Rendered report of MODEL-COMPARISON.md — the ambush pixel-art spike's outputs are previewed below as a reference strip.
Caveat up front: n = 1 prompt per medium, 1 run per arm. Rankings between adjacent arms (±1 point) are noise-level; the model-tier gaps and the effort-curve shapes are the robust part. See “What this can't tell you” at the end.
Two controlled spikes, eleven arms each (haiku, then sonnet-5 and opus-5 at low/medium/high/xhigh/max effort), one identical prompt per spike, no human art direction.
svg-mosaic skill) — live at https://phandalin-mosaic-spike.pages.dev/, local arms in spike-mosaic/<arm>/phandalin.htmlpixel-art skill) — live at https://ambush-pixel-spike.pages.dev/, local arms in spike-ambush/<arm>/shot.pngAll 22 works were viewed for this report: the 11 shot.png renders directly, the 10 surviving local mosaic scenes rendered in Chrome at 1200px, and the haiku mosaic screenshotted from the published site (its local copy was destroyed by an unrelated accident and is not recoverable; the published page still serves the original, which is what was judged).
Provenance of numbers. Tokens, tool calls, and wall time are harness-measured (USAGE.json in each spike root). Iterations, screenshots, palette size, shape/cell counts, canvas resolution, and residual-failure lists are self-reported by each arm (PROFILE.json per ambush arm; the stats block inline in spike-mosaic/gallery.html for the mosaic arms). Self-reports are treated as testimony, not fact, and are cross-examined against the pictures in the honesty section.
All 11 arms of the forest-ambush pixel-art spike (spike-ambush/<arm>/shot.png). Click a thumbnail to zoom.
| arm | tokens | tools | wall (s) |
|---|---|---|---|
| haiku | 57,726 | 36 | 332 |
| sonnet-low | 62,542 | 28 | 270 |
| sonnet-medium | 91,573 | 55 | 480 |
| sonnet-high | 88,818 | 48 | 574 |
| sonnet-xhigh | 144,665 | 86 | 1,137 |
| sonnet-max | 183,594 | 95 | 1,609 |
| opus-low | 62,505 | 31 | 371 |
| opus-medium | 107,174 | 50 | 664 |
| opus-high | 111,235 | 48 | 975 |
| opus-xhigh | 225,682 | 60 | 2,202 |
| opus-max | 218,480 | 66 | 2,007 |
| arm | tokens | tools | wall (s) |
|---|---|---|---|
| haiku | 42,452 | 22 | 188 |
| sonnet-low | 40,321 | 7 | 83 |
| sonnet-medium | 44,481 | 11 | 136 |
| sonnet-high | 64,959 | 21 | 362 |
| sonnet-xhigh | 78,458 | 20 | 517 |
| sonnet-max | 111,653 | 30 | 894 |
| opus-low | 44,655 | 9 | 157 |
| opus-medium | 61,514 | 24 | 445 |
| opus-high | 115,217 | 45 | 1,046 |
| opus-xhigh | 183,630 | 40 | 1,893 |
| opus-max | 163,952 | 71 | 1,742 |
Ambush (PROFILE.json): iterations / revisions after first render / palette / canvas.
| arm | iters | revs | palette | canvas | residual failures listed |
|---|---|---|---|---|---|
| haiku | 5 | 4 | 37 | 256x192 | 0 |
| sonnet-low | 3 | 2 | 24 | 320x200 | 3 |
| sonnet-medium | 3 | 2 | 27 | 256x176 | 2 |
| sonnet-high | 4 | 3 | 24 | 256x144 | 2 |
| sonnet-xhigh | 8 | 7 | 24 | 256x144 | 0 |
| sonnet-max | 6 | 7 | 19 | 256x144 | 2 |
| opus-low | 8 | 7 | 20 | 320x180 | 2 |
| opus-medium | 11 | 10 | 22 | 320x180 | 4 |
| opus-high | 9 | 8 | 21 | 320x180 | 3 |
| opus-xhigh | 12 | 11 | 21 | 320x180 | 3 |
| opus-max | 17 | 16 | 23 | 320x180 | 2 |
Mosaic (gallery.html stats block): iters / shapes / cells / revs — haiku 4/14/1866/3; sonnet-low 1/47/1534/0; sonnet-medium 1/37/2297/0; sonnet-high 3/45/1221/1; sonnet-xhigh 4/40/2807/3; sonnet-max 3/46/2767/1; opus-low 3/69/1261/2; opus-medium 4/86/2134/3; opus-high 7/231/5412/6; opus-xhigh 9/203/4081/7; opus-max 10/215/3254/8.
