← Academy home
Deep Dive · Sentinel · Part D of D

The Verdict Card

Four parts of anatomy, forensics and torture, condensed into one deployable contract: what Sentinel is allowed to do, what we expect of it, what we'll watch, and the exact numbers at which we pull it from duty.

D.1The verdict

DEPLOY — restricted, respected, watched

Sentinel goes into the forward campaign on USDJPY H1 only, at modest size, under the guardrails below. The architecture is sound (A), the config sits on plateaus (C), costs can't explain it away (C) — but its edge is single-pair and partly unexplained (C), and its river may be slowing (B). We deploy it the way you'd employ a brilliant specialist with an unexplained method: gladly, on their proven ground, with supervision.

D.2Identity card

Deployment USDJPY · H1 · demo · magic 7702

Signals defaults + wae_dead_zone_pip: 20 · pip_value: 0.01 (JPY — never forget)

Risk ATR×1.5 stop · 1:3 target · 3-bar cooldown · 1% risk sizing

Forbidden other pairs (transfer disproven in C.1) · H4 (failed both ways in Round 4) · size increases before the forward test earns them

D.3Baselines and wobble bands — what "on track" means

The forward test is graded against these, with Lesson 11's honesty about sample size:

metricexpectationsource
profit factor (long run)≈ 1.165y full history, real costs
profit factor (recent regime)≈ 1.06Round-4 sealed exam (2025→)
win rate≈ 35%1,876 trades
trade pace≈ 250 / yearobserved
wins after first 50 trades10–24 is all normal±2σ band, p=0.348
The patience table — luck's range when the edge is INTACT

Monte Carlo, 4,000 alternate futures with the edge fully healthy: over 60 trades (a quarter), PF can sit as low as 0.55 by pure luck — a quarter proves nothing. Over 250 trades (a year), the 5th percentile is PF 0.83, the 1st percentile 0.72 — and a healthy Sentinel still posts a losing year 23.5% of the time.

This is why the kill criteria below sit under the luck floor. Anything above them is weather, not climate.

D.4The pain contract

Signed before the campaign, so live losses are met with recognition instead of fear (numbers from C.6, at 0.1 lots on $10k):

Streaks 11 straight losses = the median year · 15 = merely unlucky · only beyond 20 is statistically alarming

Drawdown ~2.6% typical · 4.5% at 1-in-10 · 6.6% at 1-in-100 — single digits throughout

Losing months expected regularly; a losing quarter proves nothing (see patience table)

The rule during any streak inside these bands: re-read this page, check the watch metrics, change nothing

D.5Watch metrics — the dashboard's standing orders

Spread average live spread at Sentinel's entry times. Break-even at 2.6 pips (C.5); alarm at sustained > 1.5

The river rolling 6-month net USDJPY drift. Sentinel is long-skewed drift-dependent (B); a flat or reversing river (2026's PF 0.78) is the leading indicator of edge decay — yellow when 6-month |drift| < 3%

Rolling PF over the last 250 trades, compared against the patience table — the climate gauge

Reason audit monthly: read 5 random Sentinel reason strings against the chart (Lesson 12 ritual) — catches software rot no metric sees

D.6Kill criteria — the numbers, agreed in advance

triggerthresholdaction
rolling 250-trade PF< 0.72BENCH — worse than 1-in-100 luck; the edge is gone or broken
rolling 250-trade PF0.72 – 0.83REVIEW — re-run the deep dive analyses before continuing
losing streak> 20PAUSE — beyond the 99th percentile; inspect before re-arming
drawdown (at 1% sizing)> 8%PAUSE — beyond anything Monte Carlo produced
avg entry spread, 2 weeks> 1.5 pipsREVIEW costs; > 2.5 = BENCH until conditions normalise
river gauge6-mo |drift| < 3% for 2 quartersREVIEW — the habitat itself may be gone

Benching is not deleting: a benched Sentinel keeps paper-trading (steering wheel disconnected) while we diagnose — exactly how the Proving Ground would want it.

D.7Open items ledger

Verify WAE formula vs the real MT5 indicator on a USDJPY chart (A.3.5 — El's eyeball task; urgency reduced by B.1 but fidelity still matters)

Code make the SMA slope window a parameter (hardcoded diff(3); it's the dominant vote) · derive pip_value from symbol automatically (the A.3.5 trap, fixed by config today)

Engine native trailing exit (HA reversal beyond NonLagDot) still unsimulated — low stakes per C.7, revisit if live behavior diverges

Revisit this whole deep dive after the first 250 forward trades — the verdict card is a living document

That closes the deep dive: anatomy understood, myths tested and half-broken, dials proven plateaued, pain rehearsed, and a contract for deployment. The watchman goes to work — with you, at the gate, holding this card.