The method diversity-qa follows when judging whether a pack "actually produces results different from the baseline."
Boundaries of this skill (read first)
This skill does not judge source fidelity or craft.
- Judgment against the original reference →
fidelity-scoring - Render quality, alignment, contrast, spec →
craft-scoring - This skill → Differentiation from baseline + inter-pack duplication만
The three scores are not summed. High diversity cannot offset subpar fidelity, and neither can the reverse. Which gate is needed and the status derivation rules are the SSOT in the state-expression clause of design-pack-schema— do not duplicate them here.
The 5 axes used here (color/type/layout/space/shape·motion) check consistency with the meta.yaml이 claimed style. Since the person who wrote the claim is the same person who wrote the pack, these axes alone cannot judge "does it resemble the original."
Judgment vocabulary
unreviewed·unverifiable·evidence_missing은pass. If you have not directly seen the required images,needs_review다.
Quantitative metrics assistqualitative judgment; they do not replacequalitative judgment that was not performed. Insufficient time or image-viewing quota is not a reason to pass. In fact, _workspace/13_qa_premium.jsonused an axis not in this rubric (baseline_score·detail_pages_score) and passed all 20 packs while stating that some packs were not directly seen. That is not verification.
reference_modethe meaning of rejection differs depending on
The absence of this clause actually caused an incident. diversity-qa rejected the exact_document pack (HD Hyundai) and instructed **"put colors not in the original into the body,"** which directly conflicted with that deck's must_preserve signature (role separation: graphics=green / data=navy). The curator was right to refuse and escalate.
| mode | Meaning of a low diversity score | Correct action |
|---|---|---|
style_synthesis |
The pack pushed the style less hard | Rewrite — make the axes clearer |
exact_document · official_system · historical_canon |
The original itself is similar to the baseline | Replace the candidate or judge at the catalog dimension |
Do not instruct evidence-track packs to "add things not in the original to raise the score." That is an instruction that would be rejected under fidelity F3 (evidence precision), and it repeats in the opposite direction the failure this project experienced in v1–v3 ("it differs from the original").
If an evidence-track pack falls short at this gate, diversity-qa should not instruct a rewrite but escalate to pack-architectThe judgment is one of three — (a) replace the candidate, (b) pass it if it is sufficiently distinct from other packs at the catalog level, (c) keep the pack but state that similarity explicitly in the catalog.
Also, in the evidence track, the distance between packs matters more than the distance from the baseline. If the original has a common corporate IR tone, resembling the baseline is normal, and the question to ask then is "is it distinct from other packs in the catalog?"
Judgment vocabulary — enum
pass · needs_review · reject use only three. fail·escalate·draftis not used.
Key: catch both kinds of failure
- Differentiation failure — the pack-applied version is too similar to the baseline. The pack did not push the style hard enough.
- Identity failure — it differs from the baseline, but it also differs from the style the pack claims (
meta.yaml의family·axes). It is different in the wrong way.
To pass, it must be sufficiently different from the baseline and at the same time match the claimed style.
5-axis scoring rubric
baseline.png와 팩 preview.png를 Read together, and score each axis from 0 to 4.
| 축 | What is examined | 0점 | 4점 |
|---|---|---|---|
| color | Palette, contrast, emphasis handling | Virtually identical to the baseline | Clearly different in line with the claimed style |
| type | Font character, hierarchy, scale | Identical | Clearly different + matches the claim |
| layout | Grid, alignment, composition | Identical | Clearly different + matches the claim |
| space | Density, whitespace rhythm | Identical | Clearly different + matches the claim |
| shape/motion | Borders, corners, shadows, texture (for web, motion cues) | Identical | Clearly different + matches the claim |
Judgment: total score (0–20). pass = total score 13 or higher and no axis scored 0. Otherwise reject. If any axis is scored 0, whether for differentiation or identity, reject regardless of total score.
Quantitative auxiliary metrics
They reinforce the qualitative judgment but do not replace it. scripts/If a helper is available in , use it; if not, proceed with the qualitative judgment alone and indicate quantitative: unavailablein the scorecard.
- Perceptual hash distance — if it is too close to the baseline, differentiation is suspect.
- Color histogram difference — quantifies the change in palette distribution.
- Edge/ink density — a proxy for differences in layout density. Use it only as a flag for qualitative re-review when values are extreme (nearly identical/completely unrelated).
Rejection targets the axis precisely
Write it in a form that can be rewritten, not as "it's bad":
reject— color 3 / type 3 / layout 0 / space 2 / shape 1. Layout axis failure: the grid and alignment are identical to the baseline. The "layout" section of prompt.md needs to specify an asymmetric grid in absolute values. Shape axis weak: the border and shadow instructions are ambiguous.
The reason must include which prompt.md section should be fixed and how, so the responsible curator can work on it immediately.
Incremental verification
Verify immediately right after rendering each pack (or batch of 5 packs), not once after full completion — so the curator can rewrite quickly.
Inter-pack similarity matrix
In addition to individual packs differing from the baseline, adopted packs must also differ sufficiently from each other for catalog diversity to hold. At the end of full verification, create a 5-axis similarity matrix between packs per track and save it to _workspace/04_qa_diversity-matrix.md. If two packs are identical on 4 or more of the 5 axes, mark them as twins and recommend integration/redefinition to pack-architect.
Deliverables
_workspace/04_qa_scorecard.json— per slug{axis_scores, total, verdict, reasons[]}._workspace/04_qa_diversity-matrix.md— inter-pack similarity matrix + twin recommendations.
Escalation
If the same pack still falls short after 2 rejections, passis not forced. If the gate is still not met after 2 times, maintain rejectand exclude from production. If render evidence cannot be produced, record as needs_review + reason_code: render_failedand also exclude. Non-passing packs remain only in the internal review list. The status enum is only these three: pass · needs_review · reject — v2's escalate·drafthas been discarded.