@mrkt_frwd/award

A calibrated craft critic that returns "unscored" rather than a number it cannot justify.

npx @mrkt_frwd/award axes      # the rubric
npx @mrkt_frwd/award prompt    # the scorecard to hand a judge
npx @mrkt_frwd/award score axes.json

What it is for

Scoring design work with a model in the loop, without the score quietly becoming meaningless. Award does not look at your page. It holds the rubric and the arithmetic; the seeing comes from a judge you supply. Separating those two halves is the entire design — the arithmetic is deterministic and auditable, the seeing is not, and pretending otherwise is how a scoring system learns to certify its own output.

The claim, and the evidence for it

The same scorecard, where the judge could not evidence two of five axes:

| | compositionalTension | chapterDifferentiation | result | |---|---|---|---| | a critic that zero-fills | 0 | 0 | 49 | | Award | dropped, weights renormalised | dropped | 77 |

Twenty-eight points, and not one of them is about the page. Abstention is required, not optional: if the capture carries no evidence for an axis, the judge emits skip and says why. It never guesses a middle score.

And when nothing at all is scorable:

unscored — no axis was scorable. That is an answer, not a zero.

The doctrine

Seven rules. Each one is a bug that shipped here first, and each has a test that fails if the guard is removed.

1. Never tune a threshold to make a run pass. Fix the work. Gates may be optimised; quality scores may not — a quality scalar under optimisation pressure gets gamed, and this project has three recorded instances.
2. Tests must recompute. Never assert over a checked-in result file, or both sides of the comparison come from the same file.
3. The gold set must include work you did not produce. Without external anchors, separation only proves the critic agrees with your own ranking — it cannot detect that the whole set sits too low.
4. Calibration may never lower a bar. It may only move the vision blend weight.
5. Calibrate against the medium you are judging. Page anchors scored with a product-photography rubric inverted the ranking. Objects judged against screenshots of web pages produced specific, correct findings and no stable bar.
6. An abstained axis is dropped, never scored zero.
7. A set containing only your own work can never report ready for training — enforced in code, not convention.

Full text: npx @mrkt_frwd/award doctrine

Commands

award axes | prompt | score <axes.json> | verdict <verdict.json>
award brief | divergence <frames…> --against <anchors…> | providers | doctrine

divergence is the guard against reproducing a reference you were calibrating against — it compares captured frames to anchor stills and refuses a composition clone.

Requirements

Node 18+. No dependencies.

MIT © Joe Asare. Built at Joe Asare Studio, after the scoring system there learned to certify its own output.