Calibrated prune-guard confirmed — 4/5; prunes in 4/4 eligible trials at measured ratios 1.84–2.03; the fail is a turn-1 reasoning-budget force-close, unrelated to the guard

What was run

regex-log × 5, COMPACT_AT_TOKENS=15000, first run with the usage-calibrated token estimator (added after …140750 measured chars/4 undercounting real tokens 1.7–2.1× on this task's log-file content).

Results — calibration works exactly as designed

trial	result	agent	turns	peak ctx	prunes	ratio
CGUsswi	FAIL	22s	1	1.8k	0	—
fbMMmRj	PASS	4m13s	27	12.6k	22	1.84
QwGLFCh	PASS	1m51s	18	11.3k	11	2.03
tG3ZPWa	PASS	1m43s	13	12.3k	6	1.94
wAhEam3	PASS	1m35s	20	14.4k	4	4 → 1.85

• Measured ratios 1.84–2.03 — dead-center in the 1.7–2.1 band from the previous run's analysis; the calibration is reading the content correctly.
• Peak ctx now capped at 11.3–14.4k real tokens (limit 15000) vs 23–31k at the same threshold pre-calibration — the trigger now operates in real tokens, which is the property that keeps the 200k default safely under the 262144 server cap on dense-content tasks.
• Every multi-turn trial pruned (4/4 eligible); all ran full trajectories to a clean final answer. Combined with …140750: guard fully validated.

The failure: turn-1 reasoning-budget force-close (NOT the guard)

CGUsswi lasted one message: 10,384 chars of thinking cut off mid-regex (the tail ends inside (?:0[1-9]|1[0-2…), total output 4,002 tokens ≈ the server's --reasoning-budget 4000, then a single visible character (>), then a clean stop with zero tool calls. llama.cpp force-closed </think> while the model was mid-pattern; the model emitted one stray character and quit. No tool call → pi's loop correctly ended; no /app/regex.txt → the verifier fails on file-not-exists. The guard never engaged (peak ctx 1.8k, 0 prunes).

This is the second sighting of the budget force-close artifact — the first was the literal user role-leak ending in path-tracing-reverse (run …121319). Both: thinking truncated mid-thought → degenerate 0–1-char text → clean stop. Frequency so far ~1 in 10–15 trials on thinking-heavy turns. Worth tracking as its own failure mode ("budget-cut degenerate stop"); possible mitigations, none implemented: raise/sweep the reasoning budget (already a planned server experiment), or a harness-side "first message ended with no tool call and near-budget output → resend" nudge.

Next

Guard validation is complete. The real test at stock settings (COMPACT_AT_TOKENS=200000): the two untested suite 400-victims —

make smoke MODELS=qwen3.6-35b-a3b SMOKE_TASKS="circuit-fibsqrt make-mips-interpreter"

Success = zero 400 … exceeds the available context size, prune events around the 200k mark, trials run to completion (pass or fail on merit).