regex-log × 5 at budget 8000 — 5/5, force-close artifact GONE on this task

What was run

regex-log × 5 at stock settings, first K>1 validation of the reasoning-budget switch 4000 → 8000 (~19:00 same day). Direct comparison against …180910 (same task × 5 at budget 4000: 3/5, both fails were force-close → plain-text thinking-continuation runaways).

Results

• 5/5 PASS, trial times 1m32s–3m03s (vs 2m50s–10m passes plus 25–60 min runaway fails at 4000).
• Zero length-stops across all 5 trials (was 2/5 fatal + 1 lucky 62k-char survivor at 4000).
• Largest think blocks: 20.1–22.2k chars ≈ 6–7k tokens — over the old 4000 cap, under the new 8000 one. regex-log's natural reasoning length simply fits now; thinking completes with a proper </think> and the model comes out clean every turn.
• Guard silent as expected (peaks 19.6–35k), no prunes, no recoveries.

Interpretation

This is the cleanest possible confirmation of the force-close hypothesis: the task whose loops motivated the entire reasoning-budget mechanism goes 6/6 today (counting …192605) once the budget exceeds its natural thinking length. The artifact was never about the task being hard — it was the mid-thought </think> injection.

Caveat from the sibling run …205241: 8000 only helps tasks whose thinking fits. polyglot-rust-c maxes any budget and still produces the (milder, self-terminating) continuation artifact. Keep 8000; attack the residual with the extended runaway recovery, not with more budget.