{
  "schema": "cadence-two-binding-heads/1",
  "declared_date": "2026-10-22",
  "scope": "Synthetic component: two independently required shared-token bindings retrieved in one final settle; no learned semantic-code or integrated-life claim.",
  "source": {
    "baseline": "1ce3f2d300f3070feb9b975a5232b9a2485f1dcc",
    "freeze": "Commit protocol and all executed runner/anatomy/teaching bytes before the first behavioral run; record actual commit and source hashes in each receipt. Baseline is provenance, not an assertion that the uncommitted heads prototype already exists there."
  },
  "development_seeds": [0, 1],
  "confirmation_seeds": [200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219],
  "confirmation_required": 18,
  "confirmation_status": "unopened; requires all development gates and a separately declared total budget",
  "task": {
    "inputs": 19,
    "names": ["A", "B"],
    "shared_values": [0, 1, 2, 3],
    "markers": ["IS", "NEUTRAL", "AMBIGUOUS_CUE", "ASK_AB", "ASK_BA"],
    "coordinates": {"markers": [0, 1, 2, 3, 4], "names": [5, 6], "values": [7, 8, 9, 10], "exposed_A_values": [11, 12, 13, 14], "exposed_B_values": [15, 16, 17, 18]},
    "facts": "IS + name + value; two actual fact events, both orders balanced, then one neutral event",
    "query": "AMBIGUOUS_CUE + ASK_AB or ASK_BA; no values or intermediate address",
    "output_slots": [4, 4],
    "answers": "(value_A,value_B) for ASK_AB; (value_B,value_A) for ASK_BA",
    "held_out_value_pairs": [[0, 1], [1, 0], [2, 3], [3, 2]],
    "training_pairs": "All remaining 12 ordered pairs; all equal pairs and every value under each name remain in training",
    "training_cases": 48,
    "held_out_cases": 16,
    "neutral_events": 1,
    "rounds": 512,
    "batch": 16,
    "schedule": "SeedSequence([622022, founder]); each round samples 16 training cases with replacement; every arm uses identical draws",
    "evaluation": "Enumerate all training and held-out cases once; exact pair, each slot, unequal pairs and each query reported; refused/unrun answers count wrong",
    "swapped_panel": "Exchange values attached to A/B with query, fact order and token multiset fixed; held-out swaps stay held out; equal-value swaps are shams"
  },
  "brain": {
    "modules": [32], "episodic": false, "lateral": 0.0,
    "recall_next": false, "recall_address": 0.0, "recall_null": 0.0,
    "recall_recency": 0.0, "recall_amplitude": 3.0, "recall_temperature": 0.25,
    "teaching": "normalized", "learning_qualified": true, "learning_damping": 0,
    "learning_free_steps": 1024, "learning_nudged_steps": 1024,
    "learning_beta": 0.1, "learning_temperature": 0.2,
    "learning_tolerance": 0.003, "learning_momentum": 0.9
  },
  "dt": 0.1,
  "teaching": "Capture complete pre-query drive and immutable bank before greedy act; wrong-pair rows only, both final labels supplied; no head role, slot index, address, intermediate value or relevance mask taught. Corresponding rows of captured drive and bank remain together through free/nudged phases.",
  "primary_boundary": "Clear working trace and warm neural state after prefix before query in training/evaluation, keep only witnessed memory bank. Live query/read/association/motor coordinates all qualify jointly. Intact-trace comparison is later and separate.",
  "heads": "Shared bank; disjoint live query/read ports; same complete query initial drive; query-cortex projections zero initially, ordinary reciprocal read-cortex seams independently drawn. Total coupling3 shared equally over independent heads; no supplied head-to-name assignment.",
  "arms": [
    {"name": "exposed", "kind": "plain_Brain", "modules": [32], "input": "both unselected named values in dormant control coordinates"},
    {"name": "supplied_reads", "heads": 2, "modules": [32], "memory_energy": "disabled", "input": "Read ports receive 3 * actual_fact_A and 3 * actual_fact_B by fixed name, independent of query order. Witnessed snapshot kept for custody. Oracle/content exposure only, never a candidate selector or address lesson."},
    {"name": "H1", "heads": 1, "modules": [32], "coupling_per_head": 3.0},
    {"name": "H2", "heads": 2, "modules": [32], "coupling_per_head": 1.5},
    {"name": "H1-width", "heads": 1, "modules": [64], "coupling_per_head": 3.0},
    {"name": "H1-port-copies", "selectors": 1, "port_copies": 2, "modules": [32], "total_coupling": 3.0, "representation": "Duplicate keys/values/cues scaled by 1/sqrt(2); same total query/read coordinates as H2, one selector, unchanged match/read norm"},
