Capture record
Evidence, before outcome.
Every source is timestamped and hashed before a model sees it.
That creates an audit trail for what was knowable at the time. Open any row to inspect the original document.
Model judgments
The decision, not hidden.
Attention, exposure, forecast, and confidence for the same frozen document.
These are pilot traces, not benchmark results. Model names disappear when the preregistered evaluation begins.
Triage simulation
Who reads the right file?
Each model gets the same batch and an attention budget of 3 documents.
Reward is F1 against strict-majority materiality consensus, with recall capped by the available budget. The policies are frozen. Nothing learns during this pilot.
Adaptation
Judgment tuned to a house view.
Materiality is a mandate-specific call. This loop measures how much of a desk's taste transfers into the judgment.
A reference mandate stands in for a customer's research desk and labels a larger stream of captured documents — the preference data a desk produces naturally. An adapted judge then sees only those labeled examples, never the mandate itself, and judges held-out documents the way the desk would. Preference post-processing carries the taste live on this page; the same labeled corpus drives reinforcement fine-tuning on customer document sets privately.
Loading the reference mandate…
Hear when the clock starts.
Three emails only: preregistration, evaluation start, and published results.
This is an operational pilot, not a ranked benchmark or an investment product. The design may change before preregistration. No investment advice is given or implied.