Harness behavior, not model selection
Every run used qwen3.6:35b-mlx through local Ollama on the same machine.
EVIDENCE / 2026-07-11
Show the receipts.
This aggregate report describes a small local comparison and is explicit about what has not yet been published or independently reproduced.
REPORTED LOCAL RANK 019/9reported Algo CLI checker passes
Three draft tasks, three repetitions per harness, one local Qwen model, one machine, and balanced run order. Every harness occupied positions one, two, and three exactly once per task.
| Harness | Objective | Clean runs | Median | p95 |
|---|---|---|---|---|
| Algo CLI | 9/9 | 9/9 | 33.900s | 64.473s |
| Pi | 6/9 | 9/9 | 34.005s | 39.568s |
| OpenCode | 6/9 | 9/9 | 59.260s | 165.592s |
Isolated runs
Each cell received a fresh workspace and user/session state. Fixtures and source tasks were hashed before and after.
Position bias controlled
A Latin-square rotation placed every task/harness combination in each run position once.
What this supports
“Algo CLI ranked first in the reported balanced 27-run local comparison.”
What this does not support
Independent reproduction, universal superiority, meaningful speed leadership over Pi, or greater underlying model power. Raw run artifacts and competitor revisions are not yet public; the suite is also small and uses only one model and machine.
Open the aggregate JSON →