Methodology

Every rule the engine enforces, in plain language and in math. Nothing here is adjustable per customer, per tier, or per request — that is the point.

Execution model

Causality: a signal computed at the close of bar t fills at the open of bar t+1. Never same-bar, never mid-price. Repainting is structurally impossible — there is no code path that can see a bar before it closes.
Fills: buys pay the ask, sells hit the bid. Where our tick store covers the run (FX majors, index CFDs, metals, energy — real recorded Dukascopy quotes), exits are tick-exact: the actual tick sequence decides which of TP/SL traded first and at what price. A stop fills at the tick that hit it — equal to the stop when touched, worse when the market gapped, never better; a take-profit fills at its own price, never better. Ticks are re-quoted through the same cost scenario before any comparison, because the raw feed spread is several times tighter than what a retail account actually pays. Elsewhere, exits are resolved by walking the 1-minute bars inside the exit bar; both levels inside a single minute means stop-loss wins (the conservative tie-break — ambiguity always counts against the strategy). A gap through a level fills at the (worse) open, not at the level. Every result states which tier applied in its exit_resolution field — a coarser fallback is disclosed, never silent. Competitors either guess the intra-bar path from OHLC or sell this resolution as a paid add-on; here it is the default.
Costs: every result is a bracket — zero-cost / measured spread / retail floor / 1.5× stress. The retail floor binds the verdict. The zero-cost line exists only to show you the gap. An instrument the cost model cannot price is refused, never priced at zero.
Timeframes: every minute-based bar size from 1m to daily is served — all bars are built from the same 1-minute store, so 20-minute bars are real 20-minute bars, never an approximation.

The Gauntlet — every gate, with its math

Gate	Plain English	Math
Count floor	Too few trades = no verdict, in either direction. Nulls are not run, and an underpowered attempt costs you nothing.	n_trades ≥ 100 for a full verdict; ≥ 10 for any statistics
The mirror image	The exact opposite strategy — same entries, reversed direction — must do worse, or your timing is backwards or irrelevant.	`net_R(real) > net_R(flipped)`
The Noise Twin test	20 versions of your strategy with randomly shifted entry timing — same trade frequency, same exits, same costs. You must stand 3 standard deviations above that noise.	`z = (R − mean(nulls)) / sd(nulls) ≥ 3`, crc32-seeded, reproducible forever
Every year positive	A real edge pays in every year, not one lucky regime. When 2020 is in the window, the ex-2020 result is computed too — the COVID year flatters almost anything.	net > 0 ∀ years; ex-2020 reported
Recent-third positive	Edges decay. The most recent 30% of days must be profitable.	`Σ daily_R (final 30%) > 0`
The Luck Bar	The more trials you run, the better your best result looks by pure chance. Your Sharpe must beat the bar at YOUR cumulative look count — every run and every swept variant joins your ledger and none of it resets. Since engine v0.5.0 the count is effective: variants of one strategy (same market + entry mechanism + direction) pool as a family and enter as the square root of the family's size — a 44-cell TP/SL sweep on one idea adds ~6.6 effective looks, not 44 — while genuinely distinct ideas count in full. The discount is pre-registered and decoy-validated (pure-noise costume families stay below the discounted bar more reliably than independent noise stayed below the old full-count bar); receipts show raw and effective side by side.	`Sharpe > max(√(2·ln N_eff), 1) · √(1/T)`, N_eff = Σ √(family size) over your strategy families (mega sweeps: √(configs × cells)), T = years of data. The `max(·, 1)` is the cold-start floor: at N=1 the raw term is 0, but one draw of noise still has a typical magnitude of ~1 standard deviation, so the very first look is gated at z=1 (e.g. 1/√14 ≈ 0.267 on 14 years) instead of zero. From N=2 the √(2·ln N) term already exceeds it.
Beats buy & hold	Holding the instrument is free. Your timing must add at least +0.10 Sharpe over the better of holding long or short.	`Sharpe − \|Sharpe(hold)\| ≥ 0.10`
Survives a one-bar delay	One extra bar of delay must not kill the edge — if it does, retail execution will never capture it live. This is a HARD gate on fast bars (15-minute and under); on slower bars it is reported but a slow strategy is not failed for it.	delayed net > 0 and ≥ 50% of undelayed

