels and fixed buckets prevent observability from becoming operational noise.
Reserve physical claims
Manufacture, transport, material response and biology require retained external measurements.
Method stack
Current implementation, verification route and next gate.
Future-facing work is named as research until evidence admits it.
OpenUSD
Canonical design thread in Scaffold; narrow core-only 26.05 geometry qualification in Cell. Verify composition, declared dependencies, semantics and exact bytes.
Next: USDZ closure, subset semantics, retained real CAD/scanner cases and independent validation.
OpenFOAM
Real geometry-resolved execution, native case artifacts and three-grid Scaffold observation; controlled Nano runners.
Next: finer GCI closure, conservative scalars, wall-shear reductions and current sealed geometry authority.
FEniCSx
Exact 3D DOLFINx mechanics mechanism, retained MSH bytes, physical tags and MPI image smokes in Scaffold; handoff contracts in Cell.
Next: canonical full fields, exact-reference convergence, serial/MPI parity and cross-code evidence.
Surrogates
Corpus, training, registry, uncertainty and admission seams are implemented across products; no unqualified model inherits truth authority.
Next: real truth corpora, dual holdouts, calibrated uncertainty, OOD abstention and promotion evidence.
Fourier methods
The current Scaffold field baseline computes one global rFFT and feeds dense layers at a fixed resolution.
It is not an FNO/PINO. Next: equal-budget true 3D FNO/PINO, Geo-FNO/GINO and mesh-graph benchmarks.
Optimisation
Multi-objective search, quality diversity, Bayesian optimisation, active learning and checkpoint-aware scheduling are bound to explicit evaluators.
Optimality is always relative to the declared model, constraints and search budget.
AI wiring
Advisory models can parse, rank, route and explain while applicability gates and abstention protect the decision path.
AI has no laboratory, solver or dispatch authority by default.
Telemetry optimisation
The 1,000-case campaign retained 838 Prometheus series under a fail-closed cap of 1,024, with fixed buckets and no case IDs, seeds or digests in labels.
Next: information-per-expensive-run optimisation, stage latency budgets and live SLO evidence; current campaign is not a production SLO proof.