Historical prices, engineered values, validation results, and simulation ranges—organized as traceable workbooks and audit files. This room is free to read, and so are the source, protocol, and gate files behind the method. The frozen forecast releases and Step 7 simulation workbooks need a Forecasts membership.
Every later workbook is downstream of this source table. Keep its date, price policy, and warnings attached to any analysis.
Daily raw OHLCV for ASTS aligned to SPY, QQQ, and IWM, with source metadata, data-quality checks, modeling eligibility flags, and separate Phase B repair artifacts.
Move from the target definition to 23 frozen features, 740 unseen outputs, the evidence gates, and research-only forecast ranges.
The prediction clock, five-session label, field definitions, and target-quality checks.
The trailing-only price, probability, volatility, volume, and market-context features.
The workbook preserves the five Phase A candidates and adds distinct Phase B sheets for three repair paths and four audited outputs on the same 740 development dates.
The workbook preserves the Phase A gate record and adds distinct Phase B sheets. Raw Phase B files report each candidate's own nine-gate result without combining passes.
The benchmark-only controller, separate repair diagnostics, and gate-by-gate recovery plan keep the fold-local prior active.
Five-engine audit, percentile ranges, run stability, interval coverage, and tail probabilities.
These artifacts reuse the same 740 dates. Gate passes stay with the candidate that earned them and are never aggregated.
The selected family passes G4 and G8 and reaches 9 of 12 positive folds, but qcut ECE is 6.028% and median within-fold standard deviation is only 0.000569. Its downloadable family summary also contains the unselected empirical-bin member, the closest all-three diagnostic.
It posts +0.000871 Brier improvement, 0.502218 AUC, and 8 of 12 positive folds, but qcut ECE is 5.5738%, missing the gate by 0.5738 percentage points.
The state model passes qcut ECE at 4.361% and has 0.021151 median within-fold standard deviation, but loses to the prior overall and improves only 6 of 12 folds.
Only 3 of 12 folds find an eligible active configuration; the other 9 emit the fold-local prior. All-fold median within-fold standard deviation is 0.000000.
The corrected frozen-reference audit passes G4 only. Its superseded evidence file is intentionally not published here.
Every later workbook is downstream of this source table. Keep its date, price policy, and warnings attached to any analysis.
Daily raw SIP bars for HOOD aligned to SPY, QQQ, XLF, COIN, IBKR, and SCHW, plus lagged BTC and a separately dated SEC event ledger. The site now separates the original Phase A files from a development-only Phase B challenger.
Move from the target definition to 60 frozen features, 685 unseen outputs, the evidence gates, and research-only forecast ranges.
The five-session direction label, decision-time rule, benchmark, and predeclared evidence gates.
Trailing price, volatility, regime, peer, lagged-crypto, probability, and post-event values.
The original five-candidate workbook is preserved unchanged; Step 4 separately documents the three-member Phase B challenger family and selected Signed Core Ridge on the same development dates.
The workbook preserves the original Phase A audit; Step 5 applies those same gates to the verified Phase B challenger.
The preserved Phase A record and the site’s Phase B addendum keep the 55.75% fold-local prior active pending a fresh paper-forward test.
Conditional analog backtests, five-seed stability, percentile paths, and the Q2 event boundary.
Every later workbook is downstream of this source table. Keep its date, price policy, and warnings attached to any analysis.
Source-reported, split-normalized NKE OHLCV aligned to nine market, sector, peer, China, rates, and dollar instruments, with a hashed source ledger and point-in-time SEC filing events.
Move from the target definition to 38 frozen features, 1,237 unseen outputs, the evidence gates, and research-only forecast ranges.
The split-normalized five-session label, actual 40-event ex-date censor, benchmark, and nine frozen evidence gates.
Ten synchronized market histories, 68 SEC filing events, the 40-event actual ex-dividend ledger, data-quality checks, and source hashes.
Trailing price, market, peer, macro-proxy, and filed-fundamental values with the original decision-time boundary preserved.
Three compact ridge families, ten purged outer folds, nested calibration, and 252 terminal labels spanning 272 sessions.
Holdout uncertainty, calibration, flatness, fold consistency, five-run stability, null testing, and regime evidence.
The canonical no-go result, failed-gate diagnoses, and a fresh paper-forward evidence plan without reusing the holdout.
Five-, ten-, and 21-session percentile bands, CRPS backtests, daily paths, and explicit effective-support disclosure.
These files expose the exact row-level predictions, gate ledger, canonical result, executed notebook, and simulation payload behind the published no-go recommendation.
The single machine-readable record for target, data, methodology, holdout metrics, gates, runs, sources, and simulation.
An executed notebook that reloads the exported evidence, checks calculations, and reproduces the central validation figures.
The complete feature-ready NKE table with labels, eligibility flags, and the 38 selected point-in-time values.
All 252 terminal-holdout predictions, expanding-history baselines, labels, and row-level losses from the audit-corrected rerun.
The nine frozen rules, observed values, and pass/fail outcomes used to make the no-go recommendation.
Terminal percentiles, daily bands, backtests, and five model-run probabilities for the interactive simulation.
Retrieval time, exact source URLs, row coverage, source roles, snapshot hashes, and point-in-time boundary notes.