The shape counts are the single most diagnostic self-reported number in the mosaic: sonnet never exceeds 47 scene shapes at any effort; opus jumps from 69 (low) to 231 (high). Effort changes how much scene opus plans, not just how much it polishes.
Scale: 0–10, where 5 = “competent but unremarkable,” 7 = “would pass casual scrutiny as intentional art,” 9+ = “portfolio-grade for the medium.” Scored per medium after viewing all arms; half-points allowed. Art = composition, value structure, focal control, colour, atmosphere. Tech = execution against the skill's own checklist (artifacts, dither soup, tangents, broken silhouettes, rendering bugs). Detail = density and placement. Fidelity = every prompt element present and legible.
| arm | Art | Tech | Detail | Fidelity | notes tied to the picture |
|---|---|---|---|---|---|
| haiku | 2 | 3 | 2 | 3 | Flat solid-green field, castle as grey rectangles with black window squares, sprite figures standing inside the ruin facing the camera; two flanking figures face front, not “from behind”; no forest, no hill, a wavy blue band for sky. Reads as a game-tile diagram, not a picture. |
| sonnet-low | 3.5 | 4 | 3 | 4 | Handsome dawn-dither sky and sun, but the “three adventurers” are featureless brown slabs that read as tombstones; castle is a diagonal-hatched wedge; horizontal scanline banding across the whole frame; no legible goblins. Its own PROFILE admits the figures fail. |
| sonnet-medium | 5.5 | 6 | 4 | 6.5 | Clean theatre-wings composition: dark cliff/canopy frame, three genuinely readable cloaked backs, castle on the far ridge, torch pole. But it is flat-vector in texture — big untextured colour fields, minimal dithering — and the “goblin patrols” are two or three ambiguous specks with torches. |
| sonnet-high | 5 | 5 | 5 | 6 | Moody violet dusk with lit castle windows; three brown cloaked backs on the near hill. Undermined by artifacts: loose dither columns hanging in the right and left margins, a hard black swath cutting the midground, and the rim seam its PROFILE admits is a ruler-straight 1px line. |
| sonnet-xhigh | 6.5 | 6 | 6 | 7.5 | The most graphic image of the sonnet run: orange rim light on purple night, strong tree/figure/castle silhouettes, visible green goblins near the ruin. But the rim-light trick is sprayed on everything — trees, all three figures, every crenellation, even ground strips — so the device stops directing the eye. Claims zero residual failures; see honesty section. |
| sonnet-max | 6 | 7 | 5 | 7 | Disciplined 19-colour moonlit scene, ruined twin towers with a genuine breach, path leading in, dithered moon. Cleanest execution of the sonnet arms, but the midground is empty lawn, the party is three small near-black smudges, and goblins are torch-dots. Careful, slightly inert. |
| opus-low | 6 | 5.5 | 6 | 7.5 | Atmospheric mist over the ruin, party with readable shield and bow, a looming hooded figure with staff, green goblins with torches ringing the clearing. Foreground bramble texture collapses toward dither soup (self-admitted) and the big right-hand figure crowds the frame. |
| opus-medium | 6.5 | 5.5 | 6.5 | 8 | Best staging idea of the mid arms: a moonlit bowl with the castle spotlit at centre and many patrolling goblin silhouettes around the ring; birds on a branch overhead. Marred by horizontal streak-banding in the sky/mountains and a near-flat mid-forest ring — both of which its PROFILE candidly lists. |
| opus-high | 7.5 | 7 | 7.5 | 9 | The full prompt, legible: archer with bow, dwarf with round shield and hammer, hooded caster, all from behind; ruined abbey-like castle with tower; eight-plus goblin sentries with orange torches; path, wildflower meadow, framed by dark canopy. Detail concentrated at the ruin and the party, background kept quiet. |
| opus-xhigh | 8 | 7.5 | 8 | 8.5 | The best picture in the pixel set. Real aerial perspective — grey-blue pine ranks behind a sunlit ruined keep in warm stone, green hill rolling down, dark framing trunks. The shield's rim-lit circle and the drawn bow read instantly. Goblins are marginal at display size (self-admitted) and a snapped branch hangs oddly in the upper sky. |
