Updated Paper Results

[QingyunQian]

Take-home results

This PR collects the paper-ready ORBIT-Q results, now including the complete campaigns from PR #27, PR #28, and PR #29.

TensorCircuit-NG solver	Passed	Runtime / expert (Fig. 2b)
GPT-5.6 Sol	10/12	2.541×
GPT-5.6 Sol ultra	10/12	2.072×
GPT-5.6 Terra	9/12	2.702×
GPT-5.6 Luna	9/12	4.692×
DeepSeek V4 Flash	5/12	3.476×
Grok 4.5	8/12	3.844×
DeepSeek V4 Pro	5/12	2.588×
Grok 4.6	9/12	2.191×
GPT-6 Astra	12/12	0.540×

Published campaign outcomes are preserved. Astra reports 12/12 automatic passes; the maintainer subsequently accepted its Task 08 implementation. Earlier campaigns' Task 08 outcomes are unchanged. All named configurations use high thinking effort except Sol ultra.

Astra's runtime comparison uses eleven same-host measured pairs and the reviewer's estimated 50 s expert reference for Task 08 (25.05 s × 2, rounded). Astra is faster on 8/12 tasks, with 0.540× geometric-mean runtime (1.85× speedup). All non-Astra Fig. 2b coordinates retain their previous fixed-paper-reference values.

Updated Fig. 1c — agent–framework benchmark matrix

Updated Fig. 2b — failure rate versus artifact runtime

Updated Fig. 4 — benchmark resource use

DeepSeek V4 Pro, Grok 4.6 and Astra appear in every agent-side panel. Astra's solver totals are 44.72 min, 6.835M tokens, and $13.967. Bubble area represents total solver cost.

Astra artifact runtime versus expert reference

Summed artifact times are 440.81 s for Astra / 640.11 s for the expert reference, a ratio of 0.689×. The right panel shows expert/Astra speedup. Full table and replacement provenance; the original measured records are retained separately.

Human expert + AI co-optimization

The existing 12-task optimization results are unchanged. The largest retained improvements are Task 07 (45.758×), Task 01 (9.636×), and Tasks 03/09/10/12 (3.82–4.90×).

Validation

• Checked 108 unique added-campaign task outcomes and the three new source campaigns against pinned commits.
• Checked all thirteen Fig. 2b points; only Astra's runtime basis changes.
• Verified the Task 08 estimate, all per-task ratios, reciprocal speedups and aggregate totals.
• All 12 data regression tests pass; PDF masters were rendered and visually checked.
• Original framework/expert-optimization data and unrelated repository files are unchanged.

Concise report, source tables, plotting code and QA.