Market briefs

A brief is the structured output of one portfolio decision pass — the AI's full reasoning, target book and close list.

Briefs are the most important artifact in the app. Every order has a parent brief, and every brief has the inputs that produced it: window boundaries, signal count, event count, current open positions, AI model, and the raw model response. If you want to know why the agent did something, open the brief.

The 4-step cycle

Read — pull fresh news + events in the window.
Collate — combine them with the current open portfolio.
Decide — ask the AI for a verdict on each open position + a target book + a close list.
React — the executor opens, trims and closes positions accordingly.

Position verdicts

For every currently open ticker, the AI returns one of:

hold — keep the position as-is.
trim — reduce exposure but stay in the trade.
close — exit the position.

The executor only closes positions the AI explicitly tells it to close. Silence on a ticker defaults to hold. This is intentional — it prevents accidental flushes when a new target book happens to omit an existing ticker.

Target book vs close list

Target book — what the portfolio should look like after this pass. Drives new opens.
Close list — explicit closes. Plus any position with a close verdict.

Rotation rules

The brief picks closes carefully. Three runtime guards sit on top of the AI's verdicts to keep the book stable:

No rotation while the book isn't full — if open positions are below portfolio_max_positions, any "rotate out for a better idea" close is demoted back to hold. Empty slots get filled with new ideas instead.
Rotation-cost guard (only when book is full) — to replace an incumbent, the new candidate's confidence must exceed the incumbent's entry confidence by at least rotation_confidence_margin (default 0.10).
Loss-aversion guard — stale-but-not-negated theses sitting at a loss are not realised. Only thesis negation, direction flip, or locking a deteriorating winner counts as a valid close.
Pyramiding — reinforcing news on an existing position increases its target_weight_pct rather than opening a new slot.

Position sizing

portfolio_max_positions — single source of truth for how many concurrent positions the agent can hold (the older max_open_positions key was retired).
portfolio_session_budget_pct — share of the configured starting capital each pass may deploy into new exposure.

Key fields

Field	Type	Description
trigger	—	What caused this pass: scheduled session, breaking news, event cluster, or imminent event.
window_start / window_end	—	Time range of news + events considered.
input_signal_count	—	How many fresh signals fed into this pass.
input_event_count	—	How many fresh events fed into this pass.
open_position_count	—	Positions open at the start of the pass.
target_book	—	Array of { ticker, direction, target_weight_pct }.
close_list	—	Tickers the AI explicitly wants closed.
position_verdicts	—	Array of { ticker, verdict, reason } — one per open position.
narrative_md	—	Free-form AI narrative explaining the overall stance.
ai_raw_response	—	Verbatim model response, kept for auditing.