Proxy cap at K=15 — the evidence run: regex-log 93%, openssl is model baseline (0.77)

What was run

The statistical run: the two signal tasks × 15 attempts each (30 trials) through the reasoning-cap proxy (port 8021, cap ~4000), qwen3.6-35b-a3b, thinking ON.

Results

Mean 0.77: regex-log 14/15 (93%), openssl-selfsigned-cert 9/15 (60%).

Interpretation

• The budget works. A task that reliably loops and fails uncapped passes ~95% (19/20 across this run and the K=5 run) with the cap. This is the evidence that motivated making the cap permanent.
• openssl stabilizes at ~60% — its short trajectories never reach the cap, so this is the model's capability baseline on that task, not a harness issue.
• 4000 was still an untuned first guess; a proper 2000/4000/8000 sweep on the full suite remains to be done.

Next steps

The proxy is an extra moving part. llama.cpp has a native --reasoning-budget N server flag that does the same job — switch the llama-swap config to it and verify parity, then retire the proxy.