Correlation debug endpoint hides the LLM fields, so LLM activity cannot be audited #24

Closed
opened 2026-08-20 03:17:18 -04:00 by logan · 0 comments
Owner

upload.py writes corr_consensus and the corr_llm_* fields (around lines 137 and 151-153), but admin.py:65-93 does not include them in the /admin/debug/correlation response.

Consequence: the one tool built for answering "is the LLM correlation tier actually working" cannot answer it. The 2026-08-19 dump was pulled specifically to check whether the LLM tier had come back to life after the retired-model-ID fix, and the answer had to be inferred from commit dates and field-shape archaeology instead of read off the data.

Given that a dead LLM tier already survived undetected for days once (server-26#14), the debug view not surfacing LLM state is a meaningful gap and a cheap fix.

While in there, two more things the dump showed the endpoint should expose or fix:

  • corr_incident_idle_min can be negative (-4.1 observed), with the incident's updated_at earlier than its started_at. Something is writing those two timestamps out of order.
  • Calls with a skip_reason set are linking anyway and then contributing to incident summaries.

Full analysis: CORRELATION_REVIEW_0820.md (project root, not in git).

`upload.py` writes `corr_consensus` and the `corr_llm_*` fields (around lines 137 and 151-153), but `admin.py:65-93` does not include them in the `/admin/debug/correlation` response. Consequence: the one tool built for answering "is the LLM correlation tier actually working" cannot answer it. The 2026-08-19 dump was pulled specifically to check whether the LLM tier had come back to life after the retired-model-ID fix, and the answer had to be inferred from commit dates and field-shape archaeology instead of read off the data. Given that a dead LLM tier already survived undetected for days once (server-26#14), the debug view not surfacing LLM state is a meaningful gap and a cheap fix. While in there, two more things the dump showed the endpoint should expose or fix: - `corr_incident_idle_min` can be **negative** (-4.1 observed), with the incident's `updated_at` earlier than its `started_at`. Something is writing those two timestamps out of order. - Calls with a `skip_reason` set are linking anyway and then contributing to incident summaries. Full analysis: `CORRELATION_REVIEW_0820.md` (project root, not in git).
logan closed this issue 2026-08-23 01:30:09 -04:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: logan/server-26#24