The Bottleneck Moved: An Acceptance Protocol for Machine-Drafted Solver Decks

For two decades the cost structure of simulation setup was stable: a competent analyst spent most of a project's front end translating intent into syntax. Deck authoring, contact definition, solver setting selection, and the endless reconciliation of units were the tax on getting to a first run. That tax is collapsing. Assistants now emit complete, syntactically valid, superficially reasonable setup files in the time it takes to describe the problem in a paragraph.

The economics of that shift are frequently misread. If drafting falls from two days to four minutes but checking still takes two days, nothing has been saved — the schedule simply moved. Worse, the reviewers who formerly caught errors by watching a deck being written now receive a finished artifact with no visible reasoning and no audit trail of the choices inside it. The bottleneck did not disappear. It relocated to verification, and verification is where engineering time should be budgeted in 2026.

What the assistants do well, and where they fail

Machine-drafted decks are strongest where the task is conventional: an isotropic linear static case, a standard bolted interface, a familiar solver's default convergence strategy. They are weakest exactly where the model is unusual — and unusual is where most consequential simulation lives.

Recurring failure modes cluster in a few areas. Unit systems are a persistent hazard, particularly in US customary work where density arrives as weight density, mass density, or slinch, and where the assistant's confidence is unaffected by the ambiguity. Coordinate frames and sign conventions drift, especially for gravity, rotation, and pressure acting on surfaces whose normals were never checked. Contact defaults are accepted rather than chosen — penalty stiffness, friction coefficients, and interface treatment appear with numbers that look deliberate and are not. Solver settings inherit from generic examples: time-step sizes, hourglass controls, and stabilization parameters calibrated for a different problem class. And initial conditions are frequently absent altogether.

The acceptance protocol

The protocol below is deliberately procedural. It assumes the machine may propose and the engineer must ratify, and it requires evidence at each gate rather than a signature.

Gate 1: Provenance and diff

Every machine-drafted deck is compared line by line against a validated template or a prior human-authored case for the same model class. The diff is retained. Any line the assistant added, removed, or altered is listed, and each entry is dispositioned as accepted, corrected, or rejected with a reason. An undocumented deck is not a draft; it is an incident.

Gate 2: Unit and dimension sweep

The same problem is run in a second consistent unit system — SI against US customary, or a scaled variant with lengths multiplied by 25.4. Physically equivalent inputs must produce equivalent outputs to within numerical tolerance. Sweeps catch the density trap, the psi versus MPa legend, and the mixed-frame assembly that no one noticed. A deck that fails a unit sweep fails for reasons that will not be visible in the contour plot.

Gate 3: Symmetry and invariance checks

If the geometry and loading are symmetric, the response must be. Impose modeling symmetry where the physics permits and compare against the full model. Similarly, a rigid rotation of the entire model should rotate the results and change nothing else; a translation should change nothing at all. These are cheap runs that expose frame errors and misapplied constraints with unusual clarity.

Gate 4: Limit cases

Zero load must yield zero response. A rigid-body translation must yield zero stress. A uniform pressure on a fully constrained body must return the hand-calculable reaction. A linear elastic case small enough to compare against a closed-form solution should be included by default. Limit cases cost minutes and catch sign errors, double-counted loads, and contact definitions that engage before they should.

Gate 5: Equilibrium and reaction audit

Sum the reaction forces and moments and compare against applied loads. Report the residual as a relative error, not as a pass or fail adjective. This single check catches more machine-authored errors than any other, because load application errors and missing boundary conditions both show up as imbalance rather than as a suspicious contour.

Gate 6: Known-answer benchmarks

The model class is exercised against a published benchmark — a NAFEMS verification case, a Roark beam, a manufactured solution — drawn from an approved list maintained by the group. The point is not to prove the solver works; the vendor has done that. The point is to prove that this deck, produced by this workflow, reproduces a known answer.

Gate 7: Human ratification log

The engineer of record signs a short record: which gates ran, what they produced, which diff entries were dispositioned, and which assumptions were accepted without test. That log is the deliverable. In a dispute, it is worth more than the deck itself.

Where the hours go

The temptation is to treat assistant output as a productivity gain and reduce headcount per project. The more defensible response is to reallocate: accept the drafting savings and spend them on verification depth, benchmark maintenance, and the assumption register that makes a model reusable. Organizations that do this will ship faster and defend better. Organizations that bank the savings without the checks will discover the cost at the worst possible moment, in a review where the question is not whether the deck ran, but why anyone believed it.

Generation is now cheap. Judgment is not. Budget accordingly.