{
 "generated_at": "2026-11-22T13:58:03Z",
 "items": [
  {
   "actual_outcome": "UNDERPOWERED",
   "contradicting": [
    "E142 sealed OW3: NOT SUPPORTED",
    "replication OW3: UNDERPOWERED"
   ],
   "expected_information_gain": "high: n doubles toward the 30 required",
   "next": "keep running; re-test at n \u2265 30",
   "prereg_outcome": "gated \u2212 scratch > 0",
   "proposed_experiment": "more ow_transfer@0.05 experiments on the production engine, same prereg",
   "question": "Does cross-domain transfer help at a very small budget (E142 OW3)?",
   "resource_estimate": "\u2248 25 experiments \u00d7 ~1 min",
   "supporting": [],
   "uncertainty": "high (replication underpowered: n=25)",
   "why": "transfer to a different domain is a core generality test"
  },
  {
   "actual_outcome": "E143 running",
   "contradicting": [
    "E142 OW5b: NOT SUPPORTED"
   ],
   "expected_information_gain": "high",
   "next": "audit E143 results when sealed",
   "prereg_outcome": "integrate > random",
   "proposed_experiment": "E143 diagnosis-driven loop (R1-R5) with a validation gate",
   "question": "Did E142's RSI lineage beat matched random search (OW5b)?",
   "resource_estimate": "40 experiments (E143)",
   "supporting": [],
   "uncertainty": "resolved negative in E142",
   "why": "separates a real improvement mechanism from best-of-N luck"
  },
  {
   "actual_outcome": "UNDERPOWERED",
   "contradicting": [
    "H1 losses 2"
   ],
   "expected_information_gain": "high until required n reached",
   "next": "diagnose before re-running (do not re-test the same design)",
   "prereg_outcome": "director \u2212 fixed > 0 (Holm)",
   "proposed_experiment": "e144_director",
   "question": "Does self-directed prioritisation beat a fixed research schedule?",
   "resource_estimate": "20 experiments",
   "supporting": [
    "H1 wins 2 of 7"
   ],
   "uncertainty": "CI [-0.071429, 0.142857]",
   "why": "Phase 1: research without a human choosing every next step"
  },
  {
   "actual_outcome": "UNDERPOWERED",
   "contradicting": [],
   "expected_information_gain": "high until required n reached",
   "next": "diagnose before re-running (do not re-test the same design)",
   "prereg_outcome": "meta \u2212 theta0 > 0 (Holm)",
   "proposed_experiment": "e144_meta",
   "question": "Does improving the improver yield more improvement on unseen domains?",
   "resource_estimate": "12 experiments",
   "supporting": [
    "H2 wins 1 of 5"
   ],
   "uncertainty": "CI [0.0, 0.004371]",
   "why": "Phase 5 outcome B: the improvement mechanism itself improving"
  },
  {
   "actual_outcome": "UNDERPOWERED",
   "contradicting": [],
   "expected_information_gain": "high until required n reached",
   "next": "diagnose before re-running (do not re-test the same design)",
   "prereg_outcome": "calibrated \u2212 constant > 0 (Holm)",
   "proposed_experiment": "e144_selfmodel",
   "question": "Can the agent predict its own performance on unseen worlds?",
   "resource_estimate": "20 experiments",
   "supporting": [
    "H3 wins 7 of 7"
   ],
   "uncertainty": "CI [0.014799, 0.025333]",
   "why": "Phase 8: knowing what it does not know"
  },
  {
   "actual_outcome": "NOT SUPPORTED (treatment worse)",
   "contradicting": [
    "aligned vs scratch: NOT SUPPORTED (treatment worse)",
    "aligned == naive on every row: the alignment step fell back to index mapping"
   ],
   "expected_information_gain": "high: turns a negative into a diagnosis with a fix",
   "next": "E145: fix alignment timing and re-test; do not re-run the same no-op arm",
   "prereg_outcome": "aligned > scratch",
   "proposed_experiment": "delay alignment until the target has rules; align on facts/observations, then re-test on the same split",
   "question": "Why did cross-interface world-model transfer fail?",
   "resource_estimate": "O1 re-run, about 15 minutes",
   "supporting": [
    "model beats no model: SUPPORTED"
   ],
   "uncertainty": "mechanism identified, not fixed",
   "why": "O1 is the test of whether a model learned in one system can be reused in another"
