b70-optimization-lab

Laguna M8 gather-sharded Phase-A preflight aborted

Date: 2026-07-24 America/Toronto

Status: terminal infrastructure abort before candidate import, allocation, primitive execution, timing, or result production. This is not a performance result and does not authorize Phase B.

Frozen identity

Both authorization files were frozen to 0444 in a 0555 directory and passed a separate full validation invocation before Phase A.

What happened

The coordinator wrote its one-shot consumption marker:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/.laguna-m8-phase-a-preflight-d8ec51df4ea69fcf7a34ec827a882e151412dca305f79eeb6266345bc0617cb4.json

That immutable marker records:

status=one_shot_gate_started
no_retry=true
candidate_imported=false
card_root_created=false

The coordinator then entered the preregistered 65-second continuous live-idle gate. The surrounding execution wrapper terminated the process after roughly 11 seconds. The journal contains only the expected self-observing xpu-smi processes during that interval. There was no child card runner, Torch/native candidate import, XPU allocation or primitive, model load, generation, campaign root, result, or Phase-B authorization.

Because the process was terminated outside the coordinator, it could not write its normal pre-campaign failure record. The consumed marker and absence of the declared campaign root are the durable fail-closed evidence.

Decision

Do not rerun, repair, or repacket this gather-sharded treatment. Its no-retry authorization was consumed, even though no candidate work occurred. Classify the lane as infrastructure_aborted_pre_candidate, with no speed or quality claim.

The approved record remains unchanged. The next research lane is the previously preregistered runtime-only breakable command graph: retain the incumbent eager Laguna kernels and BF16 boundaries, capture only target M=8, leave draft/prefill/widths 1-7 eager, and require raw-byte identity before any endpoint work.