All selected models receive the same market snapshot and prompt version.

The product standardizes inputs on purpose so differences come from model behavior rather than ad-hoc prompt drift.

Model output must follow strict JSON schema.

Sanity checks validate target bounds and recommendation coherence.

When validation fails, controlled retries are applied.

Quotes and multi-year chart history can come from different providers. In practice, a live quote may be fresher intraday while the long-range chart remains daily end-of-day data.

When a primary market-data source fails, controlled fallbacks are used rather than silently mixing invalid tickers or broken payloads into the workflow.

Recent research on LLM stock recommendations shows that outputs can vary with rephrasing and prompt setup, so human review still matters.

Recent supervisory guidance in investment services also points to overreliance, opacity, and data quality as core AI risks.

That is why ModelAtlas is built around comparison, disagreement, and explicit risk framing before action.

Hit-rate tracks directional consistency between recommendation and realized move.

Model Edge League shows historical counts, hit-rate and directional return.

Metrics are descriptive, not promises of future performance.

Last updated: April 7, 2026.