2026-03: Dataset autonomy, self-tuning, graph propagation, and macro kill-switch

The investment stack now does more than replay a fixed dataset universe with static heuristics.

It can start widening and correcting its own research surface while staying policy-gated.

What changed

repeated uncovered theme pressure can now produce guarded historical dataset proposals
replay-safe dataset proposals can be auto-registered into the scheduler registry
the engine now keeps an experiment registry for guarded self-tuning and rollback of weight profiles
graph propagation can surface hidden candidates that do not depend only on direct trigger keywords
the live decision path now computes a macro overlay with hedge bias, exposure caps, and a kill-switch state
idea cards and direct mappings now carry explainable attribution across corroboration, graph support, beta, macro pressure, and execution-reality penalties

This still is not an unconstrained execution bot.

It is closer to a constrained autonomous research loop:

it can discover missing research inputs
it can adjust itself instead of leaving every coefficient frozen
it can reason beyond direct keyword matches
it can stand down when top-down market stress overwhelms attractive micro themes
it can explain why a recommendation survived or failed