The bar for digital product briefs

A good brief gives a product the right start. We made the standard open.

→ score a brief in the live console

Each rule is one contestable file. criteria and the examples are all the model reads; weight steers the score in code; ★ marks a rule that is also a gate requirement. The examples are also the rule's regression tests. Numbers (R01–R14) are for easy reference in discussion — link any rule directly, e.g. rules.html#anonymised; the id is the canonical name in PRs and API verdicts.

A product's outcome is mostly the team. The brief is the 20% that decides whether that team starts on solid ground — or spends its first month untangling what was meant. That untangling normally happens in workshops: expensive, slow, dependent on us being in the room. So we wrote the check down instead — grip on reality, doability, scale — and published all of it, so the bar isn't ours to gatekeep.

this page is the standard · the console scores against it · the rules live on GitHub and anyone can change them

the model judges

scope-boundariespass

conf .91 · “no native apps in v1”

success-metricsfail

conf .88 · “happy users”

One verdict per rule: status, confidence, a quote from the brief. Temperature 0.

the seam

the code decides

score.py · plain code

weights · renormalisation72/100

gate · 4 requirements4 / 4

may publishyes — with reservation

Weights, bands, gate and floor are files. Same brief in, same number out.

Not a grade — a progress bar toward publication, and the to-do list that closes the distance.

The score says how good a brief is. A separate gate says whether it may publish at all. They are different axes, so read them as a matrix — no average papers over a leaked client name.

Four hard requirements, all of which must hold:

Miss one and the brief is blocked, with the unmet ones named — even at 92/100. The line is fair because the arithmetic is reproducible from the returned verdicts — every verdict names its model and carries verbatim evidence — the breakdown shows exactly which rule flips you, and you can revise and resubmit freely. The two directory-scoped requirements exist for the blind noticeboard only: a caller scoring a generic brief deactivates that context (gate_contexts: []) and both drop out entirely — no gate, no weight, the average renormalised over the twelve rules that apply anywhere.

Fluency is free now — a brief can read beautifully and say nothing. Every rule is anchored to a source, typed by authority, pinned and human-verified in the PR:

We run the directory and the bar, and welance also pitches as a team on that directory — same fees, same anonymity, same bar. That's a conflict of interest, and open-sourcing the bar is what defuses it. Three gestures, kept separate:

the fixtures are the immune system — a corpus of labelled briefs with expected scores. quietly lower the bar to slip a weak brief through and CI rejects you; to get around it you'd have to delete fixtures, and that's a visible, suspicious diff. every verdict carries its ruleset_version, so anyone can check nobody got a different bar — and the Operator Covenant is what binds the operator itself.