Q36 mechanics-only qualification

Status: PASS on 2026-08-13; no scientific graph was launched.

The sole authorized H100 allocation, Slurm job 754602 on eligible node evc35, completed in 497 seconds with exit 0:0 and zero restarts. The exact Qwen3.6-35B-A3B mechanics passed: all 693 weights loaded; the aligned causal draft intervention changed response state and native expert routing in all 40 layers; the draft-hidden control was invariant in state and routing; one finite source-only update completed; every protected router/expert byte was unchanged; checkpoint restore was exact; and sealed holdout, product, and public access stayed zero. The immutable result is Q36_MTR_MECHANICS_QUALIFICATION_RESULT_20260813.json.

Q36 reached two distinct no-score mechanics infrastructure failures before any draft, capability evaluation, or assessor access. A future full 61-request graph is therefore ineligible until the repaired mechanics path passes once on the exact pinned Qwen3.6-35B-A3B host.

The qualification is intentionally separate from the scientific phase:

• one normal-partition H100 and the unchanged node-exclusion set;
• --no-requeue, one immutable output root, and no retry;
• the same exact 24 B1 source-only rows, NF4 host, seeds, update, restore, causal draft intervention, native-router receipt, and protected-parameter checks as the frozen Q36 mechanics gate;
• both terminal-null receipts must be byte-exact and must attest zero scoring, zero assessor reads, and no scientific gate entry;
• no assessor, source-disjoint development row, holdout, product board, or public board is an input;
• no dispatcher or submission call exists in the wrapper; and
• PASS and failure both stop for evidence preservation. Neither authorizes the scientific graph or a successor.

pipeline/authorize_q36_mtr_mechanics_qualification.py builds a write-once authorization and independently verifies it inside the allocation before model staging. train/jobs/q36_mtr_mechanics_qualification.sbatch is the sole qualification wrapper. Its authorization explicitly records scientific_graph_authorized=false, capability_scoring_authorized=false, assessor_access_authorized=false, and submission_capability=false.

The authorization is published mode 0444. Allocation-time verification rehashes the exact runtime manifest, model manifest, environment receipt, B1 source, and both terminal-null receipts, and requires their resolved paths to match the authorization. Any missing, writable, substituted, or changed authorization—and any missing, substituted, or changed custody input—fails before model staging.

This boundary proved the repaired causal-generation mechanics on the real host without consuming or weakening the publication experiment. It changed no arm, prompt, threshold, seed, source identity, training budget, or terminal rule.