File size: 1,667 Bytes
2527e91 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 | # Evaluation Protocol
## Router boundary
A router may receive only the public task, observable state, public task context,
execution memory, recent events, budget, qualitative agent profiles, and candidate
action identifiers. It must not receive admissible/progress/efficient labels,
transition definitions, action outcomes, or hidden simulator state.
## Track A
For each row in the `decision` configuration, construct the public observation and
request one action. The primary outcome is membership in `admissible_actions`.
Report at least:
- admissible next-action accuracy;
- macro template admissible accuracy;
- ambiguous-state accuracy;
- post-failure and post-rejection accuracy;
- counterfactual-pair accuracy;
- STOP precision and recall;
- distractor selection and efficient-action rates.
## Track B
For each row in the `interactive` configuration, start from the initial state and
run the router against the deterministic simulator until a terminal condition.
Report at least:
- goal completion rate;
- macro template completion rate;
- admissible and efficient action rates;
- recovery success;
- action rejection and stagnant-repeat rates;
- loop and premature-STOP rates;
- step and cost regret on successful episodes.
## Ablations
Use the same model, provider, seed, dataset revision, and split when comparing:
- `naive`: minimal instruction, role name/goal only, no guard;
- `structured`: full role contracts, failure-aware instruction, observable
rejection-repeat guard.
This two-arm comparison measures the complete orchestration-engineering package.
Additional factorial arms are required to attribute gains to individual elements.
|