Expand description
Generic expected-loss decision framework (bd-2uv9c, bd-3rss7).
DecisionCore<S, A> is the universal trait that unifies all Bayesian
decision points in FrankenTUI. Each adaptive controller (diff strategy,
resize coalescing, frame budget, degradation, VOI sampling, hint ranking,
palette scoring) implements this trait with domain-specific state and
action types.
§Decision Rule
The framework follows the expected-loss minimization paradigm:
a* = argmin_a E_{s ~ posterior(evidence)} [ loss(a, s) ]Every decision produces an EvidenceEntry that records the posterior,
evidence terms, action chosen, and loss avoided — enabling post-hoc audit
via the UnifiedEvidenceLedger.
§Calibration
After each decision, the actual outcome is observed and fed back via
DecisionCore::calibrate. This closes the feedback loop, updating
the posterior for the next decision.
§Fallback Safety
Every DecisionCore implementation must provide a safe fallback action
via DecisionCore::fallback_action. When the posterior is degenerate
or computation fails, the framework returns this action rather than
panicking.
Structs§
- Decision
- Result of a decision: the chosen action plus evidence for the ledger.
- Posterior
- Posterior belief over the state space.
Traits§
- Action
- Marker trait for an action space element.
- Decision
Core - The universal decision-making trait.
- Outcome
- Observed outcome after a decision was executed.
- State
- Marker trait for a state space element.
Functions§
- argmin_
expected_ loss - Compute the expected-loss-minimizing action given a posterior and loss fn.
- second_
best_ loss - Find the second-best loss (for computing loss_avoided).