REAL DATA · READ-ONLY BACKTEST · NOT_CAPITAL_READY — research evidence only. This run executed no order and moved no capital. SigmaX proposes; humans authorize. No capital authority, no allocation recommendation.
Research only
Kept for research. Not trusted, not review-ready, not capital-ready.
Provenance on this page is DERIVED, not taken from what the record says about itself: the declared data mode is checked against the dataset fingerprint the run hashed, and a self-declared trusted / review_ready is shown as a claim because nothing on this public surface verified it. data_mode "REAL" is corroborated by a REAL data_fingerprint => shown as declared
Rationale · Walk-forward passed on REAL data for "ETH/USD MA-crossover walk-forward OOS vs negative controls (REAL)" (OOS Sharpe 1.265142589262479), but the candidate did NOT beat its negative controls — the apparent edge is not distinguishable from a baseline. Routed to research_only; not review-ready, not capital-ready.
What this run is trying to prove
Establish, on a fair and pre-registered basis, whether "A fee-aware MA-crossover strategy, searched on train and selected on validation, retains a positive walk-forward OOS edge on ETH/USD 1d data that beats its negative controls and survives fee stress." survives real-market out-of-sample testing net of fees and across regimes — and is therefore worth advancing toward human review.
Why it matters · Capital is never authorized by the Foundry (capital_authority="none"). The objective exists to separate a genuine, durable strategy edge from an in-sample artifact BEFORE any operator review, so that only trusted evidence — never an unsupported performance claim — reaches the decision record.
- real-market (non-synthetic) data over a strictly held-out OOS window
- walk-forward OOS edge that is positive after fees across a majority of folds
- evidence the edge is not concentrated in a single regime (regime replication)
- fee/slippage resilience: net-of-cost edge survives realistic execution assumptions
- forward shadow persistence on inert/paper data, unseen at fit/validation/OOS
- synthetic-only evidence (synthetic data can never be capital_ready or review-ready)
- in-sample / validation-only support with no out-of-sample test
- failed real-data out-of-sample (OOS Sharpe <= 0 after fees) => rejected or research_only
- missing fee or regime evidence (incomplete => proof_gap, not proven)
- single-fold dependence or look-ahead/leakage in the evaluation
- a negative control that reproduces the 'edge' (the result is an artifact)
- forward decay: an OOS edge that does not persist in shadow testing
Best evidence for
The strongest fair case in favour of the claim. None of it is proof on its own — each item carries its limitations, and any in-sample / validation number is shown as supports early pattern — not OOS proof.
Proposed mechanism for "A fee-aware MA-crossover strategy, searched on train and selected on validation, retains a positive walk-forward OOS edge on ETH/USD 1d data that beats its negative controls and survives fee stress." is structurally plausible (non-degenerate parameterization).
claim_type=strategy_edge · strategy_degenerate=false · genome_fast=13 · genome_slow=62 · genome_thresh=0.002941097127524945
- plausible mechanism only — not yet causally verified
- parameterization could still be an overfit artifact
Strong in-sample fit (Sharpe 1.48). Preserved as supporting evidence — in-sample only, not trusted proof.
supports early pattern — not OOS proof
in_sample_sharpe=1.4791
- in-sample only — overfit-prone, never trusted proof
- no held-out generalization established by this receipt
Survives validation (Sharpe 1.21). Preserved as supporting evidence — validation only, not trusted proof.
supports early pattern — not OOS proof
validation_sharpe=1.2086 · degradation_in_to_validation=0.2705
- validation only — tuned-window result, never trusted proof
- validation window may share regime structure with the fit window
Survives held-out walk-forward OOS (Sharpe 1.27, return 13.84%).
oos_sharpe=1.2651 · oos_return=0.1384 · oos_max_drawdown=-0.275 · positive_fold_frac=1 · degradation=0.1446 · n_folds=5 · passed=true
- OOS pass on this dataset only — needs real-market and forward confirmation
- fold dispersion not yet stress-tested
Result is fee-resilient (not fee-fragile): OOS return stable from 13.84% (1x) to 13.23% (5x fees).
