{ "title": "Corrected AI stock-picker design and dated amendments", "initial_lock_commit": "0d65ef4", "chronology": "Initial lock 2026-09-30; amendments 1-3 precede real model results. Amendment 4 follows the computed verdict: reporting aggregates and the omega wording threshold are post hoc, not pre-registered.", "prior": "Earlier lab results were seen. No numerical probability prior is specified for the corrected run. The doubling check is confirmatory, not blind, and is outside the verdict.", "effective_parameters": { "PANEL_START": "2004-01-02", "END": "2026-06-22", "FIRST_TEST_YEAR": 2008, "LAST_TEST_YEAR": 2026, "MIN_DVOL": 5000000.0, "PRICE_FLOOR": 5.0, "TYPES": [ "CS" ], "TOP_N": 1000, "SPIKE_FACTOR": 1.5, "SPIKE_FACTOR_SPY": 1.1, "SPLIT_TOL": 0.03, "SPLIT_MIN_LOG": 0.22314355131420976, "SCREEN_GUARD": 0.001, "SEGMENT_GAP": 20, "JOIN_SAME_TICKER_MAX_GAP": 126, "JOIN_CIK_WINDOW": 30, "JOIN_RATIO_BAND": [ 0.5, 2.0 ], "FEATURES": [ "ret_1", "ret_5", "ret_21", "ret_63", "ret_126", "vol_21", "vol_63", "dma_21", "dma_200", "volz" ], "ELIG_BARS": 60, "HORIZONS": [ 1, 21, 126 ], "TRAIN_CAP": 400000, "TRAIN_MIN": 5000, "LSTM_HIDDEN": 32, "LSTM_LR": 0.001, "LSTM_BATCH": 1024, "LSTM_EPOCHS": 1, "GBT_PARAMS": { "n_estimators": 300, "max_depth": 5, "learning_rate": 0.05, "subsample": 0.8, "colsample_bytree": 0.8, "eval_metric": "logloss", "tree_method": "hist" }, "REAL_MODELS": [ "random", "logistic", "gbt", "lstm" ], "TRAINED": [ "logistic", "gbt", "lstm" ], "UNIVERSES": [ "pit", "surv" ], "SELECTIONS": [ "argmax_conf", "fixed_1d", "fixed_21d", "fixed_126d" ], "CONCENTRATIONS": [ [ "top1", 1 ], [ "top10_ew", 10 ] ], "COSTS_BPS": [ 0.0, 5.0, 10.0, 20.0 ], "DOSES": [ 0.0, 0.01, 0.03, 0.1 ], "D_STAR": 0.03, "STRESS_DELIST": -0.3, "BASES": [ 20260930, 20260931, 20260932 ], "NULL_BASE": 20260930, "N_NULL": 100 }, "model_label": "LSTM trained for one epoch; the earlier implementation used three epochs", "label": "Above own-universe median of traded return; close t+1 to t+1+h; missing only beyond the final date.", "execution": "Decide at close t, next-close fill; h trading days; gaps carry to resumption; missing-fill and empty sleeves are zero-interest cash; terminal liquidation charged.", "cost": "Half round-trip bps times one-way drift-adjusted turnover; stress changes true ends only; old-cost sensitivity charges a full-book round trip on exit.", "metrics": "Nominal price returns on 2008-01-03 to 2026-06-22; CAGR in calendar years; sample standard deviation; annual Sharpe sqrt(252), risk-free rate zero; drawdown from initial equity before charges.", "amendments": [ { "id": 1, "time": "2026-09-30T02:41:13Z", "class": "before model results", "effect": "Approves design; one-epoch disclosure and code review required." }, { "id": 2, "time": "2026-09-30T03:03:44Z", "class": "before model results", "effect": "Clarifies removed-bar screen, gap boundary, cash sleeves, liquidation costs, label median and date-mean calibration." }, { "id": 3, "time": "2026-09-30T03:43:30Z", "class": "before model results", "effect": "Ticker-wide split reconstruction; security joins; raw-volume feature; float64 scores; complete window and null checks. Same-ticker resumes within 126 days or same-CIK starts within 30 days join if split-corrected ratio is between one half and two. Current identifiers remain limited." }, { "id": 4, "time": "2026-09-30T04:47:36Z", "class": "POST HOC, NOT PRE-REGISTERED; results already seen", "effect": "Adds seed means and omega sums/ratios, plus the one-half systematic-wording cutoff; no threshold in the registered verdict changes. Historical before/after assertion execution was not recorded." } ], "source_sha256": "735bb1a58478c06e08e3a229c84fa4aa45cce87238058286b1472c9701d181a6", "read_amendments": "Amendment 3 replaces segment-level raw-price reconstruction and identity construction. Original paragraphs alone are not the effective design.", "decision_rules": [ { "branch": "REFUTED", "locked_condition": "Model-containing share at least one half, or model main effect at least the setup main-effect sum, in at least two of three seeds in either variant." }, { "branch": "BROKEN rank trigger", "locked_condition": "Model total effect ranks first, second or third among five factors in at least two of three seeds, in either variant." }, { "branch": "BROKEN size trigger", "locked_condition": "Model-containing terms at least half the setup-only terms: M >= 1/3, in at least two of three seeds." }, { "branch": "Original main-effects bar", "locked_condition": "Setup main-effect sum at least twice model main effect in every seed, both variants." }, { "branch": "INCONCLUSIVE", "locked_condition": "Neither earlier branch fires and the main-effects bar fails." }, { "branch": "HOLDS", "locked_condition": "Neither earlier branch fires and the main-effects bar passes; wording also depends on control/scope." } ], "control_rule": "PASS only if IC 0.03 fires in both variants and IC 0 fires in neither; INDETERMINATE if IC 0 fires; otherwise FAIL.", "scope_rule": "Within either stock list and either variant, rank among first two of four or M at least one third in two of three seeds narrows a HOLDS statement; it does not supersede BROKEN.", "null_arm": "100 draws; NULL-SRC and four registered twins plus screened-hindsight sensitivity. Report median, 5th/95th percentiles and beat shares against matched SPY.", "confirmatory_not_blind": "Median NULL-SRC CAGR at least twice price-only SPY CAGR, and horizon shares; outside every verdict condition.", "descriptive": "CAGR attribution, stress and old-cost attribution, cost/turnover arithmetic, best cells, PIT/high-cost/basket corner, IC, omega and data/gate counts. Post-hoc reporting conventions are labelled separately." }