Python: Sample idea: function-invocation filter for physics-validating agent actions #14785

Hi SK team - our PhysicsConsistencyScorer was recently merged into microsoft/PyRIT (PR #3507) as a rule-table-based physics-consistency module for red-teaming workflows (deterministic in execution: fixed rules, no model in the loop). The SK filter pipeline (FilterTypes.FunctionInvocation) looks like a candidate scenario for the same capability in agent workflows, and we would like to check whether an official sample would fit - to be judged by maintainers.

The gap: in the filtering samples we reviewed on main (python/samples/concepts/filtering/, as of 2026-11-12), the samples cover logging, redaction and telemetry patterns; we did not find a physics-consistency filter sample in that review. Such a filter would intercept function calls whose arguments are physically impossible - e.g. set_temperature(-300), rotate_motor(rpm=1e5) beyond rated limits, move(distance=-5) - before they reach the actual tool. By design it uses a pure rule table: no extra model calls, no new dependencies (design goals, to be verified in the implementation), with a three-state exit (pass / degrade / veto). Overhead not yet measured.

Proposal: a sample python/samples/concepts/filtering/physics_guard_filter.py following the existing sample structure (same signature: context, next; same style as function_invocation_filters.py).

Question: would this class of sample fit the official samples? If the direction fits, we can prepare a PR after your confirmation.

• PHM-Audit