Fresh independent authors with the same pinned geometry, frozen checks and original guidance, plus explicit drawing review. Nine attempts complete across six models, including three Fable attempts and a guided GPT5.6Sol retry. Three attempts pass geometry; Fable produced no layout in any attempt. Visual quality remains for expert review; no human ratings assigned.
| Model | Geometry | Blocks | Minutes | Usage |
|---|---|---|---|---|
| Sonnet 5.5 | Pass | 55 | 35.06 | $5.49 |
| Opus 5.5 | Pass | 47 | 19.27 | $3.71 |
| Fable | No valid layout | — | 40.88 | $10.06 |
| GPT 5.6 Sol | No valid layout | — | 7.55 | $1.59 API estimate |
| GPT 6 Astra | Pass | 116 | 25.53 | $5.71 API estimate |
| Fable 5.1 · retry | No layout · API error | — | 99.43 | $23.19 |
| Fable 5.1 · manual | No layout · API error | — | 54.00 | $14.54 API estimate |
| GPT 6.1 Sol | No valid mesh | 143 | 42.48 | $1.70 API estimate |
Claude Code and Codex authors use different tool adapters. Codex subscription charges and provider-resolved model IDs are unavailable; API cost estimates are below; author-log timing is approximate. Initial smoke Claude usage totals $19.26; including the uncapped Fable retry, combined Claude usage is $42.45; Fable's post-response cutoff exceeded its $10 cap by $0.0588. Shared heavy checks were limited to two slots, so elapsed time can include queueing.
API-equivalent estimates from per-author Codex token logs. Service tier was not recorded: Standard and Fast rates shown separately. These are not subscription charges. Output includes reasoning. Running snapshots omit in-flight usage; completed author totals are final.
| Model / run | Part | New input | Cached input | Output | Standard estimate | Fast estimate | State |
|---|---|---|---|---|---|---|---|
| gpt-5.6-sol (smoke) | t8_profile_01068fb0 | 81,213 | 1,960,448 | 24,111 | $1.59 | $3.18 | complete |
| gpt-6-astra (smoke) | t8_profile_01068fb0 | 103,431 | 2,589,312 | 41,683 | $5.71 | $11.42 | complete |
| gpt-6.1-sol (smoke) | t8_profile_01068fb0 | 227,818 | 4,468,992 | 79,685 | $1.70 | $3.40 | complete |
| gpt-6-astra (prior trial) | t8_profile_01068fb0 | 74,399 | 1,451,648 | 20,421 | $3.22 | $6.43 | complete |
| gpt-6-astra (prior trial) | t8_profile_0103c4d8 | 58,193 | 1,087,104 | 15,637 | $2.45 | $4.90 | complete |
Final independently verified revision 5 · click to enlarge
5 full revisions · 35.06 min · $5.49
55 blocks · SJ 0.3390 / 2× 0.3359
Frozen strict, stable, block and occupancy checks passed.
Protocol: Partial: image reads for full revisions 1, 3, 5; none for 2, 4.
Codex observation: Compact feature enclosures; thin upper-hole ring and broad skewed background transitions remain.
Human grade: none assigned.
Result evidence · Author visual reviewFinal independently verified revision 1 · click to enlarge
1 full revisions · 19.27 min · $3.71
47 blocks · SJ 0.3649 / 2× 0.3743
Frozen strict, stable, block and occupancy checks passed.
Protocol: Verified draft and full-run image reads, review before submission and after full run.
Codex observation: Compact rings around lower-left hole and capsule; upper-hole connections and asymmetric boss core remain.
Human grade: none assigned.
Result evidence · Author visual reviewInput outline only — no authored draft · click to enlarge
0 full revisions · 40.88 min · $10.06
No submitted blocks · SJ — / 2× —
no revision run (budget cap)
Protocol: Outline image read only; no draft or visual-review document.
Codex observation: No authored layout available.
Human grade: none assigned.
Result evidenceRefused 44-block draft — no full revision · click to enlarge
0 full revisions · 7.55 min · $1.59 API estimate
No submitted blocks · SJ — / 2× —
No full revision: final 44-block draft failed strict 2D corner, inversion, overlap/outside-profile and TFI checks
Protocol: Author log records final refused draft image and visual review; no full submission. One concurrent read-only query batch disclosed.
Codex observation: Author rejected oversized rings, long diagonal fans and weak local symmetry.
Human grade: none assigned.
Result evidence · Author visual reviewFinal independently verified revision 1 · click to enlarge
1 full revisions · 25.53 min · $5.71 API estimate
116 blocks · SJ 0.4326 / 2× 0.4326
Frozen strict, stable, block and occupancy checks passed.
Protocol: Author log records draft reviews, pre-submission review and final image read. Native transcript unavailable.
Codex observation: Dense boss/background subdivisions and marked thin-block chain remain.
Human grade: none assigned.
Result evidence · Author visual reviewFinal refused revision 2 · click to enlarge
2 full revisions · 42.48 min · $1.70 API estimate
143 layout blocks · SJ unavailable · no realized mesh
STRICT: C5 boundary crosses itself at q73 and q125; mesh checks not run
Author log records draft16/draft18 and full rev1/rev2 drawing reads, with review before both submissions. Two initial read deviations were disclosed: task-root listing and the next function signature following point(). The independent frozen reread refused STRICT; original rows identify boundary crossings at q73 and q125.
Author observations: excess subdivisions, diagonal wedges and imperfect local symmetry. Human grade: none assigned.
Result evidence · Author visual reviewEarlier attempts: Sonnet failed at 67 blocks; Opus failed after its bounded budget recovery at 62 blocks; Fable had no submitted revision; Astra passed at 49 blocks. The new Sonnet and Opus attempts pass, but stochastic fresh starts, budgets and visual guidance vary together: this does not isolate a causal improvement or rank models. Astra's new passing layout uses 116 blocks versus 49 previously. Passing witnesses establish geometric tractability under the frozen checks, while aesthetic quality remains open. Previous comparison gallery
0 drafts · 0 full revisions · 99.43 min · $23.19 Claude CLI usage
API response exceeded 64,000 output tokens. Session dollar/time caps were removed; the worker ended with a provider output-limit error. No layout, mesh or SJ exists.
449,567 output tokens across the session, including 448,672 thinking tokens. Outline image read only; no draft/full image or visual review. The original capped Fable attempt above is preserved.
Result and usage evidence0 drafts · 0 full revisions · 54.00 min · $14.54 Standard API-equivalent estimate
Four consecutive 64,000-output-token responses consisted entirely of thinking. Claude Code then reported the output-token-limit error. No authored layout, realized mesh, SJ or element count exists; frozen mesh reread was not run.
Native message usage: 258,025 output tokens, including 256,400 thinking tokens; 209 new input, 845,178 cached input, 71,492 cache-write tokens. Message IDs were deduplicated; prior Opus worktree setup excluded. Actual subscription charge unknown.
105 pinned files unchanged. Outline image read; no draft/full review. Interactive repository context and settings differ from automated launches; effort unknown. This is an execution failure, not a geometric impossibility result. Audit deviations: study-root listing and directory change, a point() query in another checkout, and a witness.py read extending beyond the allowed function. No human grade assigned.
Result, metrics and accounting evidence · Official API rates
0 full revisions · 8.17min · $1.90 Standard API-equivalent estimate
Final draft:50 entered blocks,40 drawn plus10 fixed-fill. Three occupied regions passed2D checks; the base region failed inversion, corner angles, outside-profile, overlap and curved-block quality. No draft passed check; no full revision or realized mesh exists. Frozen mesh reread not run.
Prompt change: generic guidance to write/check concrete candidates and reconsider the overall partition when local repairs stall. No earlier draft, successful layout, model outcome or part-specific solution supplied. The same original O-grid preferences and geometry checks remained. No per-part time/spend cap; this author chose to stop.
105 pinned files unchanged. Actual outline and candidate1/2/3 image-read calls verified. Native audit discloses a witness.py read beyond point() into neighboring functions. Actual subscription charge and resolved provider snapshot unknown; Fast API equivalent $3.8027. No human grade assigned.
Result and usage evidence · Refused draft JSON · Author visual review
Shared SJ scale 0–1; frozen floor 0.2. Rotate each mesh to inspect exposed element faces. Surface coloring does not expose all interior elements.
No full mesh generated
No full mesh generated
No realized mesh: STRICT refused the submitted layout