The filed revenue, margin, inventory, and filing-age inputs used by Model 3, preserving the original SEC filing dates used for next-session eligibility.
The 40 actual ex-dividend dates used to censor only five-session target paths that cross a recorded event.
The corrected target, filing-availability rule, development-defined regimes, exact shift null, seeds, and nine gates fixed before the rerun.
UBER uses 1,816 source-reported Nasdaq sessions and a 37-feature challenger library, with separate selection and calibration periods, two strict 21-session embargoes, and a 252-origin historical terminal comparison. The invalid first run is discarded. Four of nine gates pass; every range remains unapproved.
Source-reported UBER OHLCV aligned by completed session with SPY, QQQ, XLY, IYT, and LYFT.
Candidate selection, calibration, terminal metrics, gates, scenarios, risk inputs, and file hashes in one release.
All nine locked rules, comparators, observed values, and pass/open outcomes without threshold waivers.
The frozen forecast registry and outcome ledger are ready to accumulate genuinely new, non-overlapping evidence.
The portfolio build audits 33 supplied workbooks across v5–v59. It preserves materialized history, same-date dependence, time-ordered validation, and fail-closed gates. Manual scores, cached probabilities, mutable snapshots, and blank outcome ledgers are not imported as evidence.
V2 keeps the selected candidate's evidence frozen but displays the expanding benchmark because that benchmark scored better on the already-inspected terminal comparison.
Nine predeclared rules covering data quality, lookahead, interval score, median error, coverage, calibration, consistency, reproducibility, and genuinely new forward evidence.
The v2 policy preserves every v1 gate and switches displayed ranges to the benchmark whenever the selected candidate fails comparative skill and consistency gates.
File hashes, exact workbook roles, confirmed defects, safe concepts, and the evidence boundary for every supplied v5–v34 position-model workbook.
Six overlapping legacy rows are joined to later frozen closes. The check is reproducible, but none of the rows qualifies as independent deployment evidence.
Per-stock momentum, Wilder RSI, prior-only relative volume, market regime, breadth, and dispersion were tested under a frozen development protocol and did not beat the incumbent.
The same completed-session momentum, Wilder RSI, and relative-volume definitions now sit beside ASTS, HOOD, and NKE. They remain outside the frozen deployment models until independent evidence earns a promotion.
A fixed 252-label comparison of the shared momentum, RSI, and volume challenger against a compact price-volume benchmark.
Exact feature definitions, holdout dates, baseline and challenger metrics, promotion rules, and limitations in one JSON record.
A cached GioModel endpoint supplies separately timestamped prices and completed-session indicators, with per-stock fallbacks when a source is unavailable.
These are the descriptive layers built on top of the frozen model releases. None is gated, none forecasts anything, and several exist mainly to record what could not honestly be computed. Each links to the artifact a reader would need to recompute it.
NOPAT and invested capital are both published, so every ROIC can be recomputed by dividing one by the other.
Refused: ROIC for HOOD and FUTU (broker balance sheets are customer assets) and AMT (a REIT). No moat score at all.
ProtocolSeparates failures a modelling change could move from ones only elapsed time can fix. No model can reach a full pass today, because forward evidence has to accumulate after the freeze.
Refused: Any ranking by desirability. It measures evidence completeness, which is not a view on any company.
Facts as originally filed, keyed by the date each became public. One XBRL tag per slot, never stitched.
How CRWD, SOFI, RKLB, ONON and CELH entered the universe as data-only names. All five were run through the v4 protocol on 2026-08-16, so the tier is now empty and all twenty tickers are modelled.
Refused: The owner-supplied ratings that came with these five, which are recorded as provenance rather than published as assessments.
Revenue, gross profit, net income, EPS and share counts as originally filed. Restatements are excluded on purpose.
ATR, MACD, rolling beta, realised volatility, drawdown and relative strength, all computed causally.
Refused: Coverage beyond seven tickers, for now: the other eight price panels are not yet wired in.
The aligned TLT, UUP, HYG, LQD, USO and SPY closes the regime model reads. Proxies, not the underlying rates.
Split-adjusted, with the filing each window rests on recorded beside it.
Equity, rates, dollar, credit and oil each vote, and every vote is published so a label is never opaque.
Refused: GDP, which is quarterly, revised and lagged, with no honest daily proxy.
Minimum variance beat equal weight in 16 of 16 out-of-sample periods, but every scheme over-predicts the level of risk.
Refused: Expected returns, an efficient frontier, and any Sharpe ratio built on one.
Protocol87% of the money in one name carries 96.7% of the risk. Published with the owner's permission.
Refused: Share counts, entry prices and portfolio value, none of which the statistics need.
The 588 sessions all fifteen names share. The window is set by the youngest listing, not the oldest.
Beats the incumbent estimator at every horizon and still misses the bar that would justify using it.
ProtocolMention counts only. The corpus accumulates forward because no historical Reddit archive is open.
Refused: Sentiment polarity, which cannot be validated without a labelled sample.
ProtocolNot published. Zero of fifteen tickers have every component; four have every component except valuation.
Refused: The score itself. A 0–100 rank over four names gives each step 25 points.
These files are frozen research artifacts with target-specific price and event policies. The separately timestamped live technical layer is market context, not investment advice or evidence that any model has a deployable edge.