How we validate a signal before it ships.
SENTINEL is a public crypto-perp research desk. Most ideas don’t survive contact with the data — and we show that work. This page is the validation pipeline, the route simulations, and what’s in the lab right now.
Current figures are requested on each visit from the same endpoint our admin dashboard uses. A retained last-good snapshot or static fallback is labeled. “Research” and “reconstructed” are not a track record.
Every tier runs the same gauntlet: parity backtest, shadow collection, then prediction-outcome review. Runtime-forward public CORE rows are recorded at signal fire. Their resolved values are prediction evidence, not exchange fills, customer trades, or customer PnL.
Current CORE window
Current route-simulation evidence for the CORE tier family (sweet currently delivered; liqsqueeze shadow-only), loaded from /api/scanner/tier-performance. Bars below 1.0 stay visible; thin samples stay flat. Resolved prediction outcomes stay separate from this chart.
$1,000 through returned prediction rows
Outcomes now resolved for predictions-table CORE signals fired in the trailing 14 days, compounded in signal-time order. Outcomes may resolve after their signal time. The 14-day request returns at most 200 rows and the all-time request at most 2,000, so the returned walk can be smaller than the full count. Zero outcomes remain flat. The reconstructed backtest keeps its own tab and caveats.
The Skeptic’s Console
$1,000 through … returned CORE prediction outcomes · … resolved outcomes for CORE signals fired in the trailing 336h · API row cap 200
Outcomes now resolved for predictions-table CORE signals fired in the trailing 336 hours, ordered by signal time. Outcomes may resolve later than their signal time. This walk uses only the newest returned rows, not necessarily the full count; the all-time request is capped at 2,000. Wins come from a positive realized outcome, losses from a negative outcome, and zero is flat. Modeled at 10% of equity per row. Not exchange fills, customer trades, or customer PnL. Inspect the returned rows on /losses.
How candidates earn live delivery
Discovery
Mine candidate signals; reconstruct outcomes path-aware from history.
Shadow soak
Run live but invisible for ≥12 weeks, ≥20 entries, ≥10 resolved exits.
Latency stress
Re-test under realistic execution lag — idealized edges that collapse are cut.
Prediction outcomes
Accrue resolved outcomes for runtime-forward public CORE rows recorded at signal fire. Not execution evidence.
Graduation
Promote only after clearing predeclared profit-factor + sample gates.
Why we test under real-world latency
An exit policy that looks great at zero latency can fall apart once a real alert-to-fill delay is applied. We ship the one that holds up, not the one with the prettiest backtest.
time_up is the live policy (accent). Idealized trailing policies look good at zero latency but collapse under real alert-to-fill delay. Reconstructed from the parity backtester.
Live validation posture
What still needs evidence before promotion.
The research lane keeps the same live route telemetry visible while marking where lab evidence is preliminary, thin, or paused for sample risk.
Telemetry is still loading.
Model and alert diagnostics
Model tuning diagnostics
Public view of ML edge evidence and threshold-risk posture, with invalid or stale measurements withheld.
Loading model telemetry...
In the lab right now
HTF Supertrend
soak closed · context, not a signalHigher-timeframe (weekly) trend-rider. The 12-week live-shadow soak completed 2026-08 and closed with a negative promotion verdict: the pure supertrend flip failed out-of-sample controls (weekly OOS PF 0.44, 12h 0.70), and only the daily-confirmation read held up (OOS PF 1.28). The lanes stay live as market CONTEXT powering the Trend Board — deliberately not promoted to a standalone signal. The early reconstructed ~PF 2.0–2.5 estimate did not survive validation.
Exit policy · time_up
latency-robustRetains 87% of profit factor under a 30s execution lag, versus 35–42% for trailing-stop policies. Shipped live behind an instant env kill-switch.
HTF-trend overlay
tested · no edgeTested whether a trend filter lifts the live tiers. It did not survive a cluster bootstrap (confidence interval spanned zero). Not shipped — a deliberate negative result.
The honesty contract
- ›We don’t present prediction outcomes as exchange fills, customer trades, or customer PnL.
- ›Reconstructed and shadow numbers are labeled as such — never as a track record.
- ›Negative results ship too: an idea that fails validation is a finding, not a failure.
- ›Small samples make us more cautious, not more excited.
Verify, then decide.
The lab page explains how ideas graduate. Verify the public proof first, then connect wallet to continue in Telegram.
Need route context first? Verify the live ledger.