The system, fully drawn out.

When the home page says "5 layers", it means the math that turns four raw agent scores (fundamentals, technical, sentiment, news) into a single composite for ranking. It is not a count of agent stages — those are a different concept entirely. Each layer is a pure function and runs in order. The output is a single number per candidate plus a trace[] that the future "explain this trade" UI can replay.

Two separate pipelines run in production: discovery (find a new name, debate it, execute) and reeval (rescore everything we already hold). They share some agent names —bull_researcher, bear_researcher, consensus, execution_agent — but run on different triggers. The swap (rotation) pipeline is a third path that composes elements of both. Roles are dispatched by canStartAgent() — stage N cannot begin until every required agent in stage N−1 has reported complete.

Add the three swap-only roles — head_to_head_pm, risk_neutral_swap, risk_validator — and you get the full count of eighteen distinct agent roles across the production runtime. See §05 for how those three compose with the discovery agents to execute a rotation.

The PM is not a chatbot wrapped around a portfolio. It is the operator interface to the whole desk. Twenty-six named action tools place real orders, drive cycles, change configuration, swap LLM models, and schedule rotations. Thirteen math tools answer quantitative questions deterministically (no LLM-faking-numbers). When neither category fits, a Python execution fallback runs arbitrary calculations in a sandbox. Every destructive action returns a requiresConfirmation envelope so the operator gets a second look before anything fires.

The line on the dashboard that reads 30D forecast: +4.8 % ±3.5 % · conf 0.74 is not vibes. It's a deterministic synthesis from four primitives — composite score, direction, recent residual mean, and confidence — extrapolated along a 30-trading-day anchor and √t-scaled to any other horizon. Pure function, no I/O, returns a discriminated { ok: true | false } so callers fail-soft on missing inputs rather than fabricating.

μ_anchor = ((compositeScore − 50) / 50) × 10
           + 1  if direction = up   |   −1  if direction = down
           + 0.5 × clamp(recentResidualMeanPct, −2, +2)

σ_anchor = 6 − 2 × confidence              (range 4 % – 6 %)

μ_h      = μ_anchor · (h / 30)             √t-scaled to horizon h
σ_h      = σ_anchor · √(h / 30)

μ (mid): the expected return percentage. Composite drives the bulk of it; direction adds a ±1 % nudge; recent residual error adds a tiny mean-reversion bias.
σ (spread): shrinks as input confidence rises, but never below 4 % at the 30-day anchor.
Output confidence: re- derived separately from the synthesizer's input confidence — it's the mean of per-signal availability flags so the UI can suppress forecasts on stub inputs (no price history, no fundamentals, no catalyst).
Bias: the directional sign applied to μ_anchor — visible on the dashboard so the operator can see which way the forecast is leaning before reading the number.

Until 2026-05-18 this math only ran on ranked discovery candidates. The Position Sentinel rescores every held position too — so the same primitives now drive the per-holding cards on the dashboard.

A swap is not a sell-then-buy. It's a single atomic proposal — "exit sell-leg, enter buy-leg" — that has to pass two risk gates back-to-back, with the PM deciding first instead of last. The ordering inverts the discovery pipeline because the question being asked is different: discovery asks "is this name worth holding", rotation asks "is this name worth holding more than the one we already hold".

Four surfaces are wired but not fully fed. The page above is the fiction-free version; this section is the rest of it.

• Correlation matrix needs price history

  The correlation_matrix math tool computes pairwise ρ across symbols you supply, but a fully populated per-symbol price series isn't routinely cached. When history is missing the tool fails-soft and the PM tells you so. Wiring an opportunistic backfill is on the list.

• L3 Black-Litterman runs as a no-op on cold cells

  Until enough agent_accuracy_rolling rows accumulate for a given (sector, regime, horizon) cell, the L3 tilt is zero. The layer is still in the pipeline — it just contributes nothing to the composite for those cells. As the trade log grows, more cells light up. Cited at composite-score.ts:134-156.

• Federated posterior is opt-in

  The L2 shrinkage layer accepts an optional third anchor — the average posterior across the whole user base. The plumbing works; the feeder that aggregates per-user posteriors into a federated row runs offline. Cold-start users currently see the sector-profile prior only.

• L0 z-score skips small cohorts

  When a sector has fewer than three tickers in the cohort at decision time, the L0 layer is skipped (composite-score.ts:95-97) and the composite uses raw 0–100 scores into L1. This is intentional — a 1-ticker z-score is meaningless — but it does mean thinly-covered sectors briefly lose the regime-detrend.

If any of the above changes status, this page must move first. The code is the contract; this page is its mirror.