schema_version: 1
protocol_id: deermem-hybrid-v1-pr4789-reproduction-v1
required_policy_version: hybrid-v1
evaluation_time: "2026-08-13T00:00:00Z"

dataset:
  repository: xiaowu0162/longmemeval-cleaned
  revision: 98d7416c24c778c2fee6e6f3006e7a073259d48f
  filename: longmemeval_oracle.json
  sha256: 821a2034d219ab45846873dd14c14f12cfe7776e73527a483f9dac095d38620c

pool:
  size: 10
  capacities: [5, 7, 9]
  qa_capacity: 7
  distractors: 9
  distractor_bank_size: 40
  distractor_types: [single-session-user, single-session-preference]
  distractor_min_evidence_chars: 40
  distractor_max_evidence_chars: 700
  offset_namespace: deermem-medium-v1
  fact_order: fact-id

policies:
  confidence:
    policy: confidence
  hybrid_v1:
    policy: hybrid-v1
    weights:
      confidence: 0.65
      confirmation: 0.25
      access: 0.10
    confirmation_half_life_days: 90
    access_half_life_days: 30
    correction_reserved_fraction: 0.10
    correction_reserved_max: 10

qa:
  answer_prompt:
    path: prompts/answer-v1.txt
    sha256: 2bd4f166219946ccff7ac86f52a053c6dbf8be6d27e178c398d6bf70cf8a7423
  grader_version: deterministic-overlap-v1
  provider: openai-compatible
  api_key_env: DEERMEM_EVAL_ANSWER_API_KEY
  base_url_env: DEERMEM_EVAL_ANSWER_BASE_URL
  model: deepseek-v4-flash
  temperature: 0
  max_tokens: 2048
  stream: false
  timeout_seconds: 120
  max_attempts: 3
  workers: 3

statistics:
  bootstrap_seed: 4789
  bootstrap_iterations: 10000
  alpha: 0.05