| name | trade-decision-sim |
| description | The DECISION LENS for trading peers — recalls the CURRENT signal snapshot per market from the live ledger and renders it through the sim arsenal off ONE call: quantum factor-circuit order-optimal sequencing + decision-sim CVaR-robust sizing, sized against the MEASURED per-factor edge (from the edge-audit), correlation-aware. Emits an actionable call per market {side, sizePct, confidence, sequence}. Flat/zero-size when there is no measured edge after cost — the honest default. Sibling of trade-edge-audit (the EDGE lens); the trade-engine analog of quantum-hypothesis-simulation for a live decision. |
Trade Decision Sim — the decision lens
Sibling of trade-edge-audit. Where the edge lens asks "is there real edge in the history?", the decision lens asks "given everything firing RIGHT NOW, what should I do per market — order-optimal and risk-robust?" A capability-wrapper (the quantum-sim pattern): one call recalls the current snapshot and renders it through the whole arsenal.
Engine: _SYSTEM/Scripts/alpha-factor-library/trade-decision-sim.mjs (capability trade-decision-sim, 12/0). It ORCHESTRATES — it does not re-implement the math.
The two moves
1. RECALL (current snapshot). recallSnapshot(ledgerPath) reads the live forecast ledger and takes the LATEST row per factor (its current directional call), grouped by market. That is "what every indicator is saying now."
2. RENDER through the arsenal, per market:
- Net side — the directional vote of the current factors (flat if it cancels).
- Measured edge — averages each factor's
factorEdgeStats (mean dir-return + std) from the edge-audit. This is the SIZING input: no fabricated edge. With today's ~zero measured edge, every market correctly goes flat — consistent with the edge-audit.
- Order-optimal sequencing —
factor-circuit.optimizeFactorCircuit finds the non-commuting optimal factor order; circuitQuality > 1 = a real sequencing advantage (it amplifies edge, it cannot manufacture it).
- Risk-robust sizing —
decision-sim.robustScore (0.5·mean + 0.5·CVaR over normal shocks) gates the size, correlation-aware (effective-N dilution; flat above ρ 0.85), clamped to maxPct.
- Honest gate —
measuredEdge ≤ round-trip cost → flat, regardless of consensus (consensus on a no-edge set is the bounce trap).
Output per market: { market, side, sizePct, confidence, edgeBps, orderOptimalSequence, circuitQuality, rationale } + a verdict. Built to say "flat — no measured edge" when that's the truth (the default right now).
Method map
| Need | Call |
|---|
| Recall current snapshot | recallSnapshot(ledgerPath) → {byFactor, byMarket, factorIds, markets} |
| One market's decision | buildMarketDecision(market, signals, {maxPct, corr, edgeStats}) |
| Full battery + verdict | decideTrades({ledgerPath, maxPct, corr}) → {decisions, verdict, recall} |
| CLI (live) | node …/trade-decision-sim.mjs --decide |
| Self-test | … --decide.mjs --test |
Dual consumption (both peer lanes)
- Native Sonnet agent: "run/extend the decision lens —
.claude/skills/trade-decision-sim/SKILL.md; call decideTrades on the live ledger, return the per-market calls."
- ollama nano-swarm lane: inject this body + deltas (lanes'
read_file is repo-scoped — put any brief in a REPO path, not /tmp, or they can't read it).
Result is ADVISORY until a local run verifies it. INV-1: pure analysis, no order path.
Limitations (honest)
- Sizing is binary in v1 (flat-vs-maxPct): the value function is linear in size, so robustScore is effectively a gate, not a fractional sizer. Conservative + safe; genuine mean-variance/Kelly fractional sizing (concave value term) is a v2 enhancement.
- In-code cost floor (round-trip 5bps) — the real Coinbase taker floor is ~20bps (retail ~120bps). The gate uses the engine's in-code cost for consistency; the edge-audit owns the real-fee analysis. At real fees the bar is far higher.
- Edge from history, decision from now — sizing uses each factor's historical measured edge; the call uses the current snapshot side. Sign-aligned (a current side opposite the factor's historical edge → negative center → flat).
- No proven edge yet — so the lens is flat everywhere today. That is correct, not broken; it activates when a factor develops real, cost-clearing, multiple-testing-surviving edge (per the edge-audit).
Sibling
trade-edge-audit (/edge-audit) — the EDGE lens: is there real, overfit+FDR+fee-surviving edge in the history. Run it to find edge; run this to act on it.
Session Notes
2026-06-17
- tools: Bash, Read, Edit, Write, Agent (Sonnet red-team), capability-scan
- origin: owner directive — build the decision lens PEER-FIRST (kimi/nemotron/minimax/glm/flash/mimo swarm; "lanes do most of the work, you verify/wire/fine-tune").
- swarm reality: kimi-k2.7-code delivered a strong core (interfaces verified against source — factorVector/optimizeFactorCircuit/robustScore shapes all correct, genuinely read the source). nemotron (sizing-math) ETIMEDOUT-retry-stalled and blocked the wave barrier so glm/flash never fired; minimax couldn't read the /tmp brief (lanes' read_file is repo-scoped). I covered the missing roles myself. LESSON: no
wait barriers between independent lanes; put briefs in a repo path.
- operator fixes over kimi's draft: replaced a FABRICATED 1.5% edge with the MEASURED edge from the edge-audit (the honest core — connects the two lenses); clamped maxPct (red-team HIGH — unclamped → leverage); gauss shocks not uniform (crypto tails); market-filter on edge lookup (cross-market contamination guard).
- Sonnet red-team: FIX-FIRST on the maxPct clamp (fixed + regression-tested); edge-sign alignment, interfaces, gate logic, category slots, fail-soft all confirmed CORRECT. 12/0; live = all markets honest-flat (no measured edge), consistent with the edge-audit.
