What it measures: every time a strategy signal fires, the live engine and the shadow orchestrator both decide whether to take the trade. This panel counts how often they agree vs disagree.

Data source: orchestrator_divergence_events — one row per signal, written by engine_signals.evaluate at decision time. Scoped to your user_id only.

How to read it:

• Agreement rate — the orchestrator and live engine took the same side. Promotion requires ≥80%.
• Divergence rate — they disagreed. Not bad by itself; only matters when paired with cohort PnL.
• Dangerous divergence — they disagreed on direction OR the live path traded when the orchestrator would have blocked. Promotion requires ≤10%.
• By-symbol breakdown — useful to find symbols where the two paths systematically disagree.

What it measures: the AI model emits a confidence score per prediction (0.50–1.00). This panel groups predictions into 5-percentage-point buckets and checks whether higher-confidence buckets actually win more often.

Data source: confidence_bucket_events — one row per prediction outcome (next-bar direction match), NOT per closed trade. Scoped to your user_id only.

Important caveat: "Win Rate" here = directional prediction accuracy on the next bar, NOT trade PnL win rate. A 95% win rate in the 0.55 bucket means the model called direction correctly 95% of the time, not that you made money on 95% of trades.

How to read it:

• Monotonicity PASS — higher confidence reliably means higher EV. This is what promotion requires.
• Monotonicity FAIL — a higher bucket performs worse than a lower one. Either the model's confidence is mis-calibrated, or one of the buckets has too few samples (look at the bucket counts).
• Recommended threshold — minimum confidence where EV > 0 and PF > 1. Suggests where to set the AI gate cutoff.

What it measures: the COUNTERFACTUAL value of the orchestrator. For every signal the live engine took, this panel asks "would the orchestrator have blocked this?" and groups the resulting trades by that hypothetical decision.

Data source: divergence events joined to closed trades by symbol + time. PnL comes from the trades table (source of truth). Scoped to your user_id on both sides of the join.

How to read it:

• BLOCKED cohort — trades the live engine took that the orchestrator would have rejected. A negative total PnL here is GOOD: it means the orchestrator's filter would have saved you those losses.
• ALLOWED cohort — trades both paths agreed on. These are the orchestrator's "endorsed" trades. Positive PnL means the orchestrator's picks made money.
• Measurement state PARTIAL — some divergence events couldn't be linked to a closed trade within the matching window. Numbers reflect only what got linked; "unmeasured count" shows the gap.

Caveat: matching uses ±10 min around the trade's open_time, with a fallback to a 6 h window before close_time for broker_backfill rows (which have lost their real open timestamp).

What it measures: when the orchestrator is in sizing-only mode (AI_ORCHESTRATOR_SIZING_ENABLED=true), it scales position size by a multiplier (e.g. 0.8× or 1.2×). This panel checks whether the multiplier correlates with realized PnL.

Data source: engine_execution.get_scale_performance — joins recorded scale decisions with trade outcomes.

How to read it: if higher scales (1.2× ×) systematically out-earn lower scales (0.8×), the orchestrator's confidence-to-size mapping is doing its job. If they're random, sizing is adding noise.

Note: empty when sizing-only mode hasn't been enabled — the engine hasn't recorded any scale decisions yet.

What it measures: when the orchestrator rejects a signal, it records a reason. This panel breaks down rejections by reason and tracks the confidence gate's hit rate.

Data source: orchestrator_divergence_events.rejection_reason for blocks; confidence_bucket_events for the gating numbers.

Block reasons:

• Low confidence — model probability fell below the gate.
• Risk filter — exposure/correlation limits hit.
• Direction conflict — orchestrator's direction prediction contradicts the strategy signal.
• Probability threshold — calibrated probability fell below the per-symbol cutoff.
• Validation policy — a hard policy rule fired (e.g. blackout windows).

Confidence gating block rate — what fraction of all evaluations got rejected purely for falling below the confidence threshold. A high rate means the threshold is tight; a low rate means most decisions pass through to the strategy layer.

What it measures: the orchestrator is rolled out in four stages. Each stage has a pass/fail criterion based on the data above. This panel surfaces the exact metric vs threshold for each gate so you can see WHY a stage is failing.

Stages, in order:

• Shadow — orchestrator runs alongside the live engine but doesn't influence trades. PASS when agreement rate ≥80% AND dangerous divergence ≤10% AND ≥100 evaluations recorded.
• Sizing — orchestrator scales position size but doesn't change direction. PASS when shadow passed AND scale-fidelity mismatch ≤15% AND calibration monotonicity holds.
• Confidence — orchestrator's confidence gate can veto live trades. PASS when sizing passed AND there's a recommended threshold AND each bucket has ≥20 samples.
• Full Takeover — orchestrator replaces the live engine entirely. PASS when confidence passed AND no per-symbol degradation.

Thresholds are configurable via the PROMOTION_* env vars (see config.py) — the defaults are conservative.