A checkbox list gets ticked without anybody looking at the image. This gate has three layers that each produce evidence: a deterministic script, a structured vision pass, and a scored rubric with a shipping threshold.
Rendered-image ship condition: score ≥ 16/20, zero hard fails and applicable technical checks passed. For the script's supported photographic layouts, require its PASS. For other intentional layouts, document which heuristic is inapplicable and verify the corresponding property by inspection or an appropriate measurement.
Apply image review to every rendered image. For prompt-only requests, use prompt craft; never invent an image score.
R42 interpretation: product means advertised subject; realism means medium coherence for graphics; brand attribution may be approved text when no official logo is supplied. Subline and button-shaped CTA are optional if message and action are already clear. The script is a conservative photographic-layout heuristic: inspect flagged seams and flat graphic regions before rejecting an intentional poster. Document inapplicable checks rather than silently calling them PASS.
⚠️ A PASS without--text-boxis a partial PASS. Four of the seven checks measure the copy block; with no box declared they reportn/aand the script says so in its output. On any creative that carries text, pass--text-box(and--logo-boxwhen a logo is placed) or the gate is checking dimensions and seams only.
Things a machine measures better than an eye. No model call, no cost.
Artifacting is deliberately not here. Sensor grain, fabric weave and Voronoi webbing are the same high-frequency energy, so any threshold strict enough to catch webbing rejects legitimate photography. It is called by eye in layer 2, at 100% (
R40).
python scripts/qa.py out/ad_01.png --format 4:5 --text-box 86,900,994,1264
python scripts/qa.py out/ad_01.png --format 4:5 --text-box 86,900,994,1200 \
--logo-box 780,1210,994,1264
python scripts/qa.py out/*.png --format 4:5 --contact-sheet out/_contact.pngThe script's own behaviour is covered by python scripts/test_qa.py — thirteen
synthetic cases asserting that the canonical layouts pass and the known failure
modes fail. Run it after touching qa.py.
| Check | Rule | Needs | Fail condition |
|---|---|---|---|
dimensions |
R19 | --format |
not the declared format's exact resolution |
safe_area |
R08 | --text-box / --logo-box |
a declared box crosses the 8% margin, or the 9:16 chrome zones |
margin_activity |
R08 | — | advisory, never fails. Edge energy in the margin band. A full-bleed photo scores as high as a headline that crosses the margin, so this number tells you to look — safe_area is the verdict |
contrast |
layout §4a | --text-box |
< 4.5:1 between the glyph population and the backdrop median inside the box |
collage |
R06 | — | a seam runs the full length of the canvas → the model produced a grid |
thumbnail |
R33 | --text-box |
the copy block loses > 60% of its edge energy at 150px |
focal_regions |
R07 | — | saliency spread over > 2 strong regions |
scrim_uniformity |
layout §3b | --text-box |
< 85% of the backdrop under the copy sits in one tight luminance band (a busy photo under the text) |
Output is JSON, so it drops straight into a report:
{"file":"ad_01.png","verdict":"PASS","checks":{"dimensions":"1080x1350 ok",
"safe_area":"clear (margin 86px, keep-out 86-1264px)",
"margin_activity":"0.04% (quiet)","contrast":6.8,"collage":false,
"thumbnail":0.81,"focal_regions":1,"scrim_uniformity":0.94},"failed":[]}Send the image to a vision model with this exact prompt. Structured output is the point: it forces the model to look rather than agree.
You are the QA art director for a paid advertising campaign. Inspect the attached
ad image against the rules below. Be adversarial — your job is to find reasons to
reject it, not to be encouraging. If you cannot see something clearly, say so
rather than assuming it is fine.
Inspect the image at FULL RESOLUTION, not scaled down. Texture artifacting is
invisible at thumbnail size and obvious at 100% — zoom into every fine-texture
surface (fur, knit, foliage, crumb, stone, hair) and every flat one (walls, skies,
panels) before you answer.
