A research script can score several symbols independently and still produce an impossible combined portfolio. Each candidate may appear affordable on its own while collectively exceeding the same cash balance or exposure limit. AI Trader Paper explores this problem through an offline-first Python prototype with one shared fictional portfolio per market and session.
Its pipeline moves from input bars to technical features, a small machine-learning model and heuristic scoring. Cost-aware sizing and a second cost check come before an intent or a local fill. In paper mode, symbols are evaluated in a deterministic sequence against a shared engine, so earlier fills affect the cash and exposure available to later candidates. That ordering is reproducible; it is not a claim of optimal portfolio selection.
Portfolio state includes cash, positions and equity valued using supplied prices. Illustrative fees and fill impact affect the simulation, and a session loss cutoff latches once breached. Protective exits can remain available after new entries are blocked. These mechanisms make assumptions visible, but a threshold is still a trigger rather than a guarantee that simulated losses can never exceed it.
The dry-run mode has a different meaning. It evaluates independent proposals without changing the portfolio or reserving capital. Those proposals must not be treated as a jointly approved allocation. The offline demo uses synthetic observations and fictional balances; it may reject every candidate. Its generated movement is not a historical backtest, and the model's score is not a calibrated probability of success.
Version 0.1 is deliberately limited to a one-shot, in-memory session. Restarting creates a new account; there is no durable journal, exchange calendar, continuous position monitor or broker adapter. Stops only react when another price is supplied. Durable state and portfolio-level out-of-sample evaluation belong to the documented roadmap. The project is a transparent software sandbox for studying those decisions, with no evidence claimed for profitable returns.