Rent the judge, not the strategy.

Headless extraction cross-checked by three independent methods, confidence-scored, refund-backed.

Headless extraction cross-checked by three independent methods, confidence-scored, refund-backed

11-gate adjudication of a trading signal series.

Use when you have OHLC bars and a -1/0/1 signal series and need to know whether a backtest result is real or an artefact. Applies 11 gates most backtests fail: matched random-entry drift control, date-clustered standard errors, entry at the next bar's open, events-not-fills, absolute (not just excess) return, slippage stress, walk-forward, overlap correction, financing, and a block bootstrap that must agree on both excess and absolute return. Returns CANDIDATE or REFUTED plus every blocking reason.

Is your OHLC real, or an indicative feed pretending to be.

FREE, no payment or account needed. Use BEFORE trusting any backtest, to check whether the price data itself is real. Free and indicative feeds frequently synthesise the bar's open instead of observing it, which pins |open-close|/range near zero; measured on the same FX pairs, a free feed reads 0.020 against a real broker's 0.464. Any result that depends on the open — gap trades, overnight holds, next-bar entries — is invalid on such data. Also reports frozen (zero-range) bars, duplicate and out-of-order timestamps, and bars whose high/low do not bracket open/close. Returns BROKER-GRADE, SUSPECT or INDICATIVE.

You tested N strategies and kept the best; is it noise.

Use when you searched many strategies or parameter sets and kept the best one. Quoting that winner's solo p-value is the most common way a backtest lies. Give the number of things you tested and the winner's t-statistic; returns the family-wise p-value, the t you actually needed, and the t that pure noise would be expected to hand you from a search that size.

Your t-stat assumes independence; your trades cluster.

Use when your trades are not independent — several in the same day, the same regime, or a correlated basket. A t-stat computed as if every trade were an independent draw is inflated; in this repo the factor was about 1.9x. Give per-trade returns and their dates; returns the naive t, the date-clustered t, the inflation factor, and how many genuinely independent days you actually have.

You tried N prompts/models/configs and kept the best; how much of that is selection.

FREE, no payment or account. Use when you picked the best of several prompts, models, configs or hyperparameters from an eval. Keeping the top scorer out of N is SELECTION, and the winner's score is inflated by having looked N times — every eval harness and prompt optimiser does this and reports the winning score as though one experiment were run. Give the score each variant achieved; returns how much of the winner's margin is explained by the search itself, the margin pure noise would have handed you, and whether the excess is real. Supply n_per_variant to also check whether your variants can be ranked apart at all.

Is this Sharpe real, or the best of N tries? Deflated for the search you ran.

Is this Sharpe real, or the best of N tries? Deflated for the search you ran

Did B really beat A, corrected for how many variants you compared.

Did B really beat A, corrected for how many variants you compared

How many observations before this test can answer anything — ask before you spend.

How many observations before this test can answer anything — ask before you spend

Is this drawdown a broken strategy or a normal bad patch, vs your own shuffled returns.

Is this drawdown a broken strategy or a normal bad patch, vs your own shuffled returns

Real relationship, or two autocorrelated series drifting together.

Real relationship, or two autocorrelated series drifting together

Did the process change, or is recent weakness inside normal variation.

Did the process change, or is recent weakness inside normal variation

Every gate we have, composed in the right order, over your whole book — including the cross-instrument check that catches one edge wearing many names.

Every gate we have, composed in the right order, over your whole book — including the cross-instrument check that catches one edge wearing many names

Your FX backtest is probably running on invented opens

The same pairs, the same hours, two feeds. On one, the bar's open sits where a real trade happened. On the other it is manufactured from the close — and the ratio that exposes it is one line of code.

If you kept the best of 200 prompts, that score is inflated

Prompt optimisation has a selection problem nobody corrects for, and it is the same error that kills trading backtests. Here is how much of your winner's margin the search itself explains.

We promised a 10% false-discovery rate and measured 39%

Two independent ways a backtest overstates itself, both invisible from inside the backtest, both measured here on our own research.