REGIME-AWARE-PARAMS-1 — Plan-Only

Datum: 2026-06-05 Modus: PLAN-ONLY · 0× Code · 0× Push · 0× Orders · 0× Runtime-Apply Trigger: 180-d-Replay (MS-STRATEGY-REPLAY-20COIN-BACKTEST-1) zeigt RANGE +26 %, WEAK_TREND −30 %, BULL_STRONG −9 % Heutiger XLM-SL-Loss (−11.17 USDT in BEAR-Regime) ist ein Live-Beispiel des nicht-regulierten Risikos

1. Ziel

Per-Regime-Verhalten der MS-Pipeline einführen, um die paradoxe Eigenschaft („trend_follow funktioniert nur in RANGE, verliert in Trends") strukturell zu reparieren — ohne Coin-Allowlist/Denylist, ohne Exit-Parameter-Touch.

Drei Phasen: - P1.A — Shadow-Logger (log-only) - P1.B — Hard-Gate (erst nach 2-4 Wochen Shadow-Daten) - P1.C — Per-Regime-Params (optional, nach P1.B-Pilot stabil)

Diese Plan-Phase deckt P1.A ab und bereitet die Regeln für P1.B vor.

2. Existing Infrastructure (recon)

Im Bot existieren bereits: - MarketRegime Detector (5 Labels: STRONG_TREND, WEAK_TREND, RANGE, HIGH_VOLATILITY, CHOP) - StrategyRouter Hard-Blocks (CHOP + HIGH_VOL ohne base bereits blockiert) - Per-Regime-Strategy-Activation (STRONG_TREND→trend_follow, WEAK_TREND→breakout+trend+oversold, RANGE→mean+vwap+vol_sweep+oversold) - Decision-Log Regime im decision_logs

→ Regime-Detection ist bereits vollständig. Was fehlt ist die regime-spezifische Edge-Filter-Schicht zwischen Router und Final-Decision.

3. Empirische Datenbasis (180d-Replay)

Regime	N	WR	Exp %	Total Net %	Verdict
STRONG_TREND	23	35 %	−0.37	−9	TOXISCH
WEAK_TREND	207	43 %	−0.15	−30	HOCH-TOXISCH
RANGE	261	48 %	+0.10	+26	EDGE positiv
HIGH_VOLATILITY	klein	–	–	–	gedämpft via Router
CHOP	0	–	–	–	bereits blockiert

Kern-Befund: Die MS-trend_follow-Strategie ist eigentlich ein RANGE-Breakout-Detektor.

4. Cross-Impact-Matrix

Changed Area	Direct Effect	Downstream	Follow-up
`multi_strategy_runner.py`	neue `_compute_regime_gate_shadow()`	nur MS-Runner	Tests
`multi_strategy_runner.py`	metadata `regime_gate`	decision_logs, JSONL	KEIN DB-Migration
`settings.py`	6 neue Settings	nur MS-Runner	safe Defaults
`position_manager.py`, `live_trade.py`, `trader_core.py`	0	–	–
Strategy-Subklassen, Router, MarketRegime	0	–	–

Mainnet/MS-Live: unverändert.

5. Regime-Gate-Logik

Klassifikation (Lookup)

Bot-Regime	base_regime	Class	Default Action
RANGE	–	range_safe	ALLOW
WEAK_TREND	–	weak_trend_toxic	HARD_REJECT (P1.B)
STRONG_TREND	–	strong_trend_caution	SOFT_REJECT (P1.B)
HIGH_VOL	RANGE	high_vol_range_caution	ALLOW
HIGH_VOL	trend	high_vol_trend_toxic	HARD_REJECT
CHOP	–	chop	bereits durch Router

Gate-Mode Wirkung

Mode	Toxic	Caution	Safe
shadow (Default)	log, allow	log, allow	allow
soft	reject mit score-Penalty	reject mit score-Penalty	allow
hard	reject	reject (low conf)	allow

6. Settings (NEU)

MS_REGIME_GATE_ENABLED:             bool  = os.getenv('MS_REGIME_GATE_ENABLED', 'true').lower()=='true'
MS_REGIME_GATE_MODE:                str   = os.getenv('MS_REGIME_GATE_MODE', 'shadow')
MS_REGIME_GATE_BLOCK_WEAK_TREND:    bool  = os.getenv('MS_REGIME_GATE_BLOCK_WEAK_TREND', 'true').lower()=='true'
MS_REGIME_GATE_BLOCK_STRONG_TREND:  bool  = os.getenv('MS_REGIME_GATE_BLOCK_STRONG_TREND', 'false').lower()=='true'
MS_REGIME_GATE_BLOCK_HIGH_VOL_TREND:bool  = os.getenv('MS_REGIME_GATE_BLOCK_HIGH_VOL_TREND', 'true').lower()=='true'
MS_REGIME_GATE_SOFT_SCORE_PENALTY:  float = float(os.getenv('MS_REGIME_GATE_SOFT_SCORE_PENALTY', '1.5'))

Cutover-Default: enabled=true, mode=shadow (log-only).

