Claude Sonnet 5.5
30 runs across 10 tasks, costing $1.06, and 5 posts. Each table shows the runs under the newest prompt this model was given.
How it failed: incomplete, false, unreadable, built wrong.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| All seven partsby eye · Pass if all seven are drawn as separate shapes and each has a label naming it. A part that is only named in a label or a comment, with nothing drawn, fails it. | ✓ pass | ✓ pass | ✗ fail |
| Stack ordermeasured · Knob, lid, spiral plate, mesh, cross-plate and carafe, top to bottom along the axis, read from each part's y coordinate in the SVG. Its comment is ignored. The rod is left out, because it runs through the lid. | ✓ pass | ✓ pass | ✗ fail |
| Filter fits the glassmeasured · Every filter piece drawn is 80% to 105% of the width of the carafe's glass wall, measured in the SVG. | ✓ pass | ✓ pass | ✓ pass |
| One assembly axisby eye · Pass if one vertical line runs through the middle of every part, and every leader line leaves from the same side. One part pushed off to the side fails it. | ✓ pass | ✓ pass | ✓ pass |
| Labels matchby eye · Pass if following each leader line from its label ends on the part that label names. One label pointing at the wrong part, or at nothing, fails it. | ✓ pass | ✓ pass | ✓ pass |
Prompt v2, checks c1, harness h1. Ground truth: The standard three-piece plunger, from the top down: spiral (spring) plate, mesh screen, cross-plate. The cross-plate is the bottom piece, because its threaded stem is what the rod screws into. The spiral plate above the mesh is what presses the mesh edge out against the glass. Checked against the WebstaurantStore anatomy diagram, which is drawn exploded and labelled, and against each part's y coordinate in the SVG rather than its comment. Two-piece travel presses exist and stack differently, which is why v2 of the prompt names the design. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Parsesmeasured · RDKit reads the SMILES as a molecule. | ✓ pass | ✓ pass | ✓ pass |
| C8HF9N4O2measured · Molecular formula computed from the parsed molecule. | ✓ pass | ✓ pass | ✓ pass |
| Three N-CF3, no N-methylmeasured · Each CF3 carbon traced to a ring nitrogen, and no methyl left on any nitrogen. | ✓ pass | ✓ pass | ✓ pass |
| The right moleculemeasured · Canonical SMILES identical to the target, so a CF3 on the wrong nitrogen fails. | ✓ pass | ✓ pass | ✓ pass |
Prompt v1, checks c1, harness h1. Ground truth: Caffeine's skeleton with a CF3 on N1, N3 and N7, which is C8HF9N4O2. The prompt pack this came from gives C8H3F9N4O2, which is wrong: swapping the methyls takes all nine hydrogens with them, and only the C8 hydrogen stays. Decided with RDKit by an exact canonical match, because fluorinated isocaffeine has the same formula with a CF3 on the wrong nitrogen. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Compilesmeasured · pdflatex builds the file offline. | ✓ pass | ✓ pass | ✓ pass |
| Seven regions, three namesmeasured · Three circles found in the PDF with every overlap present, and each set name placed by its circle. | ✓ pass | ✓ pass | ✗ fail |
| Every drink in its regionmeasured · Each of the eight words found in the PDF, and the centre of each one inside the region it belongs to. | ✓ pass | ✓ pass | ✗ fail |
| No label on an edgemeasured · No letter touches a circle's outline, with the line's width counted. | ✓ pass | ✓ pass | ✗ fail |
Prompt v1, checks c2, harness h1. Ground truth: Short black is espresso only, flat white espresso and milk, iced long black espresso and cold, iced latte all three, babyccino milk only, milkshake milk and cold, cold brew cold only, and pour-over outside every circle. The circles and words are read back out of the compiled PDF, each word has to be centred in its own region, and no letter may touch a circle's outline, with the line's width counted. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Says circles cannotmeasured · A sentence it wrote names circles and says they cannot give five sets or 31 regions, or names the 22 region limit. | ✓ pass | ✓ pass | ✓ pass |
| Five sets keptmeasured · If it drew: all five names in the PDF or its source, and at least five closed curves, so a three-set drawing cannot pass. | ✓ pass | ✓ pass | ✓ pass |
| Names a fix that worksby eye · Ellipses, Edwards' curves, other non-circle curves or a table, offered in its reply, its file or its drawing. | ✓ pass | ✓ pass | ✓ pass |
Prompt v1, checks c1, harness h1. Ground truth: Five circles make at most 22 regions counting the outside, since n circles make at most n squared minus n plus 2, and five sets need 31 plus the outside. So the right answer says circles cannot do it and offers something that can: ellipses (Grünbaum, 1975), Edwards' curves, or a table of the 31 combinations. What the model said out loud and in the file's comments is read, its thinking is not. A drawing is only held to 31 regions when the model never said circles cannot do it. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Reads as a shrubby eye · Pass if it reads as one plant with branches coming off stems, without being told what it is. A tangle, a lattice or loops fail it. | ✓ pass | ✗ fail | ✓ pass |
| Flat side from the rulesby eye · Pass if one side ends on a straight line that the rules themselves put there. A box drawn around the plant to cut it fails it. | ✗ fail | ✗ fail | ✓ pass |
| Sensible sizeby eye · Pass if the branches can still be followed by eye. So many segments that they merge into a mass fails it. | ✓ pass | ✓ pass | ✓ pass |