base_oos_return=0.1384 · x2_oos_return=0.1369 · x5_oos_return=0.1323 · fee_fragile=false
- fee robustness measured on summary OOS return only
- slippage/impact beyond fee multiples not modeled here
Safety gates held: can_go_live=false, live_broker_inert=true — no live-capital pathway was reachable for this run.
can_go_live=false · live_broker_inert=true · promoted_live=false · capital_mode=none · capital_authority=none · missing_live_keys_count=3
- safety gate proves the kill-switch held, NOT that the strategy has edge
- live broker remains inert (missing keys: AETHERFORGE_BROKER_API_KEY, AETHERFORGE_BROKER_API_SECRET, AETHERFORGE_EXCHANGE_API_KEY)
Best evidence against
The strongest case against the claim. These receipts are preserved and shown even when the headline verdict already reflects them — disconfirming evidence is never buried.
Regime classifier over-calls (dominant label "trend_down" at 31.7%, accuracy 68.1%); regime context is a limitation, not yet load-bearing.
accuracy=0.681 · crisis_recall=0.8 · dominant_label=trend_down · dominant_frac=0.317 · n_regimes=5
- regime classifier over-concentrates / under-accurate — context unreliable
- regime-conditional claims cannot lean on this calibration yet
Proof gaps
What is still missing before this claim could be trusted. Each gap is a reason it is NOT_CAPITAL_READY.
PROOF GAP: real OOS held, but no live-shadow forward confirmation exists yet.
data_mode=REAL · has_real_oos=true · oos_passed=true · run_limitations_count=6
- no forward/shadow window has accrued — only historical real OOS exists
- this gap blocks any trusted / review-ready / capital-ready status until closed
Next admissible test
Fastest path to evidence · Pass the held-out walk-forward OOS gates on real data AND beat every negative control; then capture out-of-time forward data in shadow mode for replication. No promotion occurs on this lane regardless of outcome.
- Overfit collapse: strong validation/in-sample Sharpe that degrades to <= 0 OOS.
- Single-fold dependence: one lucky fold carries the OOS result.
- Fee fragility: a gross edge that turns negative under realistic fees/slippage.
- Regime concentration: edge exists only in the dominant regime.
- Look-ahead / leakage: signal uses information not available at decision time.
- Artifact mechanism: a negative control reproduces the 'edge'.
- Data-quality artifact: gaps/duplicates/non-monotonic timestamps drive the result.
Run limitations
- READ-ONLY research lane: no live orders, no exchange/broker connection, no capital moved.
- Walk-forward OOS is the ONLY proof metric; in-sample/validation are NON-PROOF.
- A candidate that does not beat its negative controls is reported as NOT proof.
- VERDICT CEILING: strongest posture is review_candidate; capital_ready stays false.
- Single CSV source + single seed; needs out-of-time forward (shadow) replication.
- NOT capital-ready: this lane holds no trading authority and cannot promote to live.
Decision record
Research Decision Record
- Run:
realdata_ethusd_1d - Claim:
claim_eth/usd_ma_oos - Decision: Do not promote strategy.
- Verdict:
research_only - Capital ready: NO
- Labels:
NOT_CAPITAL_READYNOT_TRUSTEDRESEARCH_ONLY - Reviewed by:
operator_pending - Created at: 2026-06-10T00:05:00Z
Reason
Live trading is disabled (cangolive=false); the kill-switch held — a supporting safety receipt, not a promotion path. Fee stress was non-fragile (supporting receipt preserved) but cannot rescue a negative-OOS result.
The Proof Foundry proposes; the operator disposes. This record carries no capital authority and makes no unsupported alpha claim.
SigmaX proposes; humans authorize. This record carries no capital authority and makes no allocation recommendation.
This dossier is research evidence only. Every figure above is paired with its limitations and its in-sample / out-of-sample framing — no allocation recommendation is made. Data mode: REAL. Archived record — recorded 93 days ago, not a current observation. Capital authority: none.