{ "id": "O0-CRP-021", "title": "Functional Separation Test — Auxiliary O/0 claim inverts under fair-information control", "semantic_name": "Functional Separation", "record_class": "SIMULATION", "version": "1.0.0", "status": "INCONCLUSIVE (primary) / UNSUPPORTED (fair-comparison)", "evidence_level": "computational_simulation", "replication_status": "not_replicated", "program": "track_1_auxiliary_o0_claim", "non_drift_question": "auxiliary_functional_claim", "opens_branch": "Track 1 — auxiliary O/0 claim test", "date": "2026-07-28", "source_record": "O/0 Framework §VII / §VIII", "related_records": ["O0-CRP-022", "O0-CRP-023"], "source_proposition": "O/0 §VII: 'Because functional separation serves evolutionary/practical purposes. The illusion of separation allows local processing (brains solving local problems).' §VIII: 'The boundary is functional, not fundamental.'", "operationalization": { "SEPARATED_agent": "Kalman filter maintaining explicit posterior over hidden world state", "UNIFIED_agent": "Vector autoregression on observation history, no latent state", "task": "multi-step prediction in partially-observable linear-Gaussian dynamical system", "fair_control": "Agent A' — Kalman filter with matrices learned via subspace identification from training data (same info as Agent B)" }, "design_summary": { "world_class": "linear-Gaussian state-space", "hidden_dim": 8, "observation_dims_tested": [2, 4, 8], "observability_ratios": [0.25, 0.5, 1.0], "horizons": [1, 5, 20], "n_seeds": 30, "seed_range": [12000, 12029], "T_train": 500, "T_pred": 200, "predictors": ["A (privileged Kalman)", "A' (learned Kalman)", "B (VAR)", "B+ (2× VAR)", "persistence", "zero", "oracle"], "total_evaluations": 1620 }, "primary_verdict": { "label": "INCONCLUSIVE", "reason": "Direction consistent (Agent A wins 100% of seeds in most regimes at p < 1e-10), but effect sizes 2-5% NMSE fall below preregistered 5% magnitude threshold", "delta_5pct_pass": false, "p_pass": false, "null_pass": true, "provisional_pass": true }, "fair_comparison_verdict": { "label": "UNSUPPORTED", "reason": "Direction REVERSES. Agent B (unified VAR) beats Agent A' (learned Kalman) by 3-24% NMSE in every regime, p < 0.01 in 7/9 cells.", "key_finding": "Agent A's advantage in primary comparison was due to privileged causal-structure knowledge, not structural self/world separation." }, "scientific_reading": { "supported_by_data": [ "Kalman filter with correctly-specified matrices beats matched-parameter VAR by 2-5% NMSE, direction consistent across all regimes and horizons", "This advantage is fragile: it disappears and reverses when both agents must learn from the same training data", "In the fair comparison, unified VAR outperforms learned-Kalman by 3-24% NMSE across all regimes" ], "not_supported_by_data": [ "Self/world separation, per se, is functionally valuable at matched capacity", "The O/0 auxiliary claim from §VII (that appearance of separation aids local processing) under this operationalization", "That biological cognition benefits from Kalman-filter-style separation architectures" ], "not_addressed": [ "The deep O/0 identity claim (observer = source)", "Nonlinear or higher-dimensional regimes", "Alternative separation architectures (active inference, predictive coding)", "Alternative system-identification methods" ] }, "regime_results_summary": { "primary_A_vs_B_deltas_mean": { "k/d=0.25,tau=1": -0.045, "k/d=0.25,tau=5": -0.051, "k/d=0.25,tau=20": -0.021, "k/d=0.5,tau=1": -0.028, "k/d=0.5,tau=5": -0.023, "k/d=0.5,tau=20": -0.011, "k/d=1.0,tau=1": -0.021, "k/d=1.0,tau=5": -0.020, "k/d=1.0,tau=20": -0.025 }, "fair_Alearned_vs_B_deltas_mean": { "k/d=0.25,tau=1": +0.059, "k/d=0.25,tau=5": +0.026, "k/d=0.25,tau=20": +0.031, "k/d=0.5,tau=1": +0.135, "k/d=0.5,tau=5": +0.078, "k/d=0.5,tau=20": +0.058, "k/d=1.0,tau=1": +0.240, "k/d=1.0,tau=5": +0.135, "k/d=1.0,tau=20": +0.058 } }, "adversarial_notes": [ "The result depends on specific choice of system identifier (subspace SVD). A better estimator might close the gap for Agent A'.", "Linear-Gaussian world may be too simple; nonlinear registered as R1 followup.", "Ridge regularization for both Kalman-Q/R estimation and VAR fitting was chosen conservatively; sensitivity analysis is a followup.", "The Oracle predictor confirms Agent A tracks the theoretical ceiling in partial-observability regimes." ], "replication_queue": [ {"id": "O0-CRP-021-R1", "kind": "nonlinear_extension", "target": "Lorenz-63 partial-observation with neural agents", "status": "REGISTERED"}, {"id": "O0-CRP-021-R2", "kind": "system_id_alternatives", "target": "EM / VBI / deep SSM as Agent A'", "status": "REGISTERED"}, {"id": "O0-CRP-021-R3", "kind": "training_length_sweep", "target": "T_train ∈ {500, 2000, 5000, 20000}", "status": "REGISTERED"}, {"id": "O0-CRP-021-R4", "kind": "high_dimension", "target": "d ∈ {16, 32, 64}", "status": "REGISTERED"} ], "figures": [ {"file": "figures/01_headline_primary_vs_fair.png", "caption": "Primary vs fair comparison across regimes and horizons. Direction inverts under fair control."}, {"file": "figures/02_absolute_nmse.png", "caption": "Absolute NMSE for all six predictors across observability ratios."}, {"file": "figures/03_per_seed_scatter_tau5.png", "caption": "Per-seed scatter at tau=5 showing win/lose distribution."}, {"file": "figures/04_effect_vs_horizon.png", "caption": "Effect size trajectories vs prediction horizon."} ], "files": { "preregistration": "preregistration.md", "scientific_record": "scientific_record.md", "code": ["src/run_study.py", "src/analyze.py"], "results": ["results/summary.json"], "raw_data": ["results/raw/trials.jsonl"] }, "provenance": { "philosophical_source": "O/0 Framework §VII, §VIII", "exact_proposition_tested": "Functional separation serves local processing", "conceptual_relationship": "formalization_of_auxiliary_claim", "disclaimer": "Conceptual provenance is not empirical support. Result: auxiliary claim unsupported under fair comparison; deep identity claim not tested here (see O0-CRP-022)." }, "runtime_seconds": 21.3 }