Arrow draws inside a square and always will: one art-tool decision framework from two bake-offs
The question
Verbatim: "Is Quiver's ~62% page-height composition coverage (vs. Grok's full-square output) a fixable prompt default or a hard model limitation, and does the answer reconcile with the color-rendering gap found in the earlier 2026-09-19 Quiver-vs-Grok icon bakeoff into one reusable RDCO art-tool decision framework?"
Load-bearing for the coloring-books December launch. Two one-off Quiver-vs-Grok comparisons (icons 2026-09-19, composition 2026-10-06) have never been reconciled into a standing framework.
What we already know (from the vault)
- The 2026-10-06 bake-off ([[01-projects/coloring-books/2026-10-06-bakeoff/README]]) picked Quiver Arrow 2 direct SVG over Grok-then-vectorize on line weight, retry count and cost per page ($0.27 vs $0.36), and closed with a remediation: "Before production: write prompts for a tall composition." That instruction is the thing this brief tests.
- Its implementation notes already recorded the mechanism: "Arrow (both generation and vectorization) draws in an internal 80x80 square, letterboxed in the requested 765x990 viewBox, so the art occupies a band."
fit.pywas written to re-crop the viewBox to the inked bounding box as a workaround. - The 2026-09-19 icon bake-off ([[2026-09-19-icon-set-bakeoff-quiver-vs-grok]]) found the opposite kind of result: prompt language does control content. Appending "as few lines as possible, no interior detail, something a 4-year-old could copy" visibly simplified all three weak icons and cut output tokens about 25%. That became a [[DESIGN-scribble-works]] contract rule.
- The actual gap in that 09-19 bake-off is lettering, not colour: Grok refused letters because the evaluator's
textrule fails any glyph, so BOO! and RIP came back blank, while Arrow lettered them correctly. The colour evidence lives elsewhere, in the sticker and Denominator work (below). The backlog question's premise is slightly mis-cited. - The colour gap is documented in
01-projects/personal-brand/assets/stickers/notes.md: "Arrow sometimes adds off-palette colors such as orange or gray. The palette snap handles most of them. Check the result by eye." The pipeline therefore carries a deterministic post-process that "snaps every hex color to the 5-swatch palette" and strips an Arrow-emitted semi-transparent full-canvas background rect. [[DESIGN-denominator]] adds that Quiver die-cuts print in their original colours with no tint or duotone allowed. - The same sticker notes already carry a folk-remedy for the canvas issue: "Use viewBox 1200x800 for landscape and 1000x1000 for square."
The probe (new, 2026-10-18)
Two Arrow 2 calls via the REST endpoint (POST /v1/svgs/generations), identical prompt, single variable: the requested attributes.viewBox. The prompt was the 2026-10-06 bake-off STYLE block plus the most aggressive full-bleed language I could write ("the canvas is a TALL PORTRAIT RECTANGLE, NOT a square... ink must reach within 2% of the very top edge and within 2% of the very bottom edge... Do NOT letterbox. Do NOT draw inside a centred square"). Subject held constant as bake-off page 5, Quiver's cleanest page. Ink measured as the bounding box of everything darker than 90% grey, rendered by headless Chrome at the canvas's own size. Probe cost $0.56.
Note: the MCP send_prompt tool exposes no viewBox parameter, so the decisive variable is reachable only on the REST path, which is what the production pipeline uses anyway.
| Arm | viewBox | Prompt | Transform Arrow emitted | Ink height, % of canvas |
|---|---|---|---|---|
| Archived bake-off Q pages (n=5 used) | 765x990 | bake-off STYLE, already said "fills the page, nothing cut off at the edges" | matrix(9.562 0 0 9.562 0 112.5) |
71.4 - 75.8, mean 74.2 |
| Probe A (new) | 765x990 | + maximal tall / full-bleed / anti-letterbox language | matrix(9.562 0 0 9.562 0 112.5) — byte-identical |
74.2 |
| Probe B (new) | 900x900 | same maximal language | scale(11.25) |
99.1 |
The arithmetic is exact and confirms the mechanism. Arrow normalises every drawing to an 80x80 unit square, scales it by min(width, height) / 80 and centres it: 765/80 = 9.5625, 900/80 = 11.25, and the vertical offset (990 - 765) / 2 = 112.5. The theoretical ceiling for a 765x990 canvas is therefore 765/990 = 77.3%, and Probe A reached 74.2%, which is 96% of everything available. There is no slack left for a prompt to recover.