7. Code-Spec

7.1 `multi_strategy_runner._compute_regime_gate_shadow(regime, decision)`

def _compute_regime_gate_shadow(self, regime, decision):
    s = self.settings
    if not s.MS_REGIME_GATE_ENABLED: return {'enabled': False}
    rg = regime.get('regime'); bg = regime.get('base_regime')
    if rg == 'RANGE': cls = 'range_safe'
    elif rg == 'WEAK_TREND':
        cls = 'weak_trend_toxic' if s.MS_REGIME_GATE_BLOCK_WEAK_TREND else 'weak_trend_caution'
    elif rg == 'STRONG_TREND':
        cls = 'strong_trend_toxic' if s.MS_REGIME_GATE_BLOCK_STRONG_TREND else 'strong_trend_caution'
    elif rg == 'HIGH_VOLATILITY' and bg == 'RANGE':
        cls = 'high_vol_range_caution'
    elif rg == 'HIGH_VOLATILITY':
        cls = 'high_vol_trend_toxic' if s.MS_REGIME_GATE_BLOCK_HIGH_VOL_TREND else 'high_vol_trend_caution'
    else: cls = 'unknown_pass_through'
    mode = s.MS_REGIME_GATE_MODE
    if cls.endswith('_toxic'):
        action = 'hard_reject' if mode=='hard' else ('soft_reject' if mode=='soft' else 'allow')
    elif cls.endswith('_caution'):
        action = 'soft_reject' if mode in ('soft','hard') else 'allow'
    else: action = 'allow'
    return {'enabled': True, 'regime': rg, 'base_regime': bg,
            'regime_confidence': regime.get('confidence'),
            'regime_class': cls, 'gate_mode': mode, 'gate_action': action,
            'gate_reason': f"{cls}_in_{mode}_mode",
            'would_have_blocked': action != 'allow'}

7.2 Integration in `_evaluate_symbol`

regime_gate = self._compute_regime_gate_shadow(regime, decision)
if regime_gate.get('enabled') and regime_gate.get('gate_mode') != 'shadow':
    if regime_gate['gate_action'] == 'hard_reject':
        decision = build_REJECT(..., reject_reason='regime_gate_hard', metadata={..., 'regime_gate': regime_gate})
    elif regime_gate['gate_action'] == 'soft_reject' and decision['decision'] == DECISION_TRADE_CANDIDATE:
        new_score = (decision.get('final_score') or 0) - self.settings.MS_REGIME_GATE_SOFT_SCORE_PENALTY
        if new_score < (decision.get('threshold') or 7.0):
            decision = build_REJECT(..., reject_reason='regime_gate_soft', ...)

7.3 Logging

_log_decision JSONL und emit_decision metadata bekommen regime_gate-Key.

7.4 Tests — mind. 20 neu

1-20: shadow/soft/hard × RANGE/WEAK/STRONG/HIGH_VOL × allow/reject/penalty + Kill-Switch + JSONL/DB-meta

Regression-Sweep: phase_n7/n8, ms_mtf, ms_dedup, ms_stablecoin, recon-1, sync_sanity, exit_reason_fix → ~250 grün.

8. Was NICHT geändert wird

Item	Bleibt
MULTI_STRATEGY_DRY_RUN	true
Protected-Toolkit (BE/Partial/Trailing)	unverändert
4h Hold	unverändert
Router Hard-Blocks (CHOP/HIGH_VOL ohne base)	unverändert
MarketRegime Detector	unverändert
`_is_4h_confirming`	unverändert
Strategy-Subklassen	unverändert
Live-Trade-Entscheidungen	unverändert

9. Cutover-Plan (SOT-1d)

Crontab freeze
Backup /root/regime-gate-backup-{TS}/
docker compose build clawbot
docker compose up -d --force-recreate clawbot
3-Way MD5 (2 Files)
Bot healthcheck
Crontab thaw
Live-Verify regime_gate in multi_strategy.log
DB-Verify regime_gate in decision_logs
24h-Monitor

10. Monitoring-Checkliste

Phase A (24h Smoke): 0 Tracebacks, regime_gate-Key in jeder Decision, gate_mode='shadow', gate_action='allow', would_have_blocked aggregierbar.

Phase B (4 Wochen Shadow): Total Decisions vs would_have_blocked, Verteilung über regime_class, Filter-Rate bei hypothetischem hard-Mode, BUY-Rate je Regime.

Phase C (Backtesting): 180d-Replay mit hard-mode auf Shadow-Daten, Vergleich Shadow vs hard, Akzeptanz-Check.

