MetraSlate

Standard by MetraSlate | Traceability Edition

A Defensible Record of AI-Assisted Work.

Work that a model took part in should be as traceable as work a person did: what was required, what procedure was followed, what was done, how the result was judged, and what changed afterwards. The Traceability Edition applies the platform’s record discipline to that work. It is the edition after the Public Safety Edition, and it is in design.

The chain

Five Links, One Record.

The same shape the Metrology Edition holds for a measurement, applied to AI-assisted work.

  1. Requirements

    What the work was required to produce, written down before the work, so the result has something to be judged against.

  2. Controlled Protocols

    The procedure the work followed, as a versioned record pinned to the run, so a later revision cannot change a record already made.

  3. Recorded Actions

    What was done, by whom or by what, in order, appended and never overwritten, with the model’s part attributed as the model’s.

  4. Evaluation

    The result held against its requirement with a stated decision rule. A result that cannot be assessed is recorded as such, with the reason, rather than guessed.

  5. Evidence of Change

    When a component the work relied on is found wanting, the work that relied on it is traced and declared, so a recall is a record rather than a search.

Built and planned

What the Platform Holds, and What the Edition Adds.

The foundation is in the platform today. The edition builds the model-specific parts on it.

In the platform todayBuilt
Append-only records with an audit trail under every field. Documents sealed, cryptographically signed and verifiable by whoever holds them. A versioned procedure and policy pinned to a run. A result the engine cannot assess recorded as out of verdict, with its reason. Impact traced from a component found wanting to the work that relied on it, declared as an append-only register. These are the platform’s record discipline, in the Metrology Edition today.
The edition addsSpecified
A gateway that admits AI-assisted work under a controlled protocol and records each action as it happens. A calibration record for a model’s participation in a defined task, judged against reference cases with a stated decision rule.
Designed with the people who rely on itIn design
The reference a model’s work is judged against, who holds the authority to set it, how long a calibration stands, and how repeatability is measured for work that is not a physical quantity. These are defined per task, with the people who would rely on the record.

Scope

What the Edition is For.

A record of the work, not a rating of a model.

The Traceability Edition records and judges the work a model took part in, for a defined task, against a stated reference. It does not measure a model in the abstract, and it does not rank models against each other. Its value is the same as a calibration certificate’s: someone who was not in the room can see what was required, what was done, how it was judged, and what to do if a component it relied on turns out to be wrong.

Documents

Read the Documents.

The edition document and the research paper behind it, as issued.

Standard: Traceability EditionSupplied 8 September 2026 | 10 pages | 265 KB
The Traceability Edition document: traceability for AI-assisted work through requirements, controlled protocols, recorded actions, evaluation and evidence of change, with what is built, what is specified and what is new. The edition is in design. Open the PDF.
Inference TraceabilitySupplied 8 September 2026 | 21 pages | 470 KB
The research paper: a method for making model-mediated work traceable, with the calibration record, recorded refusal and impact traversal as its instruments. Research, not a claim about a shipped product. Open the PDF.

Talk Through the Traceability Edition.

If your organization relies on work a model took part in and has to answer for it, we would like to learn what the record would have to prove.

Start a Conversation