{ "id": "O0-CRP-035", "title": "SIM-CRP-003 · Perspective Distillation — What survives when an observer is compressed", "semantic_name": "SIM-CRP-003", "record_class": "SIMULATION", "version": "1.2.0", "status": "INCONCLUSIVE (primary) · WEIGHTS-CARRY-IDENTITY (secondary, v1.2)", "evidence_level": "computational_simulation", "replication_status": "internally_replicated_by_sister_study_CRP-036", "program": "contact_and_revelation", "non_drift_question": "PERSPECTIVE", "opens_branch": "PERSPECTIVE", "date": "2026-07-30", "source_record": "OA-O001", "supersedes_versions": ["1.0.0", "1.1.0"], "design_correction_v1_2": { "reason": "v1.0/v1.1 observer had w shape (d,), so x @ w was a scalar broadcast to all action logits; softmax shift-invariance made w behaviorally invisible.", "fix": "v1.2 uses per-action weight matrix W in R^{d x K}; q(x)[a] = x . W[:, a] + p[a] is genuinely context-dependent per action.", "detected_by": "O0-CRP-036 LA1 level (preferences only, w=0) gave BC_TV=1.000 in v1.1 observer, mathematically impossible if observer were truly context-dependent.", "v1_1_findings_status": { "F1_memory_no_op": "REPLICATED under v1.2 (design-invariant, decision rule does not use memory)", "F2_preferences_dominate_weights": "REVERSED under v1.2 - was a mathematical tautology, not empirical", "F3_argmaxP_bit_lifts": "REFRAMED under v1.2 - lift is due to per-action W structure, not argmax(p) alone" } }, "related_records": [ "O0-CRP-001", "O0-CRP-004", "O0-CRP-011", "O0-CRP-013", "O0-CRP-014", "O0-CRP-015", "O0-CRP-016", "O0-CRP-018", "O0-CRP-036", "O0-CRP-037" ], "design_summary": { "n_observers_confirmatory": 30, "observer_seeds": "12000..12029", "env_seeds_train": "12500..12529", "eval_context_seed": 13000, "n_eval_contexts": 200, "probe_seed": 13500, "n_probe_pairs": 500, "vp_env_seeds": "14000..14049", "n_vp_environments": 50, "n_vp_episodes": 20, "context_dim": 8, "n_actions": 4, "observer_state_v1_2": "(W in R^{d x K}, mu_0 in R^d, eta, tau, p in R^K, memory, id)", "compression_levels": [0, 1, 2, 3, 4, 5, 6, 7], "tau_range_v1_1": [0.15, 0.50] }, "preregistered_metrics_v1_1": { "primary": ["BC_TV", "BC_AM"], "auxiliary": ["BC_cos"], "additional": ["SR", "VP", "PS", "Rec"] }, "gates_v1_2": { "identity_gate_BC_TV_0": 1.0, "identity_gate_BC_AM_0": 1.0, "metric_saturation_BC_TV_cross_0": 0.259, "metric_saturation_BC_AM_cross_0": 0.246, "identity_gate_passed": true, "metric_saturation_passed": true }, "primary_findings_v1_2": { "verdict": "INCONCLUSIVE", "reason": "No compression level satisfies plateau condition BC(L*) >= 0.7 AND BC(L*+1) < 0.4 on both TV and AM. L4 satisfies the first clause (BC_TV(4)=0.892) but BC_TV(5)=0.458 misses the second clause (0.458 > 0.4) by 0.058." }, "secondary_findings_v1_2": { "F1_memory_no_op": { "description": "Memory summary preservation is behaviorally equivalent to full memory in this observer class - REPLICATED across observer designs.", "evidence": "BC_TV(0)=BC_TV(1)=BC_TV(2)=1.000 under both v1.0/v1.1 and v1.2 observers", "design_invariance": true }, "F2_weights_carry_identity_v1_2": { "description": "In the properly context-dependent linear-bandit observer, the W matrix carries the identity-relevant signal.", "evidence": { "BC_TV_L4": 0.892, "BC_TV_L4_random": 0.282, "gap_TV": 0.610, "BC_AM_L4": 0.885, "BC_AM_L4_random": 0.266, "gap_AM": 0.619, "SR_L4": 0.433, "SR_chance": 0.033, "VP_L4": 0.927, "PS_L4": 0.986 }, "interpretation": "At L4 (W preserved, preferences and memory discarded), behavior is preserved almost perfectly. This REVERSES v1.1 F2 which was a mathematical artifact of softmax shift-invariance.", "confirmed_by_sister_study": "O0-CRP-036 verdict WEIGHTS_DOMINATE under the same observer fix" }, "F3_argmax_lifts_modest_v1_2": { "description": "Per-action argmax(W) + argmax(p) preserves modest identity signal.", "evidence": { "BC_TV_L5": 0.458, "BC_TV_L5_random": 0.299, "gap_TV": 0.159, "BC_AM_L5": 