TRACE framework

Audit any tool that gives you a material property

Nineteen checks across three levels — Traceable, Auditable, Reproducible. It takes about twenty minutes, needs no cooperation from the vendor, and works on anything that hands you a number: a selection platform, a property database, a search engine, a language model, a colleague, or your own internal spreadsheet. Score each check pass, partial or fail. The tool’s level is the highest one whose checks all pass.

Part of TRACE, CC BY 4.0. Run it on us too — our score.


Start with the thirty-second test

If you only have time for one check, do this one.

Ask the tool for a specific property of a specific material. Take the number it gives you. Try to reach the exact source of that exact number in under two minutes.

If you cannot, the value is E0 — whatever the interface looked like — and it does not belong in your documentation.

This is not a test that language models fail by construction. A model that hands you a citation you can open passes it. It is also not a test only AI tools take: run it on a search engine, on a handbook PDF, on a colleague’s recollection. E3 is entirely serviceable for screening. The point is knowing which level you are standing on when you sign.


The three levels

Level 1 — Traceable. Every number presented carries a resolvable source, a retrieval date and an evidence level. No number is shown without them.

Level 2 — Auditable. Level 1, plus the tool exports a record rather than a transcript, records what it rejected and why, enforces the evidence rules rather than suggesting them, and states its own limitations.

Level 3 — Reproducible. Level 2, plus sources are date-pinned, an identical re-run produces an identical record, and tool and model versions are recorded.


How to score it

Do not average. A tool that passes seven Level 1 checks and fails one is not “88% Level 1”. It is not Level 1. Evidence does not average.

Score the default path. If provenance exists but sits behind a setting nobody turns on, it fails. The question is what an engineer in a hurry gets, not what the tool can do when studied.

Free tools are in scope, and it is worth scoring one at least once. The point is not to disqualify a search engine — it is to see, in writing, what level your habitual sources actually sit at.


Run it

[interactive checklist component — nineteen checks, live scoring, export]


What to do with the result

If your main tool scores Level 1 or better, you are in reasonable shape. Record which level in your selection notes; the next reviewer will want to know.

If it scores below Level 1, that is not a reason to stop using it. Most engineering runs on E3 sources and always has. It is a reason to stop treating its output as documentation, and to write down where each number came from yourself.

If a value governs a design allowable, the level matters more than anywhere else. Below E5, it needs your explicit, recorded acceptance rather than an invisible default.


What this checklist does not do

  • It does not measure accuracy. A tool can score Level 3 and be wrong. Traceability and correctness are different properties, and this measures the first.
  • It does not compare feature sets. Nothing here scores usability, coverage, price or support, all of which matter and none of which are what this instrument is for.
  • It does not account for coverage. A tool with perfect provenance and no data for your material scores well and helps you not at all. Check coverage separately — ours is here.
  • It is not neutral about what matters. It is built by people who think provenance is the thing worth measuring. That is a position, not a fact, and you should weigh it as one.