Evals When the World Can Break

Software evals are golden tickets and JSON. A 90% pass rate is a model card. In hardware a 90% pass rate is a recall if the 10% is a weld, a pin, a firmware register, a stop the technician did not hit. I have spent a year telling teams to stop vibe-checking Friday demos. Atoms make the lecture cheaper: the world can break, so the eval has to be allowed to fail loud before the lot does.

This is not a plea for more ML engineering theatre. It is the missing half of skills as infrastructure. A skill without an eval is a style guide. A twin without a live diff is a render. A CoE with a bake-off slide is a museum.

Start with the failure you already had

Do not invent evals in a workshop. Pull last quarter's scrap, the firmware that regressed after a "harmless" config, the drawing that could not be cut, the twin that said OK while the cell drifted. Those are golden cases. Write them as scripts a skill can fail. If you cannot name ten real failures, you are not ready for an agent on the line. You are ready for a tour.

The same authoring loop as skills: watch the agent miss, capture the miss, make the miss un-ignorable. Anthropic says start with evaluation and write skills to close gaps. Hardware just raises the cost of skipping the first half.

A ladder, not a score

One number is how you lie. Use a ladder, and do not skip rungs:

• Unit on the model, schema, pinout, forbidden phrase, cited material present. Cheap. Fast. Not sufficient.
• Simulation, FEA that agrees with the analytic check below it, CFD with a model you named, PLC tests in a harness. If the continuum result contradicts the beam result, you do not get to average them.
• Bench, hardware-in-the-loop. Bitstream on a real FPGA. Firmware on a board. Fixture under a load cell. The render is not this rung.
• Line, first-pass yield, cycle time, the exception the operator actually sees. Exit condition of a residency lives here, not in the lab video.

Promote a loop only when the rung below is green. Skipping to the line because the twin was pretty is how you buy a shutdown.

What the script must do

Put scores in scripts/. Keep prose for judgment. The script should:

• Fail the CI. Exit non-zero. A markdown report that always says WARN is a hug.
• Cite. Materials, standards, the golden lot ID. No free allowables.
• Be cheap enough to run on every change at the bottom rungs, expensive and scheduled at the top. You will be tempted to run the continuum FEA on every comma. That is how evals die.
• Name the owner. When it goes red, a human gets the page. "The agent will notice" is not an owner.

Token spend is a cost. A skill that re-reads a 400-page standard on every call is not thorough. Point at the clause. Progressive disclosure applies to evals too.

Regulators are in the path, not in the steering pack

ISO 13849, ISO 12100, AS9100, FDA, machinery directives, ITAR, the alphabet is the work. If the eval cannot produce the packet those people read, you will reconstruct it in panic after the agent has already shipped a geometry. Non-negotiables belong in the pipeline. Preferences (which fixture to print first) belong to the triad. That is the same split as governance without a monolith.

The kill switch is physical. A feature flag is not a stop. If the technician will not hit it, you do not have a kill switch. You have a story about one.

Software teams, this is your warning label too

Most readers of this site still live in workflows, not cells. The ladder still applies. A brand skill that cannot fail a forbidden-phrase test is a vibe. A research skill that cannot fail a citation check is a hallucination engine. A GTM skill that cannot fail "this claim is not in the source CSV" is how you get a legal event. Hardware just makes the metaphor rude.

After the kill switch, evals were one of the four things still on disk. If your evals lived in a vendor's eval product you cannot export, you rented those too.

Failure modes

Leaderboard cosplay. Public model ranks. Zero cases from your scrap bin.

Twin that never sees live equipment. Green forever. Drift in the cell.

Shipping the render. The screenshot made the steering pack. The toolpath did not.

Eval as a quarterly project. Models change weekly. Your suite must be cheaper to run than to argue about.

No human path. The agent failed closed in theory and open in the cell because override was tribal knowledge.

This month

Ten golden cases from real failures. Scripts for the bottom two rungs. One bench case if you have a bench. Name the owner. Wire them to the three skills you actually use. Do not wait for the platform to grow an "eval service." A folder that exits 1 is a service.

When the world can break, evals are not a research hobby. They are how a skill earns the right to touch the work. Score it, fail loud, climb the ladder, leave a stop a human will use. Everything else is a vibe check with mass.

Related: Skills Became Infrastructure, The Last Mile Was Always Atoms, and Forward Deployed on the Line.