The EYE

Live router · shadow mode
connecting…
?

Market pattern · live

Multi-TF candle patterns + regime + ATR · refreshed every 30s
CLEAR 0/7 atr —
1m
5m
15m
WPM β-score (1m·0.65 + 5m·0.35)
Direction signal
Threshold gates

Pattern symbols: ▲ ⏶ bull strong · bull weak · ▼ ⏷ bear strong · bear weak · ◇ ◯ indecision · ⇕ ⌇ special. Hover for full name. WPM thresholds: AA β70 ≥ 0.70, AB β50 ≥ 0.50.

Regime · detector vs pattern shadow

Live comparison · candidate to replace slow detector after 24-48h validation
— agreement
DETECTOR (slow, hysteresis)
— · loading
vs
PATTERN (1m+5m+15m, fast)
— · loading

Reading: when labels disagree, pattern is usually right (it sees fast structure changes the detector misses). After 24-48h shadow data, run diag-regime-shadow-compare.js to validate and possibly flip pattern source live.

Engine status · 5m decision pipeline

Today's timescale-purity pass · Iron Rules v2 compliance
config: —

Pattern integration · Phase 0a

w_pattern_1m: — w_pattern_5m: — shadow log: —

Timescale-purity gates

chop_skip: — trend_dampen: — funding_contra: — vwap_revert: —

Retired strategies (timescale leak)

position_1d trend_cont triple_screen

Iron Rules v2: all engine config defaults OFF · plans opt-in via params.engineFeatures · per-plan kill switch via params.bypassEngineMutations. See docs/IRON_RULES.md.

Regime detector · LEGACY (slow classifier · being phased out)

— · loading…
LEGACY
— loading —
Vol ratio30m σ ÷ 24h baseline
Ret 30mclose-to-close
Ret 1hclose-to-close
Range 30m(max-min) ÷ mid
Vol 30m / 1hstddev of |movePct|
Sample5m decisions used

Phase 1a: 8-state classifier (VOLATILE_UP/DOWN, TRENDING_UP/DOWN, CHOP, POST_PUMP/DUMP, DRIFT) with 3-tick hysteresis. Updated every 1 min from btc5m_decisions. Use the history table below to verify state stability before wiring this into plan routing (Phase 2).

Regime detector · transition history

Last 30 state flips (only flips are persisted; tick-by-tick stays in realtime/market_regime).
Time (BKK) From → To Vol ratio Ret 30m Ret 1h Sample
loading…

Hourly regime attribution · supervised-learning loop

— · loading…

Each hour at :05, the system looks back at the just-completed hour, replays the detector per 5m boundary, groups consecutive same-state ticks into segments, and attributes each plan's positions to the regime that was active at decision time. Forward = label at start of hour. Hindsight (mode) = mode across detector ticks. Objective = stateless price-action label (true ground truth — independent of detector). Forward vs Objective is the real accuracy measure.

Detector accuracy
vs objective truth
hours segments last 7d
Stability filter:

Per-regime data-driven champions · sorted by frequency-of-podium (top-3 share); ★ subscript = top-1 share · last 7d

Regime Segments Total min Top 3 podium (frequency in top 3) Cumulative PnL leaderboard
loading…

Forward vs objective truth · confusion matrix · last 7d · diagonal = correct · falls back to mode-hindsight if objective unavailable

loading…

Hourly timeline · last 48h · stability badge: 🟢 stable ≤2 segs · 🟡 mixed 3 · 🔴 noisy ≥4 · Match column uses objective truth

Hour (BKK) Forward Hindsight (mode) Objective Match Segments breakdown Pos Top 3 plans Hour PnL
loading…

Removed stale sections (Shadow-mode data was empty). Active plans + per-plan stats live on /polymarket-v2.html. Regime detector + hourly attribution (above) is the active learning loop; Phase 2 routing will activate once 7d objective-truth accuracy stabilises.