| { |
| "name": "Decision sufficiency of estimated predictive representations", |
| "version": 1, |
| "scope": "Controlled finite-world study, not fresh externally authored model confirmation or operational diagnosis", |
| "hypothesis": "At matched eight-symbol capacity, retaining estimated joint hypothesis/future-observation relationships improves adaptive decision value over posterior-only and observation-only compression. Neither exact inference nor current-decision sufficiency repairs missing conditional relationships.", |
| "worlds": { |
| "development": {"families": ["independent", "context", "parity", "uninformative"], "per_family": 8, "seed": 18401}, |
| "confirmation": {"families": ["independent", "context", "parity", "uninformative", "redundant", "orientation_alias"], "per_family": 32, "seed": 18501}, |
| "held_mechanisms": ["redundant", "orientation_alias"], |
| "hypotheses": 2, |
| "current_binary_measurements": 6, |
| "retrievable_binary_measurements": 2, |
| "joint_states": 512, |
| "availability_conditions": ["full", "three_available"], |
| "unavailable_evidence": "Three hidden current channels are never retrospectively queryable. Only the two recorded future channels can be revealed after paying a cost.", |
| "freshness": "Generate confirmation joint laws and sample streams only after code, tests, development output and protocol hashes are locked. These are newly generated controlled worlds, not independent external authors or physical evidence. No opened bnRep or diagnostic cases count as fresh confirmation." |
| }, |
| "task": {"actions": ["identify_H0", "identify_H1", "abstain"], "correct_utility": 1.0, "incorrect_utility": -2.0, "abstain_utility": 0.0, "query_costs": [0.0, 0.01, 0.05], "query_budgets": [0, 1, 2], "tie_tolerance": 1e-10}, |
| "estimation": { |
| "label_budgets": [64, 256, 1024, 4096], |
| "sample_seeds": [18601, 18602, 18603], |
| "nested_samples": true, |
| "models": ["joint_jeffreys", "naive_bayes", "support_mixture"], |
| "joint_pseudocount_per_cell": 0.5, |
| "naive_beta_per_state": 0.5, |
| "mixture_weight_joint": "N / (N + 512), fixed before confirmation", |
| "primary_model": "support_mixture", |
| "target_truth_access": "Estimator and encoder accept only sampled labeled tuples and declared availability. True joint laws enter only sample generation and separate evaluator after policy construction." |
| }, |
| "representations": { |
| "symbols": 8, |
| "methods": ["identity", "query_joint", "posterior_clusters", "posterior_bins", "future_marginal", "raw_information", "raw_entropy", "hash_code"], |
| "query_joint": "Weighted deterministic eight-means of sqrt estimated P(H,Y0,Y1 | E). 32 iterations, farthest-point initialization weighted by estimated P(E).", |
| "posterior_clusters": "Same weighted eight-means and iterations on sqrt estimated P(H | E).", |
| "posterior_bins": "Eight fixed equal-width posterior bins.", |
| "future_marginal": "Same weighted eight-means on sqrt estimated P(Y0,Y1 | E), removing hypothesis relationships.", |
| "raw_information": "Three current bits with greatest estimated individual mutual information with H, all eight resulting patterns retained.", |
| "raw_entropy": "Three current bits with greatest estimated marginal entropy, all eight patterns retained.", |
| "hash_code": "Fixed probability-blind hash of current evidence coordinates to eight codes.", |
| "identity": "Retain all available raw patterns; strongest same-estimator information reference, not capacity matched.", |
| "code_model": "Push estimated law through encoder and apply the same exact bounded planner. No gold posterior, policy or future outcome enters encoding.", |
| "storage": "Count encoding lookup, resident compressed joint, fitted statistical state, temporary fit memory, planner tables, materialization and all construction work. Compression is not free and cannot add information." |
| }, |
| "decomposition": { |
| "total_regret": "V(full raw truth) - V(realized estimated policy)", |
| "information": "V(full raw truth) - V(available raw truth)", |
| "representation": "V(available raw truth) - V(encoded truth)", |
| "model": "V(encoded truth) - V(realized policy planned exactly under estimated encoded model)", |
| "inference": "Zero for main exact-planning arms; verify against an independently implemented scalar enumerator. A no-abstention action heuristic is a separate negative control.", |
| "note": "All values are ex-ante world expected utility over admissible policies, including query costs. Static and adaptive sufficiency are evaluated separately." |
| }, |
| "secondary_adequacy_diagnostic": { |
| "conditional_dispersion": "Expected total variation between each raw-state conditional P(H,Y|R) and its encoded-cell conditional P(H,Y|Z). Compute estimated and evaluator-only true dispersion separately.", |
| "bound": "Representation loss <= (3 + budget*query_cost) * true conditional dispersion, verified at every evaluated state budget and cost. This follows from the bounded utility range and is not a new theorem.", |
| "limitation": "Estimated dispersion is internal model adequacy only; it is not a certificate for the true law when the model is wrong.", |
| "registered_before_confirmation": true, |
| "primary_thresholds_or_encoders_changed": false |
| }, |
| "primary": { |
| "model": "support_mixture", "label_budget": 1024, "availability": "full", "method": "query_joint", |
| "aggregation": "Equal weight on both positive query budgets, all three costs, three independent sample streams, worlds within families, then all six families. No selected publication, cost, seed or mechanism subsets.", |
| "compact_controls": ["posterior_clusters", "posterior_bins", "future_marginal", "raw_information", "raw_entropy", "hash_code"], |
| "mean_utility_gain_at_least": 0.01, |
| "all_paired_lower_95_bounds_above": 0.0, |
| "bootstrap_draws": 10000, "bootstrap_seed": 18701, |
| "resampling": "Worlds paired within each fixed family; costs, masks, budgets, samples and derived observations are not independent cases. Simultaneous Bonferroni bounds across the six eligible compact controls.", |
| "full_information_noninferiority_margin": 0.01, |
| "note": "A compact-control advantage is not superiority to uncompressed estimation, not evidence for learned control, and not external generalization." |
| }, |
| "cost_promotion": { |
| "quality": "Lower paired 95% query_joint minus strongest full-information statistical reference > -0.01; all eligible full models reported", |
| "cost": "At least 20% measured complete cost reduction including estimation, encoding, model materialization, planning and repeated requests; resident and transient storage reported separately", |
| "repeated_request_counts": [1, 32, 1024, 65536], "timing_repeats": 11, |
| "external_confirmation_required_before_practical_promotion": true, |
| "no_production_latency_or_engineer_time_claim": true |
| }, |
| "controls": ["Faithful full explicit law", "Faithful unavailable-evidence law", "Empirical estimated laws without supplied probabilities", "Known exact posterior-only adaptive counterexample", "Same exact planner for every representation", "Independent scalar exact planner", "No-abstention inference negative control"], |
| "falsification": [ |
| "If query_joint fails its strongest matched compact controls, reject its registered advantage.", |
| "If estimated representations preserve their own modeled decisions but lose true value, attribute loss to model/representation error rather than claiming reasoning transfer.", |
| "If missing raw evidence already reduces oracle value, more computation cannot recover that information.", |
| "If posterior-only compression is static sufficient but loses adaptive value in the exact counterexample, current posterior sufficiency alone is inadequate for this query problem.", |
| "No architecture escalation, new controller training, target threshold adjustment or favorable subgroup promotion after confirmation." |
| ], |
| "execution": {"device": "cuda:2", "dtype": "float64", "new_neural_modules": 0, "backbone_executions": 0, "controller_refits": 0, "prior_results_immutable": true} |
| } |
|
|