| opus-max | 7.5 | 7 | 8 | 9 | Best scene logic: a goblin camp with campfire at the castle gate, patrols with spears, birds, low sun halo behind the keep. Fidelity is the best in the set. But the party cost it 8 of 17 iterations and still half-reads as boulders (its own words), and the dwarf's axe is a grey wedge. More picture than opus-xhigh, slightly worse picture-making. |
| arm | Art | Tech | Detail | Fidelity | notes tied to the picture |
|---|---|---|---|---|---|
| haiku | 2 | 3 | 1 | 2.5 | 14 shapes: three triangle “mountains” interleaved with three house-gables at the same scale so they read as one row of tents; the green ground is split into arbitrary darker rectangles that look like a rendering bug; a white lens shape floats in the sky. Not a village, barely a scene. (Judged from the published page.) |
| sonnet-low | 5 | 5 | 4 | 6.5 | A real village at night in 83 seconds: mountains, moon, eight cottages with lit windows, road, signpost, well. But the scene sits letterboxed in the top two-thirds of the viewport with dead space below, and the sky is three flat horizontal bands with visible seams. |
| sonnet-medium | 5.5 | 6 | 4.5 | 6.5 | Daylight version, single iteration: mountain wall, nine cottages, road, manor with red pennant. Everything is evenly weighted — no focal building, sun is a plain circle — so it reads as a pleasant pattern rather than a place. |
| sonnet-high | 6 | 6.5 | 5 | 7 | Adds what medium lacked: ringed sun, a ruined tower on the left hill (Phandalin lore point), lit windows on the central cottage, path to the door, pines. Still poster-symmetric, and a hard pale-green terrace band runs edge-to-edge behind the houses like a stripe, not a field. |
| sonnet-xhigh | 6 | 6 | 5.5 | 7 | The most landscape-like sonnet: four mountain ranges in depth, castle ruin upper right, moon with birds. But the village centre collapses into a brown house-cluster with a mud apron that reads as one blob, and a white obelisk shape by the inn is unreadable. |
| sonnet-max | 6 | 4.5 | 5.5 | 6.5 | Best sky of the sonnet run — banded violet-rose-orange dusk behind blue ridges — and a disciplined village with watchtower. But the shipped frame is truncated: the bottom fifth of the viewport is empty black and the foreground figures/road are cut mid-body at the frame edge. Most sonnet tokens (111.6k), and the one sonnet arm with a frame-level defect. |
| opus-low | 6.5 | 6.5 | 5.5 | 7 | 69 shapes in 157 s: a night village with a clear focal inn (big door, four glowing windows), nine satellite cottages, road, log wagon, ruin fragments on the right ridge, clustered clouds. First arm in the mosaic set where the eye is sent somewhere. |
| opus-medium | 7 | 7 | 6 | 7.5 | Dusk with a haloed moon over snowcaps, chimney smoke, fenced yard, path to a two-storey inn, ruined manor on the right — Tresendar Manor placement, correct lore. Cohesive light temperature (warm windows against cool blues) carries it. |
| opus-high | 8 | 7.5 | 7.5 | 8.5 | 231 shapes / 5,412 cells: thatched great-inn, stone cottage, fences, wagon, two figures on the road, tree clusters, pine treeline, layered clouds, sun sunk behind the peaks. The one risk — even-spray density — is held off because the big thatched inn and the road junction anchor the eye. The most illustrated daylight scene of the 22. |
| opus-xhigh | 8.5 | 8 | 8 | 8 | The designed one: leaded-glass idiom with glow blooms on lit windows and a lamppost, a stone path leading past two figures to the lit manor, sunset band behind three mountain ranges, ruined tower upper right, typeset title block. Best focal control in either spike — the path is a literal arrow to the brightest window. The glow gradients bend the flat-glass idiom (restraint note below). |
| opus-max | 8 | 7.5 | 7.5 | 8.5 | Densest world: inn with red roof and smoking chimneys, shrine, log cart mid-road with wheel ruts, ruined castle hill right, dozens of outbuildings, snowcaps, sunset. Richer inventory than opus-xhigh but detail is spread more evenly, so the eye wanders; the rutted road does most of the compositional work. |
Justification: every arm filed a PROFILE; comparing testimony to the visible picture is a free, fully-supported axis, and it is the axis a user of these arms most needs (can I trust the arm's own DONE-claim?).