  },
  {
   "actual_outcome": "NOT SUPPORTED (treatment worse)",
   "contradicting": [
    "vs scratch: NOT SUPPORTED (treatment worse)"
   ],
   "expected_information_gain": "medium-high",
   "next": "E145: held-out mix selection; keep the default as the control",
   "prereg_outcome": "capability > scratch",
   "proposed_experiment": "select the mix on a held-out source split and add a cross-family validation gate",
   "question": "Is the selected capability better than a good default on unseen families?",
   "resource_estimate": "O2 re-run, about 20 minutes",
   "supporting": [
    "beats random: SUPPORTED",
    "beats retrieval: SUPPORTED",
    "beats no-model: SUPPORTED"
   ],
   "uncertainty": "source-only selection did not predict target value",
   "why": "O2 tests whether a capability graph transfers, not just whether it works"
  },
  {
   "actual_outcome": "NOT SUPPORTED",
   "contradicting": [
    "vs none: NOT SUPPORTED"
   ],
   "expected_information_gain": "medium",
   "next": "E145: longer lineage; report forgetting curves",
   "prereg_outcome": "full > none",
   "proposed_experiment": "4-6 generations, more candidates, and a memory-ablation on held-out families",
   "question": "Does research memory change behaviour over more than two generations?",
   "resource_estimate": "about 1 hour",
   "supporting": [
    "beats retrieval-only: SUPPORTED"
   ],
   "uncertainty": "short lineage; memory records may be too few",
   "why": "O3 is the continual-learning test; two generations may be too short"
  },
  {
   "actual_outcome": "SUPPORTED",
   "contradicting": [
    "discovery used about 1.9x the learning interventions of the random arm"
   ],
   "expected_information_gain": "high for the scientific-discovery claim",
   "next": "E145: matched-budget replication; do not present O4 as multi-discipline",
   "prereg_outcome": "discovery > matched random",
   "proposed_experiment": "give the random arm the same intervention budget and add a second causal structure (3-input conjunctive, delayed effect)",
   "question": "Does causal discovery survive a compute-matched random baseline and a second rule class?",
   "resource_estimate": "about 30 minutes",
   "supporting": [
    "exact recovery SUPPORTED",
    "held-out prediction SUPPORTED"
   ],
   "uncertainty": "one domain, non-compute-matched control",
   "why": "O4 is E144's strongest positive; it needs a stronger control and a broader law class"
  },
  {
   "actual_outcome": "NOT SUPPORTED (treatment worse)",
   "contradicting": [
    "held-out collective vs single: NOT SUPPORTED (treatment worse)"
   ],
   "expected_information_gain": "medium",
   "next": "E145: scale + independent verifier config; publish the negative if it repeats",
   "prereg_outcome": "collective > single",
   "proposed_experiment": "scale specialists, give the verifier its own model/config, and measure quality as well as speed",
   "question": "Does a verified collective beat a compute-matched individual?",
   "resource_estimate": "about 1 hour",
   "supporting": [
    "independent verifier rejected 1 specialist claims"
   ],
   "uncertainty": "4 specialists and one verifier may be too small",
   "why": "O5 tests the collective-intelligence claim and independent verification"
  },
  {
   "actual_outcome": "NOT SUPPORTED",
   "contradicting": [
    "active design vs random: NOT SUPPORTED (both 1.0)"
   ],
   "expected_information_gain": "medium",
   "next": "E145: harder law class; keep the falsification pass",
   "prereg_outcome": "discovery > random",
   "proposed_experiment": "piecewise or sparse law class where random points leave the law underdetermined",
   "question": "Can the organism recover a mathematical law, and does active query design help?",
   "resource_estimate": "about 20 minutes",
   "supporting": [
    "held-out prediction SUPPORTED",
    "exact recovery 1.0"
   ],
   "uncertainty": "three points determine a quadratic; the law class is too easy for active design",
   "why": "O6 is the second discovery discipline"
  }
 ]
}
