docs/development-artifacts/hebbian-interface-v1/development-decision.json

{
  "decided_utc": "2026-09-29T23:51:33.901614+00:00",
  "development_seed": 142941,
  "development_result": "results/hebbian-interface-v1/development/142941/summary.json",
  "decision": "Keep configuration unchanged; run all three prespecified independent confirmation worlds. No search for a CE/MSE gap.",
  "evidence": {
    "all_four_memories_full_vocab_accuracy": [
      1.0,
      1.0,
      1.0,
      1.0
    ],
    "A_reader_accuracy_by_1000_steps": {
      "ce": 1.0,
      "mse": 1.0
    },
    "final_crossed_B_accuracies": {
      "ce": {
        "B_ce": 1.0,
        "B_mse": 1.0
      },
      "mse": {
        "B_ce": 1.0,
        "B_mse": 1.0
      }
    },
    "wrong_mapping_changed_accuracy": {
      "ce": 0.0,
      "mse": 0.0
    }
  },
  "interpretation": "Standalone vector fidelity is not necessary for successful single-hop use in this normalized, low-load task. This does not evaluate composition or establish a general loss-function advantage.",
  "formal_selection": "Fixed final 4000-step reader checkpoints; no new-mapping evaluation during reader optimization or checkpoint selection."
}