READ-ONLY RESEARCH · NO CAPITAL AUTHORITY · NOT_CAPITAL_READY — evidence runs are backtests over synthetic or real historical market bars. No run has executed an order or moved capital. SigmaX proposes; humans authorize.
Proof moves forward. Capital does not.
The Proof Foundry is where SigmaX holds every capital-linked claim to an admissible-proof standard before it can influence a decision. It surfaces the best evidence for a claim and the best evidence against it, names what is still missing, and never turns a strong-but-unproven result into a green light.
AI can analyze and propose. Deterministic checks validate. Humans authorize. SigmaX does not move capital. Everything on this page is research evidence from read-only backtests — some over synthetic data, some over real historical market bars. No run executed an order or moved capital. It demonstrates the standard; it is not an allocation recommendation, and nothing here is capital-ready.
Every claim answers the same four questions.
This is the Evidence Spine made concrete. Two of the four questions are adversarial by design — SigmaX tries to break a claim before it earns the right to influence a decision. The same discipline governs a research edge and a live capital decision.
The strongest supporting evidence a claim has earned — always tagged with its evidence class and its limitation. Support on its own is never trust.
The most damaging admissible finding — overfit, fragility to costs, or regime dependence — surfaced first, not buried. SigmaX looks for reasons a claim should not move.
What is still missing before a claim could ever be trusted — a clean out-of-sample test on real data, then forward evidence the model has never seen. Synthetic evidence can never close a real-data gap, and a historical backtest, however clean, can never close a live-execution gap.
The single test that would move the claim furthest toward — or away from — proof, pre-registered with explicit success and failure criteria before it runs.
Proof is earned, never assumed.
The Foundry is built to keep unproven claims from ever looking finished. The same posture protects every capital decision downstream: the evidence, the counter-evidence, and the limits travel with the claim.
A strong in-sample number is recorded as “supports an early pattern — not out-of-sample proof.” It is never a win and never a recommendation.
Some runs here are backtests over real historical market bars, and some of those record a passed walk-forward out-of-sample test. That is a precondition for human review — not proof, not authorization, and not capital-readiness. A backtest over historical bars is not a live-executed result: no run on this page placed an order, took a fill, paid a real cost, or moved capital.
Forward observation on data the model has never seen, an unbroken record of costs and fills, human review of the evidence and the counter-evidence, and an explicit operator authorization. No run here has any of those, which is why every run is NOT_CAPITAL_READY however strong its numbers look.
Sample and synthetic records are labeled as such and never promoted into live evidence, and a run that declares a stronger dataset than its fingerprint can evidence is shown refuted. Any capital decision remains a separate, human-authorized, operator-governed step.
The runs behind the standard.
Each entry is one research-evidence run with its verdict, evidence maturity, and data mode. Of the 16 runs on record, 6 are read-only backtests over real historical market bars and 10 are over synthetic or sample data — these counts are read from the runs below, not asserted. Every run is read-only: none placed an order or moved capital. None is capital-ready, and none is an allocation recommendation. Open a run for its objective, the evidence for and against, the gaps, and the decision record.
Data mode is DERIVED, not repeated: each run’s declared mode is checked against the dataset fingerprint the run actually hashed. A declaration that overstates its evidence is shown refuted, and a declaration with nothing to check it against is shown as an unverified claim. Every row also carries the age of the recording — these are historical records, not live readings.
The evidence-maturity ladder, L0 to L9.
Every run in the log above carries a rung. This is how far a research claim has climbed inside the Foundry — nothing skips L5, the walk-forward out-of-sample rung, on the way to human review (L8) or a frozen record (L9). A rung is a description of the evidence gathered, never an authorization: even L9 carries no capital authority.
This ladder is not the EL1–EL5 Evidence Levels published on the methodology page. The two measure different things and are not comparable: EL1–EL5 grade how far a strategy has moved toward capital (EL4 is internal-capital live; EL5 requires an approved legal pathway), while L0–L9 grade the strength of a research record that has moved no capital at all. An L6 run is not “past EL5”; every run in this log sits at EL1–EL2 on the capital ladder, because none has been live-executed.
- L0Idea
- L1Plausible mechanism
- L2Data located
- L3Pattern observed
- L4Replicated pattern
- L5Walk-forward OOSOOS rung — never skipped
- L6Anti-overfit / fee / regime resilient
- L7Shadow forward
- L8Human review candidate
- L9Frozen record
The Proof Foundry is a research-evidence surface. Every result shown is produced by a read-only backtest — over synthetic data or over real historical market bars — and is NOT capital-ready. A backtest is not a live-executed result: no run placed an order, took a fill, paid a real cost, or moved capital, so nothing here is a representation of trading results, a performance claim, or an allocation recommendation. A strong in-sample number supports an early pattern only. A passed out-of-sample walk-forward on real bars is a precondition for human review, not proof and not authorization. SigmaX holds no capital authority; any capital decision remains a separate, human-authorized, operator-governed process.