The methodology, in full.

Catalyst Runner is the first of seven specialist agents and the only one that trades the public paper-trading account. The agent runs against a narrow, audited tool allowlist:

• lookup_ticker — resolves a ticker to a canonical company record via the ticker registry.
• get_catalysts — fetches upcoming and recent catalysts (earnings, FDA, Form 4, STOCK Act) for a ticker over the configured horizon.
• fetch_filing — pulls an EDGAR filing (10-K, 10-Q, 8-K, Form 4) by accession number for direct quotation in the thesis.
• propose_idea — emits a candidate Idea for the six-layer verifier to evaluate. Only verified ideas surface to you.

Each user has a private memory namespace so prior research context informs subsequent prompts without leaking across accounts. The wire output is a single Idea shape with seven flat fields — ticker, catalyst_type, thesis, horizon, conviction, expected_30d_pnl, citation_count — wrapped in a CardCatalystSSEEvent envelope that carries the adapter attribution sha and an optional compliance flag.

Every candidate idea passes through six independent layers. A failure at any layer rejects the idea and it is never published. The layers are:

• Ticker Registry. Asserts the proposed ticker resolves to an exchange-listed, currently-trading symbol via the canonical ticker registry. Rejects delisted symbols, fictional tickers, OTC pinks unless explicitly enabled.
• Catalyst Horizon. Asserts the cited catalyst falls inside the configured horizon (default 7–30 days) and that the catalyst date is in the future relative to the proposal timestamp. Rejects post-dated catalysts and ideas referencing yesterday's news.
• Theme Taxonomy. Asserts the catalyst type maps to a known taxonomy entry (earnings, fda, form_4_cluster_buy, stock_act, options_flow_unusual, 13f_new_position, etc.). Rejects free-text "theme" labels the downstream systems cannot consume.
• Numeric Freshness. Asserts every numeric value cited in the thesis (price, market cap, revenue, EPS) has a timestamp within the configured staleness window (default 7 trading days). Rejects ideas built on stale fundamentals.
• Conviction Calibration. Asserts the conviction score (0–10) is internally consistent with the breadth and quality of citations — high conviction requires both numeric grounding and multiple independent sources. Rejects ungrounded high-conviction calls.
• Citation Completeness. Asserts every non-trivial claim has at least one citation pointing to a reachable, source-of-truth document (an EDGAR filing or a market-data record). Rejects ideas with bare-assertion theses.

A failed idea is never seen by you. The aggregate rejection rate is published at /methodology/verifier-log — refreshed every 30 minutes, prompt text is never exposed.

Factor agent is the second of seven specialist agents and the first signal-side agent to ship after Catalyst Runner. The agent runs against a narrow, audited tool allowlist:

• get_factor_scores — fetches the day's composite decile and per-factor z-scores for a ticker from the internal factor-compute service.
• expose_factor_breakdown — returns the 9-factor per-stock breakdown for the cardFactor explainer drawer.

Factor scores are computed nightly (Mon–Fri at 22:00 UTC, one hour after US market close) from source-of-truth feeds: the Fama–French 5-Factor Daily CSV from the Ken French Data Library (Tuck School of Business), plus daily and intraday price bars from multiple market-data sources. The score set is the FF5 triplet (market, size, value), the FF5 quality pair (profitability, investment), 12-2 price momentum, and the Asness–Frazzini–Pedersen Quality-Minus-Junk triplet (profitability, growth, safety). Before z-scoring, each characteristic is winsorized — clipped to its 1st-to-99th-percentile band within the day's universe — so a single extreme outlier among the broad set of names cannot inflate the cross-sectional spread and distort the published scores. The clipped characteristics are then z-scored within the day's eligible universe — a liquidity-screened set of US-listed names with >$100M 30-day average dollar volume — and the composite decile (1 through 10, 10 best) is the equal-weighted average of the 9 z-scores ranked into deciles.

Every cardFactor surfaced in chat carries the verifier envelope; the Factor agent's Idea outputs pass through the same six-layer anti-hallucination verifier as Catalyst Runner. The factor methodology is locked at services/factor-compute/methodology.yaml and pinned by version on every factor_scores_daily row — re-running the compute with a new methodology emits new rows under a new version without overwriting history. The full locked YAML is open-source — view methodology.yaml on GitHub.

The aggregate hallucination-rejection rate is public. Counts only, not prompt text. Refreshed every 30 minutes. Read the verifier log →

Per-event-type win rates on cardCatalyst describe the share of historical instances of the same event type that beat SPY over a 30-day window, not the share that simply went up. The excess-vs-SPY framing strips out market beta — a 62% absolute win rate during a bull-market window is not edge; a 62% win rate relative to SPY across the same window is.

Every edge claim on a published card carries a visible sample size n. When fewer than 30 historical instances of the event type are available, the card surfaces an EXPLORATORY chip instead of a win-rate claim. Thirty is the canonical floor for the Central Limit Theorem's normal approximation and the floor used in the post-earnings-announcement-drift literature (Chan, Jegadeesh, and Lakonishok, 1996).

Skalarity will tell the reader when it does not have an edge. Cards on cold-start event types — categories where the corpus is still accumulating — pass the verifier and reach the public portfolio, but they carry the exploratory chip and do not claim a win rate. When the corpus crosses 30 historical instances, the chip retires and the standard claim chrome takes its place.