Reference images supplied with this brief: [list them, or "none"]
Declared copy that should appear on the image:
brand: "<BRAND>"
headline: "<HEADLINE>"
subline: "<SUBLINE>"
cta: "<CTA>"
Return ONLY this JSON, no prose:
{
"transcribed_text": ["every string of text you can read on the image, verbatim,
including partial or truncated words"],
"spelling_errors": [{"found": "", "expected": "", "severity": "hard|soft"}],
"text_matches_declared_copy": true|false,
"truncated_or_clipped_text": true|false,
"focal_points": <integer: how many elements compete for first attention>,
"product_matches_reference": true|false|"no_reference",
"product_changes_observed": ["shape/color/logo/lettering/mechanism changes"],
"logo_present": true|false,
"logo_appears_redrawn": true|false|"unsure",
"invented_food_or_props": true|false,
"accent_color_placements": <integer>,
"text_sits_on_busy_area": true|false,
"anatomy_errors": ["deformed hands, extra fingers, impossible limbs"],
"physics_errors": ["floating objects, wrong shadows, missing contact"],
"texture_artifacts": ["surfaces showing cellular/Voronoi cells, webbing,
netting, noise clusters, speckle on flat areas, or a
tiled/repeating texture fill — name the surface; [] if none"],
"unbriefed_elements": ["any shape, object or colour cast present in the image
with no line in the prompt behind it — session ghosting"],
"style_coherence": true|false,
"visual_language": "the ONE language this reads as (oversized type, colour-block
still life, direct-flash editorial, documentary, native
interface, performance sticker, tactile zine...) or \"mixed\"",
"fake_functional_ui": ["play buttons on a still, close icons, notification
badges, tappable-looking checkboxes; [] if none"],
"edge_halo_or_bleed": ["objects ringed by a glow, seam or smear, or texture
leaking across their boundary; [] if none"],
"duplicated_or_warped": ["duplicated objects, melted geometry, impossible
reflections; [] if none"],
"unrequested_changes": ["ON AN EDIT ONLY: anything that differs from the
reference beyond the one change that was requested —
room, layout, furniture, face, framing, materials"],
"slop_tells_present": ["from: neon, glow, glassmorphism, gradient text,
floating particles, isometric, HUD, clip-art icons,
fake UI, excessive bokeh, plastic surfaces"],
"ad_spine": {"headline": true|false, "subline": true|false,
"cta": true|false, "brand_cue": true|false},
"readable_at_thumbnail": true|false,
"reads_as_ai_generated": true|false,
"one_sentence_verdict": "",
"hard_fails": ["R30-product|R30-logo|R30-text|R30-anatomy|R30-physics|
R30-scale|R30-chaos|R30-ui|R30-background|R30-stock|
R30-function|R30-overload|R40-texture|R40-ghosting|
R41-drift|R41-halo|R48-fake-ui"],
"score": {"hierarchy":0-2, "product":0-2, "realism":0-2, "typography":0-2,
"copy":0-2, "color":0-2, "space":0-2, "logo":0-2,
"thumbnail":0-2, "idea":0-2},
"total": <sum, out of 20>
}
Reading the result:
- Any entry in
hard_fails→ regenerate. Do not retouch. texture_artifactsnon-empty →R40-texture. Bound the detail and regenerate (artifact-control.md§3); repair locally only for a single small anomaly, never across a surface.unbriefed_elementsnon-empty →R40-ghosting. The prompt is not the bug — you generated this in a session that had already produced an image. Regenerate in a fresh one (artifact-control.md§5).unrequested_changesnon-empty on an edit →R41-drift. ThePRESERVEslot was missing or too thin. Re-edit with the preservation block for that subject; do not accept the drift because the result looks good (artifact-control.md§8).edge_halo_or_bleedorduplicated_or_warpednon-empty →R41-halo. The edit was too large to integrate. Shrink the ask, hold the original lighting, and specify contact shadows (artifact-control.md§7).style_coherence: false→ two colliding style descriptors. Pick one and regenerate; re-rolling will not resolve it (artifact-control.md§4).visual_language: "mixed"→ R45. Two dialects compete (for example direct flash plus riso grain plus stickers). Keep the language that serves the message, remove the other's devices, and regenerate (style-atlas-2026.md§5).fake_functional_uinon-empty →R48-fake-ui. Remove the control; a still image must not pretend to be tappable (platform-compliance.md§6).spelling_errorswithseverity: hard→ regenerate, or run a targeted model edit on the misspelled word (R50).total < 16→ fix the lowest-scoring criteria and re-run.reads_as_ai_generated: truewithtotal ≥ 16→ trust the flag, not the score. Redesign.
