Home › Evidence › Records › clm-0121
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

clm-0121

measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0121/ — this address never moves; the anchor /records/#clm-0121 keeps resolving

A 45-turn Vulkan cache-reuse trace completed without a logged DeviceLost or lockup signature. Each returned row records only the newly prefetched prompt tokens, which remain approximately two thousand per turn, consistent with prefix reuse. The handoff describes the conversation as about 90k tokens, but its machine-readable trace does not record cumulative context; this is not a precise depth-capability boundary or production-latency result.

verified 2026-08-24 · volatility medium
evidence cfg-0175 run-0628

Note — the record's own working

Source: the autonomous handoff's 45-turn stdout, server log, and reproduction script. The script labels its displayed depth as prompt_n for the current turn, not total retained context. It provides neither a standard guard nor independent answer grading, and no energy window. Cold-load failure remains a separate result.

Cited by — computed at build time, never stored

model pages qwen38-27b
candidate gate history qwen38-27b