Pretty equity curves lie.
A smooth curve means nothing if random entries with matched exposure produce the same result. Most of the time, they do.
Private beta forming · see build status
For MT5 traders drowning in backtests
Your backtest says it works. EdgeTruth tries to prove it doesn't.
Bring one gold strategy: a spec, JSON-shaped export, or plain English. EdgeTruth locks the assumptions, then runs a five-gate gauntlet: matched-random control, faithful execution, signal agreement, sample floor, and doubled costs. You get a graded verdict and immutable evidence report. Not a promise. A test result.
The problem
Most trading tools help you create more strategies. That was never the problem.
The problem is false-positive overload. Generators can produce thousands of robots. Backtesters make fragile ideas look convincing. Monte Carlo reshuffles your own trades and still never asks the real question:
Did the entry logic create edge, or did it ride the matched exposure a random strategy would have captured?
EdgeTruth exists to answer that question before capital does.
Costs change the verdict.
Spread, commission, slippage, and swaps turn many profitable backtests into noise. EdgeTruth simulates honest costs, then doubles the spread.
Looking good once proves nothing.
One history slice, one cherry-picked window, overlapping trades counted twice. EdgeTruth enforces faithful execution and a sample floor.
Positioning
Not a generator. A kill gate.
EdgeTruth is built on one uncomfortable premise: most strategies should not be deployed. It does not start by promising more robots. It starts by trying to kill the one in front of it. Survivors mean more because the test was hostile.
Strategy generators
Create more candidates.
Generic backtesters
Replay one version of history.
Monte Carlo tools
Reshuffle your own trades.
EdgeTruth
Tests whether your mechanism beats a matched-random control under honest costs, then shows its work.
The spine
Brutal today. Infrastructure tomorrow.
Today: a deliberately narrow, hostile gauntlet for gold breakout and channel strategies, five gates, graded verdicts, immutable evidence. Private beta forming now.
The vision: the validation layer for strategy work. Every gate deeper, your whole book tested for hidden correlation, and the truth lab reachable from web, API, MCP, with your own model key.
Everything is labeled Now, Next, Later, or Vision. If it is not built, we say so. A truth lab that inflates its own roadmap is not one.
No parameter maze
A conversation, then a locked card.
EdgeTruth is not a form with 87 fields. You start like this: “Test a gold NY opening range breakout, no weekend holds, one position at a time.”
The intake agent turns that into a structured spec, asks the missing questions, and produces a confirm card. You approve the card. The engine runs.
InstrumentTimeframeRule familyTest windowCost modelSeed set
The AI can help shape the spec. It cannot edit the locked assumptions or write the final numbers.
How it works
From idea to verdict in one session.
Bring the strategy.
Paste JSON, upload a structured export, or describe the rules in plain English.
Resolve ambiguity.
If rules are ambiguous, EdgeTruth asks instead of guessing. Unsupported inputs are rejected with a reason.
Lock the assumptions.
Confirm instrument, timeframe, test window, cost model, seed set, and data manifest before anything runs.
Run the gauntlet.
Five gates, all computed by the deterministic engine, never by the model.
Read the Truth Report.
Graded verdict, equity versus control, breakdowns, Premortem, and immutable provenance.
Survivors earn the export.
The export gate exists today. Survivor-only MT5 export is the next stage of the roadmap.
The gauntlet
Five ways to die before you risk a dollar.
Every run faces the same gates, in the same order, computed the same way by the engine.
Gate 1: Matched-random control
Same instrument, timeframe, trade count, long/short mix, and exposure. Only entry timing is randomized.
Gate 2: Faithful execution
Non-overlapping trades executed the way an EA actually would. No double-counting artifacts.
Gate 3: Signal agreement
Fast signal source cross-checked against a causal oracle. If they disagree, the run says so.
Gate 4: Sample floor
Fewer than 20 trades means no verdict. Insufficient sample is an answer.
Gate 5: Doubled costs
Spread x2. If the edge evaporates under hostile costs, it was not a robust mechanism.
Verdicts: Grades you have to earn
Today: KILLED, INSUFFICIENT EVIDENCE, and C (Scout). B and A ship with the gates that earn them.
Roadmap chips: multi-seed control bands, start-date robustness, prop-rule drawdown simulation, selection audit, and tick confirmation are tracked on the build status board.
The selection audit
Why grade A is hard.
A strategy that survives one test is not automatically robust. Selection is the hidden tax: how many variants did you try before this one looked good?
The A-grade roadmap includes a multiplicity ledger, family tracking, holdout protection, and stronger calibration. It is hard on purpose.
Audit stance
High grades should feel expensive.
C means scout. B means survived deeper robustness. A means the strategy passed a stricter evidence bar and the path that produced it is visible.
The AI, two-sided
The AI does everything except the one thing it must never do.
Every other AI trading tool lets a language model improvise your results. EdgeTruth is built the opposite way, and it still uses AI hard, just not where it can lie.
Pull quote: the model proposes. The engine disposes.
Evidence guardrail
Language model outside. Deterministic engine inside.
What the AI does
Turns messy rules into a rigorous, testable spec.
Asks ambiguity questions instead of guessing.
Explains the verdict and Premortem in plain language.
What the AI cannot do
It cannot write a single number in your report. Specs are validated, runs fire only against a locked card, and every result comes from the deterministic engine.
Every report is bound to a hashed strategy version and provenance manifest: cost model, test window, seed set, data manifest, gate results, engine version, and report ID.
Gold kill demo
The gold breakout that made +33% and still could not beat random.
On first pass it looked tradable: a gold breakout mechanism, cost-honest simulation, multi-year out-of-sample window, roughly +33% net.