Ten criteria, 0/1/2 each. This is what score in the vision JSON refers to, and what a human uses when reviewing by eye.
| # | Criterion | 0 — fail | 1 — passable | 2 — good |
|---|---|---|---|---|
| 1 | Hierarchy (R07) | two or more elements compete; eye doesn't land | one focal point but weak separation | one obvious focal point in <1s, clear 2nd and 3rd level |
| 2 | Product / subject (R02) | small, obscured, or ambiguous | visible but not dominant | large, lit, sharper than surroundings, attractive angle |
| 3 | Realism (R04, R40) | wrong physics, deformed anatomy, fake materials, or visible texture artifacting | mostly plausible, one soft tell | reads as a real photograph; correct light, shadows, contact; every surface clean at 100% |
| 4 | Typography (R17, layout §2) | >3 sizes or >2 families; default font; mid-word clipping | correct sizes, unremarkable pairing | named pairing, correct tracking, hierarchy ratio ≥3× |
| 5 | Copy (headline-system) | interchangeable headline, banned words, misspelling | specific but long or flat | passes the specificity test; within budget; spine reads as one voice |
| 6 | Color (layout §4) | default gradient, >3 accent placements, contrast <4.5:1 | brand palette, contrast ok | brand palette, one accent used ≤3×, deliberate contrast |
| 7 | Space (R08) | frame full, elements touching edges | margins present but tight | 8% margins respected, deliberate negative space |
| 8 | Logo (R30-logo) | AI-redrawn, distorted, or a text substitute | official file but cramped or low contrast | official file, correct clear space, reads on its backdrop |
| 9 | Thumbnail (R33) | headline unreadable at 150px | headline readable, detail lost | full message survives at 150px |
| 10 | The idea (R01, R06) | pretty picture, no message | a message, weakly expressed | one idea, expressed visually, would work without the copy |
Thresholds
- ≥ 18 — ship.
- 16–17 — ship if the deductions are on criteria 7–9 (fixable in one more generation or targeted edit); otherwise fix.
- 12–15 — fix and re-score. Usually typography or copy.
- < 12 — regenerate from a new prompt. Don't patch.
- Any hard fail at any score — regenerate.
- Build a contact sheet from the batch (exclude previous contact sheets from the glob — a verified footgun).
- Score every image individually; a batch verdict hides the weak ones.
- Check series consistency (R20): the product/venue must be identical across the set. Only context, frame, mood and light change.
- Check angle diversity (R21): count distinct headline archetypes and creative types.
5 creatives, 5 archetypesis the target;5 creatives, 1 archetype, 5 palettesis a failed batch. - Report per image: score, verdict, failed rule IDs, and what changed on the second pass.
Report format:
ad_01 18/20 PASS —
ad_02 14/20 FIX typography (mid-word clip, 88px→64px), copy (generic headline)
ad_03 9/20 REDO R30-logo (redrawn crest), R30-stock
ad_04 17/20 PASS space (subline 40px from edge — padded)
ad_05 16/20 PASS —
batch: 3 ship, 1 fixed, 1 regenerated · archetypes used: concrete, place, number, proof, deadline
For the written deliverables around the image. This does not check images — see layer 1 for those.
Prompts: run the linter before every generation. It fails on unresolved placeholders, quoted copy without "no other text", copy slop (R52) and fake functional UI (R48-fake-ui), and warns on long headlines, density adjectives (R40), decayed trend cues (R53), visual slop vocabulary (R05), evidence words (R55) and a missing aspect ratio. Negated mentions ("no neon glow") pass.
python scripts/prompt_lint.py prompts.txt # prompts separated by a line with ---
python scripts/prompt_lint.py prompts.txt --strict # WARN also failsThe greps below remain a quick check for captions and prose:
# Visual-slop vocabulary leaking into prompts
grep -Eiq "purple gradient|glassmorphism|neon glow|glowing orbs|floating particles|isometric" "$f" \
&& echo "SLOP" || echo "CLEAN"
# AI copy tells in captions/prose
grep -Eiq "delve|seamless|empower|elevate|robust|tapestry|game-changer|revolutionary|unlock|unleash|🚀" "$f" \
&& echo "SLOP" || echo "CLEAN"
# Placeholders that survived into a final prompt (R25)
grep -En "\[[A-Z_ ]+\]|<BRAND>|HEADLINE\"" "$f" && echo "UNFINISHED PROMPT"
⚠️ False positives: CSS variables and class names (--purple,.card-lp) are legitimate. Check context.⚠️ Meta note: don't quote banned words in a file's own QA section — the grep will catch them. Write "stop-slop gate — PASS" instead.
gate script PASS + vision JSON + score ≥16/20 + zero hard fails
layers 1. scripts/qa.py — dimensions, safe area, contrast, collage, thumbnail
(always pass --text-box, or four of seven checks report n/a)
2. vision prompt (§2) — structured JSON, adversarial, transcribes all text
3. rubric (§3) — 10 criteria × 0/1/2
redo any hard fail · score <12 · reads_as_ai_generated
fix score 12–17 → lowest criteria first, re-score
batch contact sheet · per-image scores · series consistency · archetype diversity