2026-10-10 local: initial 0.5B qualification server exited NTSTATUS 0xC0000409 during GPU discovery before any task inference; full owned-job cleanup confirmed. Vulkan explicitly disabled for subsequent planned preflight; prior incident record retained, no main-study retry or stop reset. No model output implicated.
0.5B qualification 2: returned valid bounded schema, but failed exact A/true/solved instruction. Initial fixture conflated instruction-following correctness with schema compatibility and failed before saving response; this lost response is disclosed. Corrected before evaluation: save raw first, report compliance separately, require valid finite schema and exact tokenizer counts. No evaluation selection or score changes.
Primary 1.5B failed two new guarded preflights because current host available RAM fell below the unchanged 1GiB reserve. No unrelated applications terminated. Main design changed BEFORE any evaluation outcome: compare general Qwen2.5-0.5B and Coder-0.5B if both qualify. Same-size variants; this is not evidence about model size. Larger model exclusions are resource-based, not evaluation-score-based.
Development completed 16/48. Kernel owned-job cleanup reported active=0, but immediate socket probe connected during listener release; batch conservatively stopped before assignment 17. Pre-freeze lifecycle fixed: wait up to 10s for port release, never adopt/kill other listeners. Continue only remaining 32 development assignments in separate retained phase; no prior outcomes rerun.
