{
  "schema_version": 1,
  "batch": "2026-10-21-same-bulletin-followup",
  "papers": [
    {
      "key": "agent-s4",
      "domain": "agent-research",
      "title": "Evolving Agent Interfaces",
      "paper_url": "https://arxiv.org/abs/2610.28601",
      "detail_path": "agent-research/2610.28601-agent-s4/README.md",
      "published_at": "2026-10-20T17:47:37Z",
      "priority": "P1",
      "component_implemented": true,
      "implementation_status": "component_only",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "same-bulletin-recent-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2610.28601",
        "scale": "Qwen3.8-27B, 30 evolutionary generations, 30/48 training and15 validation tasks, long GUI rollouts and cloud VMs; no full OSWorld evaluation.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    },
    {
      "key": "ask-to-distill",
      "domain": "post-training",
      "title": "Ask to Distill: Learning to Synthesize Data for On-Policy Distillation",
      "paper_url": "https://arxiv.org/abs/2610.27450v1",
      "detail_path": "post-training/2610.27450-ask-to-distill/README.md",
      "published_at": "2026-10-20T01:00:17Z",
      "priority": "P1",
      "component_implemented": true,
      "implementation_status": "component_only",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "same-bulletin-recent-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2610.27450v1",
        "scale": "Multi-model ask/answer/distill training and long on-policy generations; no teacher/student checkpoint or dataset downloaded for this component.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    },
    {
      "key": "projected-distillation",
      "domain": "post-training",
      "title": "Projected Distillation: On Learning to Solve Hard Problems from Close, Correct Targets",
      "paper_url": "https://arxiv.org/abs/2610.27467v1",
      "detail_path": "post-training/2610.27467-projected-distillation/README.md",
      "published_at": "2026-10-20T01:38:01Z",
      "priority": "P1",
      "component_implemented": true,
      "implementation_status": "component_only",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "same-bulletin-recent-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2610.27467v1",
        "scale": "Original eight-A100 training and paid API projection/verifier matrix; no full checkpoint training or projector service.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    },
    {
      "key": "race-rope-compression",
      "domain": "foundation-models",
      "title": "RACE: RoPE-Aligned Attention Compression Based on Query-Key Activation Energies",
      "paper_url": "https://arxiv.org/abs/2610.28385v1",
      "detail_path": "foundation-models/2610.28385-race-rope-compression/README.md",
      "published_at": "2026-10-20T15:34:02Z",
      "priority": "P1",
      "component_implemented": true,
      "implementation_status": "component_only",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "same-bulletin-recent-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2610.28385v1",
        "scale": "Original multi-billion-parameter model calibration, recovery training and long-context benchmark matrix; no complete checkpoint experiment. Native projection folding does not require a custom CUDA kernel.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    },
    {
      "key": "residency-rl",
      "domain": "agent-research",
      "title": "ResidencyRL: Reinforcement Learning in Simulated Clinical Environments",
      "paper_url": "https://arxiv.org/abs/2608.07418v1",
      "detail_path": "agent-research/2608.07418-residency-rl/README.md",
      "published_at": "2026-08-07T17:04:41Z",
      "priority": "P0",
      "component_implemented": false,
      "implementation_status": "deferred_high_cost",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "historical-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2608.07418v1",
        "scale": "Internal Gemini3.5 Flash GRPO, Gemini3.1 Pro simulation/judging,57,453 scenarios and up to60 dialogue turns plus8 tool calls; author code explicitly not released.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    },
    {
      "key": "stratified-sparse-attention",
      "domain": "foundation-models",
      "title": "Stratified Sparse Attention: Reducing the Sparse-to-Dense Gap at Extreme Sparsity",
      "paper_url": "https://arxiv.org/abs/2610.28441v1",
      "detail_path": "foundation-models/2610.28441-stratified-sparse-attention/README.md",
      "published_at": "2026-10-20T16:02:12Z",
      "priority": "P1",
      "component_implemented": true,
      "implementation_status": "component_only",
      "review_done": true,
      "full_experiment_status": "not_reproduced_high_cost",
      "intake_class": "same-bulletin-recent-backfill",
      "automatic_retry": false,
      "automatic_implementation_queue": false,
      "cost_assessment": {
        "source_url": "https://arxiv.org/abs/2610.28441v1",
        "scale": "Original multi-billion-parameter long-context model/kernel benchmark campaign; no checkpoint, throughput claim or serving system.",
        "available_resource": "One NVIDIA A100 or A30, or local Mac",
        "basis": "paper configuration, not local runtime measurement"
      }
    }
  ]
}