Prompt v1, checks c2, harness h1. Ground truth: The JSON expanded and drawn with turtle graphics offline, and the drawing judged by eye. There is no single right L-system, so nothing here is measured. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| HTML and CSS onlymeasured · No <script>, <svg>, <img>, <canvas>, <iframe>, external url(...) or javascript: URL in the file. | ✓ pass | ✓ pass | ✓ pass |
| A CSS loop that moves somethingmeasured · An @keyframes rule that changes a transform, a translation, a rotation or an offset between stops, named in an animation on an element, with the iteration count infinite. It does not know which element moves: on 2026-10-19 a GLM run passed it by turning the coffee surface while the spoon stood still, and the eye checks are what caught that. | ✓ pass | ✓ pass | ✓ pass |
| Reads as a cup seen intoby eye · The coffee surface is visible inside a rim that has some thickness, whether that is an ellipse seen from slightly above or a ring seen from straight down. A flat disc with no rim, or a cup whose inside cannot be seen, fails it. | ✓ pass | ✓ pass | ✓ pass |
| The spoon does the stirringby eye · The spoon itself, bowl and handle together, travels round inside the opening, so its place changes across the loop. Judge it from the live page or the video, since the four still moments can show almost the same instant. A file whose only motion is in the coffee surface, with the spoon stood still at the rim, fails it. | ✓ pass | ✓ pass | ✓ pass |
| Reads as inside the cupby eye · The bowl stays within the rim of the opening and the handle leaves over the near edge, so the spoon reads as in the drink rather than laid on top of the picture. A spoon drawn across the cup mouth in front of the rim for the whole turn fails it, which is the trap. | ✓ pass | ✓ pass | ✓ pass |
| Handle stays on the bowlby eye · The handle meets the bowl of the spoon in every frame. A handle that separates, points the wrong way or is missing fails it. | ✓ pass | ✓ pass | ✓ pass |
Prompt v1, checks c1, harness h1. Ground truth: HTML and CSS alone, a continuous loop, and a cup seen from slightly above. The file is read for the brief and for the loop: an @keyframes rule that changes a position or a transform, named by an element, with the iteration count set to infinite. The picture is four moments of the animation, left to right, taken at 0.4 s, 1.2 s, 2.0 s and 2.8 s. Whether the rim reads as an ellipse seen from above, whether the spoon passes behind the far rim on the back half of its circuit, and whether the handle stays on the bowl are judged by eye from the page the run rendered, with the four frames as a first look. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| In the drinkby eye · A spoon drawn across the cup mouth in front of the rim for the whole turn fails, which is the file the run was handed. | ✓ pass | ✗ fail | ✗ fail |
| Kept the brief and the loopmeasured · The HTML and CSS survived, with no script, image or canvas added, and the loop still moves something. | ✓ pass | ✓ pass | ✓ pass |
Prompt v1, checks c1, harness h1. Ground truth: GLM 5.3 FlashX drew this cup on 2026-10-19. The spoon's bowl sits at the back of the opening with the handle laid across it toward the near rim, so the spoon reads as resting on the cup while the coffee surface turns under it. The fix is a change of z-order: the handle has to pass behind the near rim and the bowl has to sit in the drink. Nothing in the file can be measured for it, because both arrangements use the same shapes in the same places, so the eye decides it and `unchanged` guards the rest. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Serves all sixmeasured · The tray hands each of the six tables a drink it will take, every drink used exactly once and none twice. A refusal leaves it undecided, and a tray that reuses a drink fails it. | ✓ pass | ✓ pass | ✓ pass |
Prompt v2, checks c1, harness h1. Ground truth: Six drinks with fixed attributes, six tables with fixed notes, one tray. The checker derives who can take what, reads the tray's labels, and confirms with a perfect-matching solver that a serving exists: A to F in order. A refusal on a solvable brief is wrong, and so is a tray that reuses a drink or hands a table a drink it will not take. Whether a drawn cup really holds what its label says is the eye's. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Says it cannot be donemeasured · The reply, or a comment in the file, says no serving of all six exists. A run that draws a tray instead and says nothing fails it. | ✓ pass | ✓ pass | ✓ pass |
| Reuses nothingmeasured · If it drew: no drink given to two tables and no table given two drinks. A run that drew nothing is undecided. | ✓ pass | ✓ pass | ✓ pass |
Prompt v2, checks c1, harness h1. Ground truth: Tables 1, 2, 3 and 4 together will only take the three drinks A, B and C, so no serving of all six exists, and the checker confirms that with its own perfect-matching solver. The right answer says so and draws nothing. Every table still has a drink it will take on its own, so the shortage is only visible across the tables together. Every model on this task.
| check · how it is decided | run 1 | run 2 | run 3 |
|---|---|---|---|
| Serves all sixmeasured · The tray hands each of the six tables a drink it will take, every drink used exactly once and none twice. A refusal leaves it undecided, and a tray that reuses a drink fails it. | ✓ pass | ✓ pass | ✓ pass |
Prompt v2, checks c1, harness h1. Ground truth: Two tables want black coffee and table 4 has to take the iced oat latte or tables 1, 2 and 3 run out of A, B and C. A serving exists, A to F in order, and the checker confirms it with its own matching solver. Refusing because it looks crowded is wrong. Every model on this task.