Two secondary measurements matter for the framework:
- The Grok arm is double-letterboxed. Grok Imagine honours its
aspect_ratio: "3:4"parameter natively, so its raster is 3:4; Arrow's vectorizer then fits that 3:4 image inside the same 80x80 square. Measured ink width on all six Grok pages is 71.8 - 75.2% (exactly 3/4 of the square), and ink height is 58.0 - 74.4%, mean 65.5%. fit.pycannot fix this, and the fitted files prove it. After re-cropping to the inked bounding box, the five Quiver pages have aspect ratios of 0.969 to 1.078 — still square. The five Grok pages come out at 0.792 to 0.916, genuinely taller and closer to the 0.750 target of a 7.5x10in live area. The composition itself is square, not just the canvas, so cropping cannot recover a tall picture from it.
A correction to the 2026-10-06 README. I could not reproduce "about 62% of the page height" for the Quiver arm under any measurement I tried: raw canvas (74.2%), fitted output, or placed on the 8.5x11 sheet (65.5% of sheet, 72.0% of live area). The figure tracks the Grok arm instead: 65.5% of canvas, 63.6% of the 10in live area, with its weakest page at 58.0% and page 1 at exactly 62.7%. The README's headline recommendation (Quiver) still holds on its other four grounds, but the defect number quoted against Quiver appears to belong to the arm that lost.
What the web says
- There is no official QuiverAI documentation on viewBox handling, a fixed internal canvas, or letterbox behaviour. Two searches surfaced nothing from Quiver's own API reference, so this finding rests entirely on the reproducible measurement above rather than on vendor confirmation.
- The closest third-party corroboration is ComfyUI's node documentation for Quiver's text-to-SVG node, which states that width and height set the viewBox and control aspect ratio, and that leaving either at 0 lets the model choose, "which usually gives a square canvas." That is a third-party integration's description, not a vendor spec, and it describes the default rather than the override, but it is directionally consistent with a square-native model.
- Arrow 2 and a higher-capability sibling, Arrow 2 Telos, both shipped 2026-09-16.
list_modelsconfirms Telos is live on our account for generate, edit, vectorize and animate, tagged "Excels at complex SVG creations and drains usage much faster." Telos has never been tested at RDCO. Vendor tagline, not evidence.
Convergences and contradictions
- Convergence. The vault's implementation note (80x80 square), the sticker notes' folk-remedy (pick a square viewBox for square work), the ComfyUI third-party description, and the new probe all describe the same behaviour. The mechanism is settled.
- Contradiction, and it is the useful one. The 09-19 bake-off proved prompt language moves Arrow's output; the 10-06 remediation assumed that generalises to composition. It does not. Prompt language moved detail by roughly 25% on 09-19 and moved canvas coverage by zero on 10-18, against the strongest wording available.
- Contradiction on the 62% figure between the 10-06 README and this measurement, resolved above in favour of the measurement, which is reproducible from the archived SVG files.
Synthesis for RDCO
The answer to the first half is: hard limitation, and the 10-06 remediation is disproven. Arrow 2 composes inside a normalised square under every condition tested, and the emitted transform is byte-identical whether the prompt mentions the canvas or screams about it. Any production plan that depends on "write prompts for a tall composition" will fail, and the December coloring-book launch currently depends on exactly that line. This is the single most actionable finding here: the fix is not in the prompt, it is in the pipeline or in the page design.