Verdicts

PASS	cleared every gate. Rare by design.
FAINT	beat its Noise Twin and the 2-sigma floor, but short of a full pass — promising, with a named gap.
DEAD	killed at a named gate, with the reason.
NOT JUDGED	too few trades (VETO_UNDERPOWERED, COUNT_FAIL), missing data (NO_DATA), couldn’t compile (UNCOMPILABLE), or a rule that never fires (VOID_DEGENERATE_SPEC) — “we couldn’t honestly test this” is never conflated with “it failed,” and it never costs credits.

Path risk (every run)

The max drawdown a backtest shows is one draw from a distribution: the same trades in a different order produce different worst paths. Every result includes a stationary block bootstrap (1,000 resamples, mean block 10 days — preserving short-range autocorrelation, seeded and reproducible) of the daily P&L, reported as drawdown percentiles with your observed path located inside the distribution. Size to the 95th percentile, not to the path you happened to get. Two honest notes: this is sizing information and gates nothing (reshuffling a series cannot change its mean), and regime clustering longer than the block length is partially broken by any bootstrap — so the tail percentiles are a floor, not a ceiling.

The Skeptic audit (premium)

A mechanized adversarial battery run against a positive result: day/year concentration, drop-the-best-year jackknife, the TP/SL plateau grid, sibling-instrument placebos, a cost ladder to 2× the floor, and delay fragility. The verdict is REFUTED or NON-REFUTED with mandatory downgrades — restrictions on how the result may honestly be described. The audit re-scores your strategy about a dozen ways, so about a dozen lines land on your ledger: the skeptic raises your Luck Bar like any other trial.

Custom strategies (generated code)

Ideas beyond the grammar are written as code by the engine, shown to you in full, and accepted only if they pass a safety guard (numpy/math only, no I/O) and the mechanical look-ahead test: the data is truncated and the signal re-run — if any entry changes, the code used the future, and the run is refused before a credit is spent.

Data honesty

169 instruments served: FX majors (2010+), crosses, exotics, index CFDs, metals, energy, rates CFDs, and crypto (fee-based costs, synthetic spreads disclosed) — each with its point unit, its conservative retail floor, and its known limitations stated on the card (roll splices, unmodeled CFD financing, indicative volume).
2024–2026 are sealed: the data exists on disk and no ordinary backtest can reach it. The one exception is the sealed look — priced, budgeted at two per lifetime, and spent forever per window — which is exactly what makes its result evidence. Your receipt says so on every run.
Anything we can’t test honestly is listed as unserved with the reason — never approximated. If data is missing you get NO_DATA, never a made-up number.
We are not a data vendor — no bulk price export, no feed. But every verdict is auditable against the data that produced it: the free query_bars tool returns the recorded bid/ask quotes around any timestamp (a small window, enough to check any fill in your trade log against the market that was actually there). Ask the chat for the quotes around any of your trades.
Fill-resolution downgrades are disclosed, never silent: every result carries exit_resolution (tick / 1m / bar), and any run below tick carries a tick_note saying exactly why — no tick store for the instrument, named missing years, or the operator kill switch. Resolution tiers are never mixed inside one run.

The commitments

Realism is not an upsell. You cannot pay us to turn costs off.
Every non-scored outcome states “no credits were spent” — and it’s true. Engine failures on our side are refunded, automatically.
Nothing runs until you’ve seen the strategy card and confirmed — enforced by the server, not by politeness.
Your trial ledger is yours to inspect, line by line, any time.
No subscription: you buy credits once, they never expire, and there is nothing to cancel. Credit purchases are final; failed runs refund themselves automatically.
Every verdict is reproducible from its receipt: spec hash, engine version, data version, trial count.

Open the app →