Σ₀ EV Gate — the trader's Converge stage

Plain-language summary

The trader collects evidence from several detectors — Grok's read, Claude's read, whether price is at a real support/resistance zone, a 1-minute structure shift, a candle pattern, the higher-timeframe trend, and ticker news. The old design ran these as pass/fail gates (Riley's WAIT/GOOD/PERFECT tiers, plus a Claude HOLD veto), which exists to stop a human from overtrading. unisona.ai doesn't need that crutch — it can weigh every piece at once and act on expected value.

The EV gate turns each candidate trade into a Convergence Record (one of unisona.ai's four core objects: hypothesis + evidence + confidence + result) and makes a single, auditable ENTER/SKIP call.

The model

Each signal is normalised to [0, 1] where 0.5 = neutral. The win probability is a transparent linear model:

p_win  = base_rate + Σ wᵢ · (signalᵢ − 0.5) · 2

EV (R) = p_win · target_R − (1 − p_win) · 1R     # risk 1R to make target_R

ENTER  iff has_evidence AND EV ≥ EV_MIN AND p_win ≥ P_MIN

• base_rate — the realized win-rate for this ticker+direction (≥5 closed

trades), else the 90-day backtest hit-rate, else a 0.5 prior. This is the External-Reality anchor: the prior is grounded in measured outcomes, not optimism.

• target_R — reward:risk (tp% / |stop%|).
• EV_MIN = 0.15, P_MIN = 0.45 — require a real +0.15R edge and a 45%

hit-rate, so a huge target alone can't carry junk.

• has_evidence — the evidence floor that replaces Riley's discipline tiers:

at least one grounding signal must be real (a zone, a structure shift, a graded pattern, trend agreement, or a meaningful news lean). A bare +EV from an optimistic reward:risk is not enough to act.

Evidence weights

Weights are how many probability-points a fully-confirming signal (1.0) adds. They need not sum top_win(p_win is clamped to [0.05, 0.95]), and the council re-weights them from realized edge over time.

backtest is folded into base_rate rather than carried as a signal.

Grok + Claude are council members, not gates

Grok and Claude used to be a single opaque llm signal, and Claude additionally ran after the EV gate as a hard HOLD veto that could kill a trade the EV had already approved. As of 2026-06-30 both are absorbed into this one council:

1. Split signals. grok and claude are separate, individually-graded

signals (llm_conf stays a back-compat alias → grok_conf; claude defaults to neutral 50). The per-signal realized-edge table now shows each model's edge on its own, so a model with no predictive lift gets re-weighted down.

1. Σ₀ EV is the sole decider. After Claude returns, the EV re-scores with

both convictions — Claude BUY/SELL agreeing with direction → 78, HOLD → 35 (a lean against, not a veto), opposing → 15 — and that combined EV makes the ENTER/SKIP call. A Claude HOLD now only lowers p_win; it no longer unilaterally vetoes a +EV trade.

1. Risk stays enforced. Portfolio risk (max positions, delta limits) is still

enforced downstream by agent_risk_manager — which is what Claude's HOLD was redundant with.

A cheap Grok-only pre-screen (Claude neutral) runs first as a token-saver: clearly-below-bar candidates SKIP before Claude is ever called. Set SIGMA0_EV=0 to restore the legacy Claude-as-hard-gate behavior.

The Convergence loop (Observe → … → Converge)

Every executed ENTER persists a ConvergenceRecord to the shared store (data/convergence/records.jsonl); on close, a Brier-graded outcome is appended to data/convergence/trader-outcomes.jsonl:

brier = (confidence − outcome)²        # outcome = 1 win / 0 loss

The Σ₀ council (sigma0-trader-council.js) reads those outcomes and computes:

• calibration — canonical Brier / ECE / skill grader. Grades **only

conviction-bearing rows** (conviction_recorded !== false): a trade backfilled with a 0.5 no-information prior (e.g. a cross-project trade whose system never logged its conviction) is excluded so it can't pile into the 0.5 Brier bin and fake calibration. Such rows still count toward win_rate / avg_pnl_pct as realized-outcome volume. The snapshot reports graded (all rows), graded_with_conviction (calibration set), and graded_with_signals (per-signal-edge set).

• per-signal realized edge — for each signal, the win-rate when it fired

strong (>0.55) vs weak; lift = strong_winrate − weak_winrate. A positive lift means the signal genuinely predicted wins and earns (more) weight; ≈0 or negative means it's noise. This is the evidence for re-weighting the EV.

The council is warming until ≥20 conviction-graded trades, then ready. It surfaces at GET /api/trading/sigma0/calibration.

Warm-up from historical data

scripts/warm_sigma0_council.py backfills the council from real closed trades so it starts ready with a live per-signal edge table instead of cold. Two externally-grounded sources, every row tagged source and idempotent (re-runnable; remove with grep -v '"source": "warmup'):

• lanternOS lessons.db — closed trades joined to records.jsonl by ticker +

nearest timestamp for the full signal vector and the Σ₀ p_win. Outcomes are recomputed from the actual Alpaca fill prices, because the stored pnl_pct column is corrupted (sign-flipped on the TP-zone / EOD rows). These rows carry the signal vectors that drive per_signal_edge (the re-weighting evidence).

• Independant AI Trader trading.log —EOD closed TICKER: X% outcomes

over ~7 weeks (2026-05-11 → 06-30) from a separate system. No signal vector and no recorded conviction, so they carry confidence = 0.5 (no-information prior) and conviction_recorded: false: they ground the win-rate/volume but the council structurally excludes them from calibration and the maturity gate (and they carry no signals, so per-signal edge already ignores them).

External-Reality grounding. Nothing synthetic is injected — each row is a real realized outcome. The intraday fills are the broker's own record (the authoritative external anchor); a second independent system corroborates the win-rate regime. The council's strongest emergent signal — news sentiment (lift ≈ +0.27) — is directionally consistent with the literature on short-term/intraday return predictability (e.g. news sentiment ranking second only to volume in LSTM studies), though that literature also cautions the economic significance is contested — fitting a thin, early council. Treat the warmed edges as a prior, not a verdict, until live closed trades accumulate.

External-Reality Rule

Every record carries why — the signals that moved the decision, strongest first — so the call is auditable: [claim, evidence, confidence, source]. Nothing is accepted without evidence; the evidence floor and the realized-win-rate base rate are where this rule bites.

Tuning knobs

Weights, EV_MIN, P_MIN, and P_CLAMP are module constants in convergence_ev.py; the council's per-signal lift table is the data you use to adjust them.

Tests

tests/test_convergence_ev.py pins the contract: confirming evidence raises p_win monotonically, the gate only fires on a real positive edge, news is signed to direction, Grok and Claude are graded separately, llm_conf back-compat holds, and every decision carries an auditable record + why.