The two gaps reconcile into one law, and it is the framework's spine: Arrow controls the drawing, not the frame. Arrow obeys semantic, content-level instructions reliably. Tell it to simplify and it simplifies (09-19, measured). Tell it to letter "14.8%" and it letters it. Tell it which parts are blue and it mostly complies. Arrow does not obey global, non-semantic constraints, and both known failures are of that type: canvas aspect is geometry the model never sees as content, and exact hex values are a global property the model approximates rather than enforces, which is why the sticker pipeline needs a deterministic palette-snap and a background-rect strip before any asset is usable. Grok is the mirror image. It obeys the frame natively (aspect_ratio works, colour is consistent) but hands back a raster with no editable geometry, and its content-level constraints are enforced by an external evaluator that actively fights lettering. The operating rule that falls out: prompt the content, post-process the frame. Anything global, aspect or palette, must be solved deterministically in code or chosen at the tool-selection step, never asked for in a prompt. This also explains why the two bake-offs reached opposite-feeling verdicts without contradicting each other: 09-19 tested content and Arrow won; 10-06 tested a frame property and nobody told the scorer that neither tool could win it.
For the December launch specifically, there are three honest paths and one of them is a design decision, not an engineering one. (1) Grok composes 3:4, Arrow vectorizes: the only route to a genuinely tall drawn composition, but it re-imposes the square at the vectorize step, broke 2 of 5 pages on 10-06 (a dropped wreath outline, a clipPath that cut the star), produces busier art that scored worse for ages 4 to 8, and cannot letter. (2) Arrow generates square objects and code composes the page: the 2026-10-06 composition test already proved this reads as a scene, built a reusable object library for $1.26, and is the on-contract answer to the founder's standing rule that real art tools draw pictures while code only does geometry. (3) Accept a square art block on a tall page with a caption or title band below, which is a common commercial coloring-book layout, costs nothing to engineer, and de-risks December immediately. My read is that (2) is the strategic answer because the object library amortises across every future book, and (3) is the fallback that protects the date. Path (1) is dominated: it loses on line quality and adds a lossy step. The genuine unresolved tension is composition-aspect (Grok) versus line quality (Arrow), and the test that settles it is a side-by-side of path 2 against path 1 judged on the founder's actual criterion, "bold and easy for ages 4-8", not on coverage percentage.
The reusable framework
Tool by job class
| Job class | Tool | Why, and the catch |
|---|---|---|
| Icons / isolated subjects, <=12 | Arrow singles | Best quality, drop-in SVG, letters correctly. Letterboxing is harmless here because isolated assets get trimmed to their bounding box anyway. |
| Icons, >12 subjects | Arrow one-call sheet + bbox split | About 3x cheaper per icon. Output has no groups (243 flat elements on 09-19), so splitting needs bounding-box clustering. |
| Full-page composition, non-square | Neither tool alone | Arrow cannot compose tall at all. Grok composes tall but is raster and vectorizing re-imposes the square. Use code composition of Arrow objects, or redesign the page around a square art block. |
| Exact brand colour | Arrow + deterministic palette snap | Arrow drifts off-palette (orange, grey). Never ship raw output. Strip the background rect too. |
| Anything with letterforms | Arrow | Grok's evaluator refuses glyphs. Note the 09-19 ruling on model-lettered text is still open; the customize spec says letterforms stay code-drawn permanently. |
| Scene-class raster art | Grok | Cheapest per call ($0.02), honours aspect_ratio, consistent colour. |
Standing bake-off checklist. Any future art-tool comparison measures these nine, because 10-06 measured only a subset and shipped a wrong remediation as a result:
- Frame fidelity. Request a non-square canvas; measure ink bbox as a percentage of each axis; record the emitted transform. Never infer this from eyeballing.
- Colour fidelity. Request exact hex; diff the output's hex set against the requested palette; count off-palette colours.
- Structure. Semantic groups or flat elements? Determines whether sheets are splittable.
- Closure. Percentage of open paths. Colour leaks are invisible on screen and fatal in print (10-06: open ox and donkey outlines).
- Lettering. Does requested text render correctly?
- Detail dial. Does prompt language move detail? Establishes what the tool will take direction on.
- Retry rate and failure modes under a fixed prompt.
- Cost per accepted page, not per call. Retries changed the 10-06 ranking.
- Post-process burden. How many deterministic fix-up steps before the asset is usable.
