Skip to content

specsmith Governance Efficiency — Model Comparison

Run: 29650915022

Date: 2026-07-18

Cross-model ranking withheld

GPT-4o-mini completed all 182 requested cells. Qwen3-Coder completed only 74; 108 cells were unavailable because the Hugging Face provider stopped serving requests. A partial provider run cannot support model rankings, relative cost-of-pass claims, or governance-lift comparisons.

Artifact completeness

Model Requested Valid Provider errors Passes among valid Comparative status
gpt-4o-mini 182 182 0 111 (61.0%) Valid single-model evidence
Qwen/Qwen3-Coder-30B-A3B-Instruct 182 74 108 36 (48.6%) Invalid for comparison

Qwen failures consisted of 107 HTTP 402 credit-depletion responses and one HTTP 429 token-per-minute response. They are infrastructure failures, not failed task solutions.

Complete-model findings

For GPT-4o-mini, ungoverned achieved 64% aggregate pass rate. specsmith LIGHT, FULL, and DISPATCH each achieved 57%. The results are task-dependent: LIGHT and FULL improved T10 from 0% to 100%, while governed conditions underperformed on T1 and T13. T6 and T7 were non-discriminating because nearly every condition passed.

See Governance Efficiency Benchmark for complete condition and task tables.

Next comparison matrix

Role Model Rationale
Historical baseline gpt-4o-mini Preserves continuity with prior runs
Open-weight managed route Qwen/Qwen3.6-35B-A3B:deepinfra Current Qwen coding/repository model with a live tool-capable HF route
Cost-conscious current model gpt-5.6-luna Modern closed-model screening baseline
Frontier anchor gpt-5.6-sol Upper-bound comparison for complex coding and reasoning

The FP8 repository Qwen/Qwen3.6-35B-A3B-FP8 is suitable for a separate self-hosted experiment. It currently has no managed Hugging Face Inference Provider mapping, so combining it with managed API models would also combine model-quality and serving-infrastructure effects.

Publication rule

Future reports are generated only when every selected model has the identical task/condition/repetition cell set and every cell is valid. Incomplete artifacts remain downloadable for diagnosis but cannot produce a successful comparison.