    {"name": "H1-work", "heads": 1, "modules": [32], "coupling_per_head": 3.0, "maximum_sweeps_per_phase": 2048, "comparison": "Measured full neuron+transport work matched to H2; unmatched work gives no efficiency/head-specific claim"},
    {"name": "frozen-shuffled", "kind": "acquired-copy selection-interface interventions and frozen newborn controls"},
    {"name": "uniform_random", "seed": 622023, "expected_exact_pair": 0.0625}
  ],
  "candidate_order": ["agreement", "linear"],
  "maximum_candidate_rounds": 2,
  "candidate_selection": "One equation round includes all heads/resource controls; try linear only after retained agreement failure. No per-arm parameter grid or tuning on held-out/confirmation. Keep simpler H1 if it passes.",
  "gates": {
    "exposed_acquisition_floor": 0.95,
    "exposed_panel": "All 48 training cases: exact pair, each slot, unequal-value pairs and each query; both development founders. The 16 held-out cases are reported separately.",
    "content_use_floor": 0.95,
    "content_use": "supplied_reads uses the H2 anatomy with memory energy off, exposing both actual fact events by fixed A/B identity at its read ports, independent of query order. Require .95 exact pair, each slot, unequal-value pairs and each query on the 48 training cases, on both founders; report the 16 held-out cases separately. Final-answer teacher only; no candidate address lessons. H1 is not given two independent facts through an invented extra read.",
    "held_out_exact_pair_floor": 0.90,
    "swapped_unequal_exact_pair_floor": 0.90,
    "relevant_record_erasure_slot_loss_minimum": 0.20,
    "donor_pair_follow_minimum": 0.80,
    "sham_or_record_reordering_answer_change_maximum": 0.05,
    "retention_loss_maximum": 0.05,
    "retention_rounds": 256,
    "retention": "Continue balanced training on the same 12 pairs, then read the held-out panel; retain .95 on the training panel. No competing-task or P7 claim.",
    "causal_controls": ["queried-record erasures", "irrelevant-neutral-record change", "both-values donor transplant", "one-head erasures", "duplicated heads", "head permutation with downstream transport custody explicit", "frozen/shuffled acquired selection interface"],
    "promotion": "All development gates must pass on both founders 0 and 1 before confirmation; retain failures, refusals and caps. No library or default promotion from this experiment."
  },
  "representation_phase": {
    "start": "Only after primary acquisition/content/retrieval/causal gates pass",
    "variable": "Free settled association activation, not context-free held LearnedCode output",
    "probe": "Two diagnostic ridge readouts to one-hot requested values, fit on trainingpairs only; centered features, ridge0.01; freeze before held-out and swapped panels; never installed as answer path or teaching rule",
    "pair_transfer_floor": 0.90,
    "gain_over_raw_query_and_frozen_newborn_minimum": 0.20,
    "loss_after_relevant_memory_erasure_minimum": 0.20,
    "claim": "Task-context accessibility and held-out readout transfer, not general meaning",
    "controls": ["current raw query only", "fixed random observation code separately labeled if run", "frozen newborn association", "record erasure", "same ambiguous cue with swapped witnessed facts"],
    "later_semantic_gate": "Separately freeze learned-code recipe/source/budget and compare raw, fixed-random, learned, no-top-down; require held-out alias/group members without raw overlap, misleading raw overlap, member-specific facts and early facts after competition. Geometric shifts alone do not pass."
  },
  "budget": {
    "serial_local_processes": 1, "training_seconds_per_arm": 60,
    "development_total_host_seconds": 600,
    "charged": "All training, evaluation, oracle diagnostics, retention/probes, attempted phases/transports and refusals; deadlines between bounded trials",
    "stop": "Capped/unrun cells fail to establish gate and stay in denominator; separately declare later-stage budget before runs if total will not fit; no paid/cloud campaign"
  },
  "preservation": "Retain previous failed admission/coupled-reader receipts and existing gates. Reward custody, private imagination, complete checkpoint continuation, continuing context/recovery and learned semantic codes remain separate requirements before integration."
}
