engine room / answer trace

wrun_01M1VSYM9CJAQ4BVDEVQ8C8NCW

The guardrail refused this message (off topic) before any model ran; the visitor got a scoped redirect in 7.51s.

Visitor identity is a one-way hash; message content never enters traces.

Timeline

Guardrail

off_topic

6.86s
0first token 7.19s7.50s

The gray spans are durable-workflow dispatch: each stage is an independently retried, checkpointed step, and the handoffs are what that crash-safety costs.

What happened

  1. Admitted at the edge

    A 19-character message from caller 3e48ef9d passed schema validation, the per-caller rate limit, and both token budgets (charged an estimate of 1,029 tokens, settled later against actual usage).

  2. Became a durable workflow

    The endpoint never calls a model. At 16:48:44.848 the request was enqueued as run wrun_01M1VSYM9CJAQ; tokens stream into the run’s persistent log, which is why a page reload can reattach mid-answer.

  3. Refused by the guardrail

    gpt-oss-safeguard classified the message against policy 6bb17bbe0908 in 6.86s: off_topic.

    The request asks for private salary information, which is not covered by the allowed scope.

    The visitor got a scoped redirect; no retrieval ran and no generation was paid for.

Record

Status
finished (refused-off_topic)
Admission
admitted
Guardrail reason
The request asks for private salary information, which is not covered by the allowed scope.
Tokens in / out
0 / 0
Cost
$0
Model
zai/glm-5.3-flash
Prompt version
45b79cdd7ce5
Policy version
6bb17bbe0908
Input size
19 chars
Caller (hashed)
3e48ef9d22e096da
Request id
d00517df-5335-41d0-99ec-28b1f07d752f
Enqueued
2026-09-06T16:48:44.848Z
Finished
2026-09-06T16:48:52.353Z