clm-0045

The KV dequant patch removes most of quantised KV's agentic cost, not just its speed cost: on identical seeded tasks, patched q8_0 takes +9.3% more turns than patched f16 (234 vs 214 over 11 paired tasks) where the stock build cost +39% (clm-0038). Reward is near-identical (10/11 vs 11/11). Separately, the non-termination marathons that were attributed to q8_0 strike f16 too — this run's 3.3-hour blowup was on f16 while q8_0 solved the same task in 18 turns — so task blowups look stochastic, not KV-caused.

First run under the fixed methodology: PATCHED build (ce7689f asserted), identical 12-task subset via --task-ids 0-6,8-12, --seed 42, --max-steps 200, protocol-checked, windows recorded by instrumentation (run-meta.jsonl), Qwen3.5-122B both sides (within-model lever test — simulator confound does not apply, clm-0043). ## Paired table (reward, turns, minutes) | task | patched f16 | patched q8_0 | |---|---|---| | 0 | 1.0, 20, 2.0 | 1.0, 12, 1.6 | | 1 | 1.0, 24, 2.0 | 1.0, 24, 3.2 | | 2 | 1.0, 29, 3.5 | 1.0, 22, 3.2 | | 3 | 1.0, 14, 1.2 | 1.0, 18, 1.7 | | 4 | 1.0, 12, 1.4 | 1.0, 24, 2.7 | | 5 | 1.0, 15, 2.1 | 1.0, 24, 3.3 | | 6 | 1.0, 10, 1.5 | 1.0, 12, 1.4 | | 8 | 1.0, 26, 3.3 | 1.0, 32, 3.8 | | 9 | 1.0, 12, 2.3 | 1.0, 16, 2.6 | | 10 | 1.0, 30, 6.3 | 1.0, 26, 11.9 | | 11 | 1.0, 22, 2.9 | **0.0**, 24, 3.3 | | **Σ common** | **214** | **234 (+9.3%)** | | 12 | unscored — killed at the arm's 4h bound after ~3.3h | **1.0, 18, 3.4** | ## The three findings **1. The patch closes ~75% of the agentic turn gap.** Stock: +39% turns (clm-0038, matched to the percentage point by energy, clm-0042). Patched: +9.3%. Direction on signs: q8 longer on 6 tasks, shorter on 3, ~tie 2 — consistent but weak at n=11; the residual may be real or may be noise. Combined with clm-0041 (patch = 42% energy saving on throughput work), the patched picture is: quantised KV costs ~nothing in speed, little in reward, and possibly a single-digit-percent turn overhead on agentic work. **2. Task blowups are not a q8_0 property.** Stock runs kept drawing multi-hour non-terminating tasks on q8_0 arms (4.4h, clm-0040), which fed a "quantised KV fails to terminate" narrative. Here f16 drew the blowup — 3.3h on task 12 without converging — while q8_0 finished it correctly in 3.4 minutes. Same model, same seed, same subset. Non-termination looks like a stochastic simulator-path phenomenon that any arm can draw, which also means single-arm wall-times are a poor basis for KV conclusions. **3. Reward stayed flat.** 11/11 vs 10/11 — the one q8 miss was a wrong answer at normal length (24 turns, user_stop), not a runaway. Within small-n noise. ## Scope · n=11 pairs, one domain, self-play (fine for a within-model lever), one seed. The +9.3% residual needs repeats (--num-trials) before it is a number rather than a direction. · Wall-clock asymmetry (f16 4h bound-cut vs q8 62 min) is dominated by the single f16 blowup and must NOT be read as f16-is-slower. · Per-task energy from the recorded windows to follow as eng- records; arm-level energy is marathon-skewed and deliberately not quoted here (clm-0042's lesson).

Note — the record's own working

Cited by — computed at build time, never stored