0.505, "BC_AM_L5_random": 0.242, "gap_AM": 0.263 }, "interpretation": "Preserving per-action structure of W (which context feature each action attends to) retains substantial identity signal even after most numeric information is discarded." }, "F4_sub_argmax_collapse_v1_2": { "description": "Compression below the per-action argmax level (L6 one-bit invariant, L7 identifier) collapses to the random baseline.", "evidence": { "BC_TV_L6": 0.326, "BC_TV_L6_random": 0.317, "BC_TV_L7": 0.319, "BC_TV_L7_random": 0.319 } } }, "supported_claims_v1_2": [ "Opens the PERSPECTIVE non-drift branch of CRP-001.", "Memory summary preservation is a behavioral no-op (F1, design-invariant).", "In the properly context-dependent linear-bandit observer, the W matrix carries the identity signal (F2 v1.2, REVERSING v1.1 F2).", "Per-action argmax structure of W preserves modest identity signal beyond a random baseline (F3 v1.2).", "Sub-argmax compression (one bit or less) collapses to random (F4 v1.2).", "v1.0/v1.1 F2 was a mathematical tautology (softmax shift-invariance); documented and corrected here.", "Sister study CRP-036 (which exposed the bug) also gives WEIGHTS_DOMINATE under the fixed observer." ], "not_established_claims_v1_2": [ "Whether the preregistered plateau condition is meaningful at all - registered followup CRP-035-R4 will recalibrate.", "Whether F1, F2 (v1.2), F3 generalize beyond the linear-bandit observer class (CRP-037 registered).", "Whether the W-vs-p asymmetry holds for other environment classes.", "Any claim about identity or perspective in a philosophical sense." ], "adversarial_notes_v1_2": [ "The design correction was detected by a sister study (CRP-036), not by internal review. This is documented publicly.", "The random-parameter baseline at L4 uses W matrices drawn from N(0,1); different distributions would move the baseline.", "The plateau condition is a specific curve-shape rule; the observed data is closer to a declining ramp with L4 plateau." ], "replication_queue": [ {"id": "O0-CRP-036", "kind": "alternate_compression_path", "target": "preference-preserving path (invert L4)", "status": "EXECUTED_UNDER_V1_2", "verdict": "WEIGHTS_DOMINATE"}, {"id": "O0-CRP-037", "kind": "cross_architecture", "target": "PP observer with beliefs+precisions", "status": "REGISTERED"}, {"id": "O0-CRP-035-R3", "kind": "environment_generalization", "target": "nonlinear reward", "status": "REGISTERED"}, {"id": "O0-CRP-035-R4", "kind": "verdict_rule_recalibration", "target": "gap-from-random criterion", "status": "REGISTERED"} ], "figures": [ {"file": "figures/01_bc_curves.png", "caption": "Primary BC_TV and BC_AM curves with random-parameter and cross-observer baselines (v1.2 observer)"}, {"file": "figures/02_all_metrics.png", "caption": "All preregistered metrics (BC_TV, BC_AM, SR, VP, PS) across compression levels (v1.2)"}, {"file": "figures/03_gap_from_random.png", "caption": "Non-triviality gap (compressed BC minus random-parameter BC) across levels (v1.2)"}, {"file": "figures/04_heatmap.png", "caption": "Metric heatmap (v1.2)"}, {"file": "figures/05_v10_vs_v11.png", "caption": "v1.0 diagnostic vs v1.1/v1.2 comparison"} ], "files": { "preregistration": "preregistration.md", "scientific_record": "scientific_record.md", "code": ["src/observer.py", "src/compressors.py", "src/measurements.py", "src/run_study.py", "src/analyze.py", "src/summarize.py"], "results": ["results/summary_exploratory.json", "results/summary_confirmatory.json"], "v1_0_diagnostic": ["data/v1.0_diagnostic/summary_exploratory_v1.0.json"] }, "provenance": { "philosophical_source": "OA-O001 - 'end credits contain only your name' - a perspective reduces to a single identifier.", "conceptual_relationship": "operationalization", "disclaimer": "Conceptual provenance is not empirical support. This study operationalizes a compression pipeline. It does not evaluate metaphysical claims." }, "runtime_seconds": 184 }