TRUTHFUL-V2 / LAST UPDATED AUG 22, 2026
How we keep the math
from lying to us.
The old experiment made too many numbers look more certain than they were. We preserved it, flagged it as legacy, and started a new epoch. These are the rules now. If a requirement is unavailable, the answer is "Insufficient evidence." It is never quietly turned into zero.
THE WHOLE PIPELINE
Register, search, freeze, prove.
INPUTS AND EXECUTIONA model cannot trade information it did not own.
FINALIZED BARS
Signals read immutable finalized bars only. Forming candles cannot reach inference.
REQUIREDSIGNAL TIME
Features and signals are calculated after finalized bar close t.
CLOSE tFILL TIME
The order executes at the next bar open, t+1. Reversals are an exit plus a new entry and pay both sides of costs.
OPEN t+1PROVENANCE
Every score owns an epoch, methodology version, artifact hash, data snapshot hash, instrument and cost configuration.
HASHEDFINANCING
Missing BTC funding or SPY short borrow makes the short unavailable. Missing Treasury data makes excess-return metrics unavailable.
NO FALLBACKCURRENT INSTRUMENTSWhat the model is actually trading
- SPYThe SPY equity, with splits, dividends, fees and observed borrow costs for shorts.
- BTCUSD_PERPThe exchange-discovered Kraken BTC/USD perpetual, including contract metadata and funding.
- BTCUSD_LEGACYThe old spot identity. Preserved for audit, excluded from the new long and short experiment.
- TREASURYThe official 13-week coupon-equivalent rate, used only after it was known.
ONE PERFORMANCE ENGINE
One curve.
Everywhere.
Backtests, forward ledgers, charts, APIs and portfolio views use the same performance engine. Equity starts at exactly 1.0. Returns compound. Drawdown includes the initial point. Gross P&L, fees, slippage, dividends, borrow, funding, Treasury cash return and net P&L stay separate.
1.0 × 1.5 × 0.6 = 0.9. Percentage returns are never added.
The sleeve is ruined, closed and invalidated. Loss cannot run below the capital.
METRICS WITHOUT SENTINELS
Undefined means
null.
No NaN. No infinity. No fake 999 profit factor. No zero standing in for "we do not know yet." Every non-finite result fails validation before persistence.
Matched to time.
Sharpe uses matched-period Treasury excess returns, sample standard deviation and the correct market calendar. Sortino measures downside deviation over every observation.
Samples come first.
Profit factor is null with no losses. Kelly is null below 30 closed trades, stays informational, and never affects ranking or promotion.
Abstention is visible.
Directional accuracy includes its numerator, denominator, total bars, active bars and abstention rate. Annualized return is null below one year.
HISTORICAL QUALIFICATION
Useful research.
Never a promotion by itself.
| INTERVAL | MINIMUM HISTORY | EVALUATED PERIODS | CLOSED TRADES | ACTIVE EXPOSURE |
|---|---|---|---|---|
| Intraday | 20 completed sessions | 500 | 20 | At least 10% |
| Daily | 126 evaluated periods | 126 | 20 | At least 10% |
| Weekly | 52 evaluated periods | 52 | 10 | At least 10% |
The search reads development data only. Anchored walk-forward folds train on earlier observations and execute on the untouched following fold without resetting an open test-period position.
FORWARD MATURITYAll requirements must mature before a promotion test.
INTRADAY
20 SPY sessions or 30 BTC calendar days, 500 paired periods, 20 closed trades and at least 10% active exposure.
ALL FOURDAILY
180 calendar days, 120 paired periods, 20 closed trades and at least 10% active exposure.
ALL FOURWEEKLY
52 weeks, 52 paired periods, 10 closed trades and at least 10% active exposure.
ALL FOURNO INCUMBENT
The nominee is paired against a Treasury-funded flat sleeve.
TREASURYPROMOTION PROOFA result has to beat chance and risk limits
- 10,000Deterministic moving-block bootstrap samples on paired net log returns.
- BlockLength is max(2, ceil(n^(1/3))).
- AlphaThe kth test spends 0.05 / (k(k+1)).
- LCB > 0The one-sided lower confidence bound of nominee minus incumbent mean return must exceed zero.
- DD > -5%The nominee forward sleeve must remain above the drawdown ceiling.
HOW A MODEL LOSES ITS PLACE
No vibes.
No second chances.
A champion is immediately dethroned at a 5% forward sleeve drawdown, invalid input, artifact mismatch, or unavailable financing. Otherwise, replacement requires a mature nominee to pass the paired superiority test. A rejected frozen artifact cannot be retested.
SEE THE CURRENT EVIDENCE STATUS ↗Nothing ranks until the evidence earns it.
Visible for inspection, clearly labeled immature.
LEGACY RECORDSPreserved, flagged, excluded
- Raw dataOriginal rows and hashes are retained for audit.
- Rebuilt historyRecoverable calculations are labeled legacy_simulated.
- Old pagesLegacy model pages carry a visible red warning.
- No rankingLegacy evidence cannot seed promotion, trust, portfolio results or truthful rankings.
HYPOTHETICAL OPTIONSScenario only
- Visible inputsIV, spread, commission, strike and allocation assumptions stay on screen.
- Short signalA short signal means buying a put in the scenario, not writing one.
- Zero influenceOptions outputs never enter evidence, trust, ranking, portfolio return or promotion.
DO NOT TAKE OUR WORD FOR IT
Open the ledger.
Check the models.
Create a free account to inspect the frozen artifacts, market charts, signals, costs, sample counts and maturity warnings yourself.
CREATE MY FREE ACCOUNT ↗