Justification: both skills have explicit discipline rules (palette caps, “1–2 shimmering shapes,” dither hygiene), so restraint is measurable against the skills' own texts, not taste.
Justification: iterations are self-reported but screenshot counts and token totals corroborate them; the spread (1 to 17 iterations) is wide enough to mean something.
Justification: the PROFILE narratives plus the pictures let you distinguish arms that fixed a wrong composition from arms that buffed one.
Justification: same 11 arms, two media — the only cross-cutting comparison this design supports.
Does more effort buy better art, and where does the curve flatten or reverse?
On opus: yes, monotonically to xhigh in both media (mosaic Art 6.5→7→8→8.5; ambush 6→6.5→7.5→8), then flat-to-slightly-down at max (8 and 7.5) while cost stays at the ceiling (opus-max: 218k/163k tokens vs opus-xhigh 226k/184k). On sonnet: the curve rises to high/xhigh then goes flat (ambush 6.5→6) or negative (mosaic: sonnet-max shipped the truncated frame at the highest sonnet cost). The reliable buy-zone is opus low→xhigh; past xhigh you pay max prices for max's failure modes. One-run caveat applies especially to the max-vs-xhigh comparison — the gap is within single-run variance.
Opus-at-low vs sonnet-at-max?
Absolute terms: opus-low wins the mosaic (6.5 vs 6, and no frame defect vs a truncated frame) and ties the ambush (6 vs 6, trading opus-low's atmosphere against sonnet-max's cleaner execution). Tokens: opus-low used 40% (mosaic) and 34% (ambush) of sonnet-max's; wall time 157 s vs 894 s and 371 s vs 1,609 s. Dollars depend on the opus/sonnet per-token price ratio, which this report does not assert — at any ratio under ~2.5x opus-low is cheaper outright; at higher ratios sonnet-max is cheaper per run but buys equal-or-worse art much slower. There is no quality case for sonnet-max over opus-low in this data; the only case is a per-token price ratio steep enough that you'd rather have sonnet-medium anyway.
Where does haiku land?
Last, in both media, by a wide margin (Art 2/2), and it is not cheap enough to excuse it: 42.5k tokens in the mosaic vs sonnet-low's 40.3k for a picture two full tiers better, and 57.7k in the ambush vs sonnet-low's 62.5k. Strictly dominated by sonnet-low on quality-per-token, and the only arm whose self-report misstates what is on the canvas. Not the bargain of the set; the bargain is sonnet-low (mosaic) / sonnet-medium (ambush).
Do the two skills rank the models the same way?
At tier granularity yes (haiku < sonnet ≤ opus; opus-xhigh on top twice). Within tiers, no: sonnet's best pixel arm (xhigh) is mid-pack in glass; opus's effort dial pays off harder in glass (where effort became 3–5x the scene shapes) than in pixels (where it became figure-legibility grinding). Divergence reading: mosaic quality is planning-bound, pixel quality is judgment-bound — effort converts to plans better than it converts to taste.
Over/underspenders and value picks?
Overspent: opus-max in both media (ceiling cost, no gain over xhigh; in the ambush, 8 of 17 iterations sunk into one still-unresolved element); sonnet-max in the mosaic (111.6k tokens, shipped the only broken frame); opus-medium in the ambush (11 iterations of churn for a mid result). Underspent: haiku in both (though more effort would not have fixed its perception problem); sonnet-low in the ambush (quit with figures it had itself flagged as failing). Value-optimal: mosaic → opus-low, budget floor sonnet-medium; ambush → opus-high, budget option sonnet-xhigh. Quality-optimal in both: opus-xhigh.
spike-ambush/USAGE.json, spike-mosaic/USAGE.jsonspike-ambush/<arm>/PROFILE.json; mosaic stats block at spike-mosaic/gallery.html (lines ~329–339)spike-ambush/<arm>/shot.png; mosaic arms screenshotted locally from spike-mosaic/<arm>/phandalin.html at 1200px; haiku mosaic from https://phandalin-mosaic-spike.pages.dev/haiku/phandalin.html (local copy destroyed)