11. Akzeptanz für P1.B-Aktivierung

Schwelle	Wert
Shadow-Daten ≥ 200 TRADE_CANDIDATEs	200
WEAK_TREND-Toxizität bestätigt	WR < 35 %
RANGE-Edge bestätigt	WR ≥ 50 %
Total Net bei WEAK_TREND-Block positiv	≥ +5 % über Shadow-Period
Keine Coin-Allowlist/Denylist	(Boundary)
Operator-GO	ja

12. Risiken

#	Risiko	Severity	Mitigation
1	Regime-Detector falsch klassifiziert	Mittel	seit Phase-N1 produktiv
2	RANGE-Edge im 180d-Sample war Markt-spezifisch	Hoch	Shadow > 4 Wochen
3	Toxic-Klassifikation blockiert echte Gewinner	Hoch	Shadow misst genau das
4	Doppel-Defense mit Router	Niedrig	Router-Blocks bleiben, Gate ist additive Schicht
5	Shadow-Period zu kurz	Mittel	4–6 Wochen
6	env-flip shadow→hard vergessen	Niedrig	Roadmap-Pin
7	metadata_json wächst	Niedrig	~300 Bytes/Row
8	Score-Penalty 1.5 heuristisch	Mittel	Sweep nach Shadow

13. Rollback-Plan

Soft: git revert + SOT-1d rebuild. Hard: env MS_REGIME_GATE_ENABLED=false oder MS_REGIME_GATE_MODE=shadow, recreate. Phase: env MS_REGIME_GATE_MODE=shadow, recreate, weiter Daten sammeln.

Backup: /root/regime-gate-backup-{TS}/.

14. Operator-Entscheidungen

Q	Frage	Default
Q1	Cutover im shadow-Mode?	A Ja, 4 Wochen log-only
Q2	Block WEAK_TREND in P1.B?	A Ja (−30 %)
Q3	Block STRONG_TREND in P1.B?	B Nein (Caution statt Toxic)
Q4	Block HIGH_VOL mit base in Trends?	A Ja
Q5	Soft-Score-Penalty?	A 1.5
Q6	Kill-Switch enabled?	A Ja, default true
Q7	Shadow-Period?	B 4 Wochen
Q8	Cutover-Zeitpunkt?	Wochenende
Q9	Bundle mit MS-MTF-D-SHADOW-LOGGER-1?	A Empfohlen
Q10	P1.B-Flip nach 2 Wochen erlaubt?	B Nein, mind. 4 Wochen

15. Aufwand

~145 LoC + 39 Tests + Cutover-Prep = ~5.5 h.

16. Verhältnis zu anderen Phasen

MS-MTF-D-SHADOW-LOGGER-1: Bundle empfohlen, gleicher Cutover
BTC-MACRO-CONFIRMATION-1: unabhängig, später kombinierbar
LOSS-STREAK-CIRCUIT-BREAKER-1: ergänzend
Strategy-Loss-Backtest-Audit: paralleler P1, blockiert NICHT
MS-Live-Pilot: DEFERRED

17. Boundaries

0× Code-Touch (außer Plan-Phase) · 0× Trading-State · 0× Orders · 0× MS-Live · 0× Mainnet · 0× Env-Änderung (außer optional 6 safe-Default Settings) · 0× DB-Schema/Migration · 0× ConfigProfile-Apply · 0× Bot/Worker-Recreate · 0× Push · 0× Roadmap-Commit · 0× Coin-Allowlist · 0× Coin-Denylist · 0× Protected-Toolkit-Param-Change.

18. STOP-Bedingungen

NICHT umsetzen wenn: Operator-GO fehlt, Bot/Worker unhealthy, State-Coherence-Conflicts aktiv, MS-Dry-Run-Anomalien letzte 24h, Bundle-Partner hat frische Tracebacks.

19. Plan-Closure (per CLAUDE.md §13)

Root Cause: 180d-Replay RANGE +26/WEAK_TREND −30; XLM-BEAR-Live-Loss heute
Files: 2 (multi_strategy_runner.py, settings.py)
Tests: 20 neu + Regression
Cross-Impact: §4
3-Way MD5: post-Cutover Pflicht
Mainnet: weiter verboten
bewusst nicht: Protected-Toolkit, Subklassen, Router, Regime-Detector, Live-Pipeline
Follow-up: P1.B Mode-Flip, P1.C Per-Regime-Params, Score-Penalty-Sweep, BTC-Macro, Loss-Streak

20. STOP

Plan abgeschlossen. Keine Umsetzung. Operator-Entscheidung zu Q1–Q10 + GO EXECUTE REGIME-AWARE-PARAMS-1-P1A-SHADOW erforderlich für Code-Phase.