Work you can prove.
Producing work became cheap. Accounting for it became expensive. An agent can now produce a plan, a report, a change, a document — and nothing about the output tells you what it was measured against, who checked it, or who accepted it.
This is a published standard for evaluated work: declare the bar before the work, evaluate it independently, retain the record with the artifact. It is a specification, not a product, and it is adoptable without us.
Clarify good. Prove done.
The problem, in four lines
- 01A document of principles. It says what good practice is. It cannot tell you whether any particular piece of work met it.
- 02A reviewer nobody can vouch for. The review happened. There is no record of what it was measured against, or of who signed it.
- 03Output that arrives already finished. Nothing in it distinguishes a judgement from a guess, or a checked claim from an unchecked one.
- 04A gate that has never said no. A gate that cannot fail is not a gate. It is a formality with a green tick.
The four moves
Declare
The bar is written down — dimensions, anchors, thresholds — before anything is produced, and it carries the name of the buyer and of the specialist who approved it. No provenance, no authority.
Produce
Work happens against the declared bar. The instrument that produces it is pinned to a version, so a later reader knows exactly what rulebook was in force.
Evaluate independently
A party other than the producer reads the work against the declared bar. A control runs in the same batch, because the same input scored twice does not score the same.
Retain the record
The record is kept with the artifact: the bar, the evidence the work produced, the independent reading, the violations, and the name of the person who accepted it. It outlives the tool that made it.
The refusal
When a judgement cannot be grounded in independent evidence, this standard does not produce a number. It returns unobserved.
That is not a shortcoming we are working around. A score with nothing behind it is the exact failure this standard exists to prevent, so the instrument declines rather than guess. It is the page we would put first if honesty were a marketing strategy — it is not, so here it is, three quarters down, where a sceptic arrives.
The honest state
This is early, and it is software-first. One conforming implementation exists, exercised on a small number of real jobs; the evidence base is narrow and we publish its limits rather than implying they do not exist. Several quantities a reader will reasonably ask for are not yet recorded at all.
Two things are true at the same time: the specification is versioned, citable by commit, and independent of the implementation — and the implementation is honestly young. The problem it addresses, however, is not speculative.