adaptive-gate-tuning

name

adaptive-gate-tuning

description

Tune or audit ML Engine gate thresholds as a coordinated adaptive system. Use when confidence, momentum, risk, or agent-consensus thresholds are being changed; when one static gate bottlenecks an otherwise adaptive pipeline; or when the scanner alternates between over-trading and total suppression. Focus on coupled gate behavior, regime awareness, drawdown state, virtual trades, and safe one-change-at-a-time tuning.

Adaptive Gate Tuning

Use this skill when adjusting gates that decide whether a setup becomes a trade.

Core principle

No gate is independent in production. Confidence, momentum, risk, and agent-consensus thresholds interact, and the system must be tuned as a coupled control loop.

Goals

• identify the true bottleneck gate
• prevent stacked suppression across multiple gates
• avoid over-loosening in response to a single dry spell
• preserve regime-aware and drawdown-aware safety

Workflow

1. Establish the current control surface

• Identify every active gate and threshold involved in approval.
• Separate:
  • static thresholds
  • adaptive floors
  • drawdown-dependent tightening
  • regime-dependent adjustments
  • execution-quality penalties

2. Use raw evidence before tuning

• Read recent virtual trade outputs to see where setups are dying.
• Confirm whether the dominant failure mode is:
  • confidence
  • momentum
  • risk
  • agent consensus
  • post-gate execution rejection

3. Check for asymmetric adaptation

• If some gates are adaptive and one remains static, test whether the static gate became the bottleneck.
• Explicitly audit weighted_vote_threshold and any other hardcoded consensus rules when adaptive floors exist elsewhere.

4. Quantify interaction effects

• Measure how multiple penalties and floors combine.
• Look for:
  • confidence penalties stacked with disagreement penalties
  • drift or calibration penalties
  • regime multipliers
  • drawdown tightening
  • execution-quality adjustments
• Treat total suppression as a systems problem first, not a single-threshold bug.

5. Apply the smallest safe tuning change

• Prefer one bounded change at a time.
• Prefer completing missing adaptive logic over broadly lowering thresholds.
• Preserve hard safety invariants around portfolio risk and minimum R:R.

6. Verify by regime and recent outcomes

• Check whether the proposed tuning behaves differently in LOW, NORMAL, HIGH, and EXTREME regimes where applicable.
• Confirm the change would have helped recent false negatives without obviously admitting low-quality setups.

Repo-specific anchors

• src/scanner/config.py
• src/scanner/agents/
• src/scanner/automation/continuous.py
• src/scanner/automation/orchestrator.py
• src/scanner/automation/qa_pipeline.py
• .claude/learnings.md
• .claude/rules/improvement.md
• virtual trade and gate-attribution artifacts

Output contract

Return:

1. Active gate map
2. Dominant bottleneck with evidence
3. Interaction analysis across gates
4. Recommended bounded tuning change
5. Verification plan by regime and recent trades

Guardrails

• Do not loosen multiple gates at once unless the evidence clearly shows coupled suppression.
• Do not weaken risk controls to compensate for confidence or consensus issues.
• Do not propose tuning without naming the exact gate and expected directional effect.