{
 "backlog": [
  {
   "approach": "ingest E113 raw rows",
   "cost": 1,
   "evidence": "G11 delegated to E113 (model-backed arms A-J), not yet published",
   "experiment": "re-verify E113 gate from raw rows with the E114 independent verifier",
   "failure": "they do not",
   "id": "B3",
   "importance": 0.9,
   "limitation": "no model-backed collective result inside E114",
   "priority_score": 0.81,
   "resources": "E113 publication",
   "risk": 0.0,
   "success": "E113 claims reproduce",
   "uncertainty": 0.9
  },
  {
   "approach": "context-gated skill invocation (only when task features match the skill's training support)",
   "cost": 1,
   "evidence": "{\"no_skills\": 0.6, \"with_skills\": 0.35, \"wins\": 0, \"losses\": 10, \"mean_evals_no_skills\": 1842.75, \"mean_evals_with_skills\": 1903.475}",
   "experiment": "gated vs ungated library on shift and sealed",
   "failure": "sealed gain lost",
   "id": "B2",
   "importance": 0.8,
   "limitation": "skills hurt on shifted task families",
   "priority_score": 0.51,
   "resources": "CPU, ~20 min",
   "risk": 0.05,
   "success": "no shift loss, sealed gain retained",
   "uncertainty": 0.7
  },
  {
   "approach": "shared-trunk + expert heads with replay",
   "cost": 1,
   "evidence": "zero-shot mean (experts) 0.573 ~ chance",
   "experiment": "trunk+heads vs experts, sealed seeds",
   "failure": "no transfer",
   "id": "B5",
   "importance": 0.6,
   "limitation": "continual learners show no forward transfer",
   "priority_score": 0.28,
   "resources": "CPU, ~10 min",
   "risk": 0.02,
   "success": "zero-shot > 0.55 with forgetting <= 0.05",
   "uncertainty": 0.5
  },
  {
   "approach": "learn rule grammar extensions (program synthesis over rule DSL) instead of fixed hypothesis class",
   "cost": 2,
   "evidence": "shift Brier causal 0.117 vs mlp 0.075",
   "experiment": "grammar-growth arm vs fixed grammar on sealed shift worlds",
   "failure": "no shift gain or in-class regression",
   "id": "B1",
   "importance": 0.9,
   "limitation": "world models degrade under out-of-class shift",
   "priority_score": 0.22,
   "resources": "CPU, ~1 h",
   "risk": 0.05,
   "success": "shift Brier of causal <= mlp with in-class Brier unchanged",
   "uncertainty": 0.6
  },
  {
   "approach": "run official protocols with a local judge after E113",
   "cost": 4,
   "evidence": "datasets + LLM judge not available offline; CPU shared with E113",
   "experiment": "AMA-Bench synthetic subset with official scorer",
   "failure": "protocol cannot be reproduced",
   "id": "B4",
   "importance": 0.8,
   "limitation": "external memory benchmarks not run (AMA-Bench, LongMemEval, LoCoMo)",
   "priority_score": 0.13,
   "resources": "dataset download, local model, several CPU hours",
   "risk": 0.05,
   "success": "official-protocol score reported with version",
   "uncertainty": 0.9
  },
  {
   "approach": "run the mission engine under a supervisor for 24 h",
   "cost": 3,
   "evidence": "mission elapsed 16.09 s",
   "experiment": "24 h governed mission with scheduled kills",
   "failure": "any duplicate effect or lost decision",
   "id": "B6",
   "importance": 0.7,
   "limitation": "missions run minutes, not 24 h",
   "priority_score": -0.0067,
   "resources": "24 h host time after E111 endurance finishes",
   "risk": 0.1,
   "success": "continuity checks hold after >=3 kills over 24 h",
   "uncertainty": 0.4
  }
 ],
 "questions": {
  "Q1": {
   "depends_on": [],
   "hypotheses": [
    "H1"
   ],
   "id": "Q1",
   "objective": "O1",
   "text": "Does EIG experiment selection beat random selection on sealed worlds?"
  },
  "Q2": {
   "depends_on": [
    "Q1"
   ],
   "hypotheses": [
    "H2"
   ],
   "id": "Q2",
   "objective": "O1",
   "text": "Does a prior from validated memory of earlier worlds reduce experiments further?"
  },
  "Q3": {
   "depends_on": [
    "Q2"
   ],
   "hypotheses": [
    "H3",
    "H3s"
   ],
   "id": "Q3",
   "objective": "O1",
   "text": "Does that memory prior transfer to a different environment size, and does it hurt under shift?"
  },
  "Q4": {
   "depends_on": [],
   "hypotheses": [],
   "id": "Q4",
   "objective": "O1",
   "text": "Which world model predicts interventions best in- and out-of-distribution?"
  },
  "Q5": {
   "depends_on": [],
   "hypotheses": [
    "H5"
   ],
   "id": "Q5",
   "objective": "O2",
   "text": "Which continual learner retains best per unit compute?"
  },
  "Q6": {
   "depends_on": [],
   "hypotheses": [],
   "id": "Q6",
   "objective": "O3",
   "text": "Do mined skills raise held-out solve rate; does environment diversity matter?"
  },
  "Q7": {
   "depends_on": [
    "Q4"
   ],
   "hypotheses": [
    "H7",
    "H7b"
   ],
   "id": "Q7",
   "objective": "O4",
   "text": "Can a candidate configuration beat the frozen world-model baseline on sealed data?"
  },
  "Q8": {
   "depends_on": [
    "Q6"
   ],
   "hypotheses": [
    "H8"
   ],
   "id": "Q8",
   "objective": "O3",
   "text": "Does context-gating make a broader skill library safe under shift? (backlog B2)"
  },
  "Q9": {
   "depends_on": [
    "Q4"
   ],
   "hypotheses": [
    "H9"
   ],
   "id": "Q9",
   "objective": "O1",
   "text": "Does growing the rule grammar after misspecification fix shift prediction? (backlog B1)"
  }
 },
 "schema": "cain.e114.backlog.v1",
 "source": "results.director",
 "uncertainty": {
  "H1": {
   "mean": 0.9569,
   "sd": 0.0188,
   "status": "SUPPORTED_HYPOTHESIS"
  },
  "H2": {
   "mean": 0.6,
   "sd": 0.0655,
   "status": "UNSUPPORTED_HYPOTHESIS"
  },
  "H3": {
   "mean": 0.75,
   "sd": 0.0804,
   "status": "SUPPORTED_HYPOTHESIS"
  },
  "H3s": {
   "mean": 0.5,
   "sd": 0.2887,
   "status": "UNSUPPORTED_HYPOTHESIS"
  },
  "H5": {
   "mean": 0.5,
   "sd": 0.2887,
   "status": "STALLED"
  },
  "H7": {
   "mean": 0.4715,
   "sd": 0.0448,
   "status": "UNSUPPORTED_HYPOTHESIS"
  },
  "H7b": {
   "mean": 0.3048,
   "sd": 0.0447,
   "status": "UNSUPPORTED_HYPOTHESIS"
  }
 }
}