tjarvis91

/

vfai-x-3.5-9b-options

like 2

We are working on the next phase of our financial AI research stack, with the
near-term focus on evaluation discipline rather than public performance claims.

The core direction is simple: financial AI systems need better foundations
before they need bigger models.

The areas we are prioritizing:

• point-in-time data handling, where every row is tied to when it became
  knowable;
• contamination-resistant benchmarks with explicit held-out windows;
• deterministic replay, so the same as-of timestamp produces the same result;
• calibration and abstention metrics, so "do nothing" can be evaluated as a
  first-class decision;
• audit trails for data, configs, benchmark versions, and model outputs;
• risk-aware output formats that are easier to validate and reject safely.

We are intentionally not publishing live trading signals, proprietary datasets,
private training recipes, or unreleased implementation details. The public goal
is to discuss the evaluation and safety discipline around financial AI, not to
ship a black-box trading system.

If you work on financial ML, time-series evaluation, model calibration, or
benchmark contamination, we would be interested in feedback on what you consider
the minimum bar for a trustworthy finance benchmark.

This is research infrastructure and discussion, not financial advice.