Then it met its matched-random control. Random entries, same instrument, trade count, long/short mix, and exposure, did just as well. On the closest variant, the random control's mean return per trade actually beat the signal.
Verdict cardNo certified edge
Best fork earned: C (Scout). Several forks: killed. The return came from gold-bull exposure, not timing skill.
Join the waitlistMatched-random explainer
Monte Carlo asks whether the path was lucky. Matched-random asks whether there was edge at all.
Monte Carlo reshuffles your own trades to test sequence risk. Useful, but it assumes the trades themselves were skillful.
EdgeTruth builds a control strategy with your exposure and none of your logic: same long exposure on gold, same number of trades, random entries. The question is whether the breakout timing added anything random timing did not.
Two different questions
Generation asks “what can I create?” Validation asks “what should survive?”
EdgeTruth is not trying to be StrategyQuant, an optimizer, or a robot factory. It is the adversarial second opinion after a candidate looks good enough to tempt you.
StrategyQuant-style workflow
Generate, optimize, filter, and tick-test thousands of robots. Powerful, but heavy. Easy to drown in false positives.
EdgeTruth workflow
Bring one candidate. Lock the assumptions. Run the truth check. Kill it or keep it.
Workflow speed
The truth check takes seconds, not hours.
Strategy generators are heavy because they create and optimize thousands of candidates. EdgeTruth takes one strategy you are considering, locks the assumptions, and tries to falsify it against realistic costs and a matched-random control.
On the current internal engine, a full six-fork gauntlet on the gold demo ran in about 9 seconds on a laptop-class machine. The extended report, including breakdowns and correlation checks, finished in under a minute.
Full gauntlet~9 sec
Measured internal run: six forks, strategy signal plus matched-random baseline.
Extended report< 1 min
Breakdowns, robustness views, and correlation checks.
HardwareLaptop-class
Designed to run falsification without a dedicated research box.
Not a like-for-like benchmark. StrategyQuant and EdgeTruth do different jobs. The point is workflow speed: adversarial validation becomes cheap enough to run on every promising candidate.
Vision: portfolio truth
Your 15 robots might be one bet.
The deepest lie in a folder of backtests is not any single curve. It is that together they look like diversification. In one analysis, real returns collapsed to 0.99 correlation: one gold bet wearing fifteen costumes.
The insight is measured. The feature is future. This page keeps them separate.
Where this goes
Vision, not shipped
One strategy at a time answers whether this is edge. The bigger question is whether your strategies are actually different bets.
StrategyQuant refugees
Built for traders with too many robots and not enough truth.
If you use StrategyQuant or similar generators, your problem is not candidate supply. It is candidate trust. EdgeTruth is the second opinion before deployment.
Test the spec before you code the EASeparate timing edge from market betaSee what honest costs and spread x2 doKeep an immutable report for every decisionStructured JSON-shaped exports supported todayUnsupported shapes get flagged, never silently approximated
Later: workflow infrastructure
Wire the truth lab into your own loop.
Later, an EdgeTruth MCP server will let your own agent submit a strategy, run the gauntlet, and read back the verdict and manifest inside your own loop.
Later, bring your own model key. You pay your provider's metered rate. We never mark up inference. The rigor is ours; the AI spend and data path stay yours.
Privacy and trust
Your strategies stay yours.
Serious traders will not upload strategies into a black box. So here is exactly where we stand, split into what is enforced by architecture and what we commit to as policy.
Enforced today: reports are immutable. A database trigger prevents edits after write.
Enforced today: every run is locked to a hashed strategy version and manifest.
Enforced today: the AI layer cannot write or alter results.
Policy: your strategies are private to your workspace.
Policy: we do not trade, resell, or train models on your strategies.
Policy: deletion honored on request; aggregate telemetry only where explicitly allowed.
Build status and beta
See what is shipped before you join.
A truth lab that inflates its own roadmap is not one. The beta page shows the current shipped surface, what is merging, what is next, and the waitlist form.
Open the build status boardFAQ
Is EdgeTruth financial advice?
No. EdgeTruth provides testing tools, evidence, and reports. It does not tell you what to trade, manage capital, or promise profitability.
Does a passing grade mean the strategy will make money?
No. A grade means the strategy passed the specified gates under the recorded assumptions. Future performance is never guaranteed.
What grades exist today?
KILLED, INSUFFICIENT EVIDENCE, and C (Scout). B (Survived) and A (Certified) ship with the gates that earn them.
How is matched-random different from Monte Carlo?
Monte Carlo reshuffles your own trades to test sequence risk. Matched-random builds a random-entry control with your exposure to test whether the entry mechanism added anything at all.
What can I test today?
Gold breakout and channel strategies as JSON spec, structured JSON-shaped export, or plain English. Other mechanisms get a clear not supported yet.
Can I upload a StrategyQuant strategy?
Structured JSON-shaped exports, yes. Native SQX project files are a fidelity problem we refuse to hand-wave. Certified import rolls out family by family.
Can I upload an MT5 EA?
Not yet. Arbitrary MQL is arbitrary code. Certified import comes later, family by family.
Can I upload a trade log?
Not for a Truth Test. A trade log cannot answer the mechanism question. Trade logs will get their own Triage Report.
Does EdgeTruth generate or improve strategies?
No. That is deliberate. An improvement loop bolted onto validation becomes an overfitting machine.
What if my strategy gets killed?
Then the lab worked. You get a Premortem explaining the cause and keep the capital you were about to spend finding that out live.
Does anything ever pass?
Rarely, and that is the point. A gate everything passes is not a gate.
What does the beta cost?
Free during the founding cohort. Join the waitlist and we will invite you when the private beta opens.