From closed bars to governed signals.

The system moat is visible. The alpha moat is not proved.

Crypto Signal Lab implements deterministic signal definitions, point-in-time evidence, a ten-gate decision path, cost-aware labels, uncertainty controls, replay, and promotion governance. A separate research lane can produce descriptive insights, but it cannot create a production alert.

The value of this result is falsifiability. The system can now distinguish an implemented mechanism, a prospective runtime event, a retrospective experiment, and an unproven product claim. Those categories are not added together to manufacture confidence.

The difficult part is not another indicator. It is preserving authority across the whole evidence chain.

Every mathematical ingredient has precedent. The present differentiation is the orchestration: data availability, deterministic hypotheses, independent challenge, calibrated authority, cost-aware utility, uncertainty, attention, immutable outcomes, and delivery remain separate but reproducible.

Demand is established across four adjacent product categories. Their centers of gravity remain different.

This is a category scan based on official product pages, not a feature-by-feature competitor scorecard. It validates that users pay attention to alerts, derivatives state, on-chain intelligence, and attention analytics. Signal Lab is positioned around the governed process connecting those observations to a local research record.

Official pages accessed 28 July 2026. Product surfaces change. The table describes stated category emphasis and does not claim that any named product lacks unlisted capabilities.

There are two lanes, and only one can create a production alert.

The separation is the most important architectural fact. Production detectors create governed candidates. Pattern Studio, VIC, Chronos, and Qwen create research artifacts or explanations that cannot cross into production authority.

Four registered hypotheses, normalized for risk and challenged after detection.

Time-series momentum

The implemented score averages five volatility-normalized momentum horizons: 24, 72, 168, 336, and 720 hours. Each component is capped before averaging; a candidate requires absolute score of at least 0.25.

Donchian breakout

Prior-range breaks use 24, 168, or 360 hourly bars. Efficiency ratio must reach 0.30 and governed cross-venue breadth must reach 0.60. Strength combines path efficiency, breadth, and distance beyond the channel.

Compression-expansion

A 24-hour hypothesis requires a prior squeeze, a 24-bar Donchian break, current volume at least twice its trailing median, and breadth of at least 0.60. The 90-day compression percentile creates a 2,160-hour warm-up requirement.

Derivatives fragility

A weighted score combines funding extremity, positive open-interest change, basis magnitude, and liquidation intensity. It emits typed risk-off context at 0.70, with at most one missing component and weight renormalization.

Three real research emits. Zero production alerts.

At the 28 July 2026 00:00 UTC closed bar, the research lane emitted down-continuation insights for BTC, ETH, and SOL. They were delivered at information priority and remained provisional because historical same-side outcome and utility evidence were unavailable.

Production telemetry recorded 120 detector evaluations: 33 each for 4-hour TSMOM and Donchian, and nine for each 24-hour or 72-hour policy. Every row ended precondition_failed / detector_row_missing. There were no production candidates, gate decisions, promoted champions, or alerts.

The implemented TSMOM score is weak in this first public-data benchmark.

The experiment used 22,536 contiguous Binance spot hourly bars for each of BTC, ETH, and SOL from 1 January 2024 through the 28 July 2026 boundary. Canonical raw-row hashes are retained in the public manifest.

An episode is the first candidate bar after a non-candidate state or direction change. Net return uses a transparent research proxy: 20 bps round-trip taker fees, 2 bps round-trip fallback spread, and two square-root-impact legs at USD 10,000. It is less complete than the governed minute-fill label.

What the evidence permits us to say.

Use HFT-style correlation discipline, then demand calibration and economics.

Information coefficient evaluates the continuous score before a threshold creates selection effects. It is the right first test, but not a complete product metric.

Primary statistical tests

• Pearson and Spearman score-to-forward-return IC by asset, horizon, side, regime, volatility, and data-quality state.
• Joint timestamp-block bootstrap across assets so BTC/ETH/SOL dependence is retained.
• Residualization against broad market direction, realized volatility, and BTC beta to distinguish exposure from incremental information.
• Brier skill, reliability slope/intercept, selective-risk curves, and interval coverage for probabilistic outputs.
• Realized net return after fees, spread, impact, funding, gas, and user-specific local costs.
• Romano-Wolf correction for production claims; false-discovery controls only for exploratory discovery.

Minimum prospective proof

Current governance requires at least 30 unchanged days and 50 closed shadow outcomes, feature parity of at least 99.5%, bounded probability deltas, decision-to-completion p95 no more than 60 seconds, p99 no more than 180 seconds, modeled-versus-observed cost deviation no more than 25%, and no corruption incidents.

The methods have lineage. The composition still has to earn evidence.

1. Moskowitz, Ooi, and Pedersen (2012), Time Series Momentum. Continuation hypothesis.
2. Moreira and Muir (2017), Volatility-Managed Portfolios. Volatility scaling and risk normalization.
3. Liu and Tsyvinski (2021), Risks and Returns of Cryptocurrency. Dedicated crypto factor and risk evidence.
4. Makarov and Schoar (2020), Trading and Arbitrage in Cryptocurrency Markets. Cross-venue fragmentation.
5. Cont, Kukanov, and Stoikov (2014), The Price Impact of Order Book Events. Flow and impact motivation.
6. Page (1954), Continuous Inspection Schemes. Two-sided CUSUM lineage.
7. Newey and West (1987), HAC covariance estimation. Overlapping-return inference.
8. Politis and Romano (1994), The Stationary Bootstrap. Dependence-aware resampling.
9. Brier (1950), Verification of Forecasts Expressed in Terms of Probability. Probability scoring.
10. Bailey and Lopez de Prado (2014), The Deflated Sharpe Ratio. Search-aware performance evidence.
11. Romano and Wolf (2005), Stepdown Methods for Multiple Hypothesis Testing. Production-claim multiplicity control.
12. Howard et al. (2021), Time-uniform Confidence Sequences. Prospective monitoring without fixed-horizon peeking.
13. Franc, Prusa, and Voracek (2023), Optimal Strategies for Reject Option Classifiers. Abstention as a first-class decision.
14. Ansari et al. (2024), Chronos. Probabilistic pretrained forecasting lineage; research context only here.

The downloadable PDF preserves this selected research grounding. The public evidence bundle publishes the experiment metrics, source hashes, runtime counts, and derived event datasets used by the argument.

Public evidence is published beside the argument.

The experiment used frozen public Binance hourly snapshots with 22,536 rows per asset and 2,000 moving/stationary block-bootstrap resamples. The evidence bundle contains no keys, tokens, wallet data, private endpoints, or execution records.