dillylang-rank

name	dillylang-rank
description	Scores and orders artifacts by criteria. Binding required. Trigger= /dillylang-rank CRITERIA
argument-hint	'criterion [, criterion_1, ...]' [--absolute] [--top-k N]

Dillylang rank

[[THIS is_grounded_by: urn:unique_reference:dillylang::spec-primer]]

Score and order artifacts by named criteria. The ranking statement is required — ranking without criteria is undefined.

Rank is a selector, not a generator. Its output is a permutation of input IDs with scores, not new content.

Arguments

Required: A ranking statement — one criterion or comma-separated list.

• Single: /dillylang-rank 'security impact'
• Multi: /dillylang-rank 'security impact, effort, reversibility'

Flags:

Flag	Default	Effect
--relative	yes	Scores are high | medium | low per criterion. Allows ties.
--absolute	no	Strict ordinal (1st, 2nd, 3rd). Forces total ordering, no ties.
--top-k N	omitted	Mark top N items in output. Full ranking always shown.

Artifacts to rank come from context: preceding conversation, prior operator output, or @file references. If fewer than 2 artifacts are present, emit a diagnostic — do not produce a degenerate single-item ranking.

Prompt constraints

• Score first, order second. Score every artifact on every criterion before determining the ordering. Never decide the ordering first and rationalize scores to match — that produces confabulated rationale indistinguishable from genuine judgment.
• Quote-ground every score. Each score must point to specific content in the artifact that evidences it. "Scores high on novelty" is rejected. "Scores high on novelty — proposes a constraint-propagation approach not present in any other artifact" is accepted.
• Comparative, not absolute. Score artifacts relative to each other, not against an abstract standard. rank is not map(evaluate, artifacts). If every artifact scores identically on a criterion, that criterion contributed nothing to the ordering — say so.
• Criteria fidelity. Use the criterion the user supplied, not a proxy the model finds easier to judge. If "structural novelty" scores are justified by "provides detailed examples," that's criterion drift — the score is against a substituted criterion.
• Operationalize before scoring. Before the score matrix, restate each criterion as a concrete, falsifiable question. "Best from a sweetness perspective" → "Which delivers the highest perceived sweetness intensity per unit weight?" This makes criterion drift visible: if the operationalization doesn't match the scores, the skill drifted.

Calibration examples:

Rejected (flat): "A: medium, B: medium, C: medium. Rationale: all three artifacts address the problem adequately." (No ordering information. Uniform scores with vague rationale is the most common rank failure — it looks scored but carries no signal.)

Rejected (confabulated): "Ranking: B, A, C. B scores high on impact because it is the strongest option." (Ordering stated before scores. "Strongest option" is a restatement of the ranking, not evidence for it. Rationale was generated to justify a pre-decided order.)

Accepted: "Scores — Impact: A=high (proposes audit trail that catches drift at source), B=medium (detects drift but post-hoc), C=low (no drift mechanism). Effort: A=high (requires schema migration), B=low (config change only), C=low (no change). Ordering: B, A, C. B ranks first: medium impact at low effort dominates A's high impact at high effort. C ranks last: low on both axes."

Output template

Artifacts

Assign each input artifact a short label and one-line identifier. Use only these labels in all subsequent output.

Label	Identifier
[A]	(first line, title, or brief identifier)
[B]	...

Criterion operationalization

Restate each criterion as a concrete question before scoring.

Criterion	Operationalized as
(user's criterion)	(falsifiable question the scores will answer)

Score matrix

Mode: relative

Artifact	Criterion 1	Criterion 2	...
[A]	high / medium / low	...
[B]	...

For --absolute mode, replace high/medium/low with ordinal ranks (1st, 2nd, ...) per criterion.

Each cell must include a parenthetical evidence pointer.

Two input modes determine evidence standards:

• Artifact-inspection (input is text — documents, proposals, code): evidence must quote or point to specific content in the artifact.
• Concept-ranking (input is bare nouns or ideas, not text with quotable content): evidence cites domain knowledge. State the factual basis and its source (established science, industry convention, etc.). Do not fabricate textual evidence from artifacts that have no text.

Ranking (RK-n)

The ordering, derived from the score matrix above.

Each entry must include:

• Artifact: label
• Position: ordinal rank
• Rationale: how the scores across criteria produce this position — connect the dots, don't just restate the scores

When --top-k N is specified, mark the top N entries.

Criterion notes (CN-n)

Per-criterion observations. Required when:

• A criterion produced uniform scores across all artifacts (contributed nothing to the ordering)
• Multiple criteria produced near-identical score patterns (coupling — effectively double-counting one dimension)
• In --relative mode, more than 2 artifacts share the same score on a criterion (bucket collision — the score discriminates poorly at this granularity; note that --absolute mode would force separation)

Each entry must include:

• Criterion: which one
• Observation: what happened

n is sequential starting from 1 in each section.

Self-review

After generating the ranking, check two things:

1. Score variance. For each criterion, do the scores discriminate? If all artifacts received the same score on a criterion, that criterion did no work. If every criterion is flat, the ranking is a non-ranking — strengthen the scores or state that the artifacts are genuinely indistinguishable on these criteria.

2. Score-order consistency. Does the ordering follow from the score matrix? Read the scores and derive the ordering independently. If your derived ordering disagrees with the stated ranking, the rationale was confabulated — the ordering was decided before the scores. Fix the ordering to match the scores, or fix the scores to match genuine judgment.