Recipe feasibility: glm53-k3-dflash2-dual-pro6000-wsl2-default

Required tokens: 258,192

Candidate	Classification	Estimated policy T_max	Policy VRAM margin	Reason
purtell-k3-dflash2-k5-fp8-1m-vision	benchmark-survivor	[1,154,604, 1,542,488]	[21.706, 28.706] GiB	resource bounds pass; behavioral evidence remains incomplete
brandon-v84-k4-dflash2-vision-98k	proven-infeasible	unknown	unknown	runtime context limit is below the required token budget
cardillo-k4-adaptive-mtp-replayssm-262k	empirically-disqualified	unknown	unknown	Local repeated structured-tool gate passed only 12/20 and produced degenerate repeated output.
0xsero-glm53-flash-exl3-3bpw	empirically-disqualified	unknown	unknown	Publisher-held-out quality evaluation rejected the 3.0 bpw quant; no complete supported dual-GPU server path was published.

Missing evidence

• purtell-k3-dflash2-k5-fp8-1m-vision: deterministic_pass_rate, no_spec_warm_e2e_seconds, quality_score, reference_quality_score, reference_tasks_per_hour, successful_tasks_per_hour, warm_e2e_seconds
• brandon-v84-k4-dflash2-vision-98k: deterministic_pass_rate, measured_max_stable_context_tokens, no_spec_warm_e2e_seconds, quality_score, reference_quality_score, reference_tasks_per_hour, successful_tasks_per_hour, warm_e2e_seconds
• cardillo-k4-adaptive-mtp-replayssm-262k: deterministic_pass_rate, measured_max_stable_context_tokens, no_spec_warm_e2e_seconds, quality_score, reference_quality_score, reference_tasks_per_hour, successful_tasks_per_hour, warm_e2e_seconds
• 0xsero-glm53-flash-exl3-3bpw: deterministic_pass_rate, measured_max_stable_context_tokens, no_spec_warm_e2e_seconds, quality_score, reference_quality_score, reference_tasks_per_hour, successful_tasks_per_hour, warm_e2e_seconds

Unbounded variables

• tracked_variables.wsl2_collective_compatibility
• candidates[0].metrics.deterministic_pass_rate
• candidates[0].metrics.quality_score
• candidates[0].metrics.reference_quality_score
• candidates[0].metrics.warm_e2e_seconds
• candidates[0].metrics.no_spec_warm_e2e_seconds
• candidates[0].metrics.successful_tasks_per_hour
• candidates[0].metrics.reference_tasks_per_hour
• candidates[1].per_token_components.unmodeled_kv
• candidates[2].per_token_components.unmodeled_kv
• candidates[3].runtime_context_limit_tokens
• candidates[3].per_token_components.unmodeled_kv

benchmark-survivor and math-qualified do not authorize production promotion.