Skip to content

Routing computed fact blames routing for filter denials and missing devices (#399 follow-up) #439

Closed

Description

@ARSH871-bot

Read on main (ccb2591, #399's merge):

  • The flat branches at ai/explain.py:424-425 and :442-443 say "the network's current routing is what disagrees" for every disposition. DENIED_IN and DENIED_OUT are filters, not routing. EXITS_NETWORK is, per routing.py:34-41, most often a device missing from the upload. The model is told this fact is "already verified", and tests/test_explain_safety_net.py:441 pins the wording for a NO_ROUTE + DENIED_IN case.
  • Producer drift: tests/test_explain_safety_net.py:370 still hand-writes the "Expected … for a flow from" prefix, copied from routing.py:485-486. Build the detail through routing.run() with a fake Session instead.
  • The mixed-result text says "(one ended in X)" even when more than one path failed (:414).

Suggested fix: state the observed disposition without a cause, or say "left the part of the network that was uploaded" for EXITS_NETWORK. Measured on 94bc75c, before the merge: routing was blamed in 8 of 8 runs on both shapes. The blame wording is unchanged at the merged head.

Activity

  1. added
    bugSomething isn't working
    layer: aiLayer 2 — local LLM explanation and Q&A
    safetyTouches F-4, grounding, or the refuse-rather-than-guess boundary. Review with care
    on Nov 3, 2026
  2. added a commit that references this issue on Nov 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't workinglayer: aiLayer 2 — local LLM explanation and Q&AsafetyTouches F-4, grounding, or the refuse-rather-than-guess boundary. Review with care

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions