Quant Core, Human Gate: An AI Investor-Committee for Daily Stock Plans

Abstract. This paper describes a daily stock-selection system with a deliberate division of labor: a quantitative engine ranks a universe of stocks by a validated 0-100 confidence score, and an AI acting as an "investor committee" researches the shortlist for catalysts, risks, and reasons to say no. The output is a fully allocated buy/sell/hold plan - never an order. The design question it explores is where language-model judgment genuinely adds value on top of a quant core, and where it must be fenced out.

The problem

Pure quant rankings know what has had statistical edge but nothing about why now: an earnings date tomorrow, a downgrade this morning, a lawsuit nobody has priced. Pure discretionary stock-picking has the opposite failure: stories without a base rate. The system pairs them - and then confronts the governance question that most "AI trading" projects skip: what is the AI allowed to do?

Method

The daily loop runs in five steps. A collector pulls about a year of daily history for the universe. The engine computes features and a confidence score per name, validated by backtest before deployment, and emits a shortlist of roughly fifteen candidates. The committee step is where the language model works: for each candidate it web-researches current catalysts and tail risks and writes a conviction record - a 0 to 1 conviction, a one-line catalyst, the key risk, and a verdict from BUY to AVOID:

{ "NVDA": { "conviction": 0.7, "verdict": "BUY",
            "catalyst": "earnings beat + guidance raise this week",
            "risk": "priced for perfection; any datacenter pause hits hard" } }

The allocator then merges quant score and committee conviction into a plan that is always fully invested across five to fifteen names, with per-name caps (30% for BUY-rated names, 10% otherwise) and weights that must be earned from data plus a stated catalyst.

The governance rules are hard-coded into the operating prompt: the system never places, cancels, or modifies a real order (a human executes every plan); it touches only a designated cash account; and it is required to state realistic expectations in its own output - the prompt itself instructs the agent that strong weeks are single-digit to low-double-digit percentages and that anyone implying 50% weekly returns is lying.

Results

The system runs daily in production as a plan generator, with its output feeding a reporting journal. No audited performance record has been published, and this paper deliberately reports none: the plan-quality data that exists has not passed through the same audit discipline this library requires of performance claims (chronological validation, cost realism, and a sample large enough to mean something). What can be reported is operational: the pipeline produces a complete, sized, reasoned plan daily, and the plan-only boundary has held by construction - the executing account's order permissions were never granted to the agent.

Limitations

The committee layer inherits every weakness of language-model research: it can be confidently wrong about a catalyst, it reads the same headlines everyone else reads, and its conviction scores are calibrated by instruction rather than by a measured track record. The quant core's validation predates live deployment and decays like all such validations. And the honest structural point: a system without an audited live record is a design study, whatever its daily outputs look like.

Full implementation available on request.

Key definitions

Confidence score - A quantitative ranking from 0 to 100 assigned to each stock by the system's backtested statistical engine, reflecting the strength of historical edge before any catalyst analysis.

Catalyst - A near-term, identifiable corporate or market event (earnings announcement, regulatory filing, competitor action) expected to move a stock's price or valuation.

Conviction - A 0-to-1 probability estimate assigned by the language-model committee layer reflecting confidence in the investment thesis after catalyst and risk research.

Backtest - Historical simulation of a trading strategy using past price and fundamental data to validate that its ranking rules had statistical edge before live deployment.

Plan-only boundary - An operational constraint that restricts the AI agent to generating fully reasoned investment plans while forbidding it from executing, placing, or modifying real orders.

Base rate - The historical frequency or statistical probability of an outcome in a defined population, used to anchor discretionary judgments against pure narrative reasoning.

References

• CME Group, "Backtesting Best Practices", CME Education. Https://www.cmegroup.com/education
• CFTC, "Automated Trading Systems", Division of Market Oversight. Https://www.cftc.gov
• Federal Reserve, "FRED: Federal Reserve Economic Data", Board of Governors. Https://fred.stlouisfed.org
• Investopedia, "Backtesting", Investopedia Dictionary. Https://www.investopedia.com/terms
• Investopedia, "Catalyst", Investopedia Dictionary. Https://www.investopedia.com/terms
• SEC, "Principles-Based Approach to Capital Markets Regulation", Division of Corporation Finance. Https://www.sec.gov/divisions/corpfin

Educational research on historical data only - not investment advice, not a signal, and never a performance promise. Past results do not predict future performance. Drafting uses AI assistance; every citation is link-verified before publication and every paper is re-audited weekly against the library's editorial standard. Last reviewed by the PropLedger research pipeline: 2026-08-26. Educational research on historical data; not financial advice.