Known failure modes. Arrow: square-only composition; off-palette colour drift; open outlines; flat ungrouped sheet output; emits a semi-transparent full-canvas background rect; the MCP path returned model_error then insufficient_credits on 09-19 while REST worked, so scripts use REST and MCP stays a convenience. Grok Imagine: raster only; refuses lettering under the evaluator's text rule; evaluator max_tokens too small for 6-cell sheets; requires a vectorization step that re-imposes Arrow's square and damaged 2 of 5 pages.
Minor finding worth fixing regardless. The 10-06 run requested a 765x990 viewBox, an aspect of 0.773, while the live print area is 7.5x10in, an aspect of 0.750. Even a perfectly full-bleed result would have been slightly mismatched to the page.
Why this is in the vault
It kills the stated pre-production remediation in the 2026-10-06 coloring-book bake-off README ("write prompts for a tall composition"), which the December Shopify launch plan depends on, and replaces it with three costed paths plus a correction to that README's 62% figure. It also converts two one-off tool comparisons into the standing checklist that any future RDCO art-tool bake-off runs, so the next one does not again score a property neither tool can satisfy.
Open follow-ups
- Is the 80x80 square normalisation a property of Arrow's training representation, and therefore inherited by every Quiver model including Telos and future versions, or a serving-layer default that could change without notice? No vendor documentation exists either way. The answer decides whether to engineer around it permanently or treat it as temporary.
- Do competing production vector-generation models (Recraft, SVG.io, Illustrator's vector generation) impose the same square normalisation? If none do, the fix for tall compositions is tool choice rather than pipeline engineering, which would be much cheaper.
- What is the commercial norm for art-area aspect in published 8.5x11 coloring books? If a square art block with a caption band is standard, the constraint is not a defect and path (3) becomes the obvious answer rather than the fallback.
- Is Arrow's off-palette colour drift bounded and predictable enough that palette-snapping is lossless, or does it ever collapse two intended swatches into one? This determines whether multi-colour Arrow work can be trusted at volume or needs per-asset human review, which is the current practice.
Related
- [[2026-09-19-icon-set-bakeoff-quiver-vs-grok]]
- [[01-projects/coloring-books/2026-10-06-bakeoff/README]]
- [[01-projects/coloring-books/2026-10-06-composition-test/README]]
- [[2026-10-06-shopify-launch-plan]]
- [[DESIGN-scribble-works]]
- [[DESIGN-denominator]]
- [[2026-09-01-customize-spec]]
- [[2026-09-11-unit-economics-cost-basis]]
Sources
Vault
~/rdco-vault/01-projects/printables-product/2026-09-19-icon-set-bakeoff-quiver-vs-grok.md~/rdco-vault/01-projects/coloring-books/2026-10-06-bakeoff/README.md~/rdco-vault/01-projects/coloring-books/2026-10-06-bakeoff/tools/IMPLEMENTATION-NOTES.md~/rdco-vault/01-projects/coloring-books/2026-10-06-bakeoff/tools/gen.py,fit.py,render.py,calls.jsonl~/rdco-vault/01-projects/coloring-books/2026-10-06-composition-test/README.md~/rdco-vault/01-projects/personal-brand/assets/stickers/notes.md~/rdco-vault/01-projects/personal-brand/art-recipes/2026-10-06/recipes.md~/rdco-vault/02-sops/design-denominator.md,design-scribble-works.md
Primary measurement (this brief)
- Archived SVGs measured:
2026-10-06-bakeoff/quiver/raw/*.svg(6),grok/raw/*.svg(6), and the fittedquiver/*.svg,grok/*.svg(10). - New probes: Arrow 2 via
POST https://api.quiver.ai/v1/svgs/generations, 2026-10-18, viewBox 765x990 and 900x900, identical prompt. Artifacts and ledger in/tmp/art-probe-out/(probe-tall-aggressive.svg,probe-square-control.svg,probe-ledger.jsonl). Probe cost $0.56.
Web
- QuiverImageToSVGNodeV2 / QuiverTextToSVGNodeV2 node docs (ComfyUI, third-party integration)
- Arrow 2 Telos model listing (Vercel AI Gateway) — vendor listing, release date 2026-09-16
- Arrow 2 release coverage (i-scoop) — vendor-derived, not independent evidence
- No official QuiverAI API documentation on viewBox or canvas handling was found in two searches.