Classification pipeline

Outputs

category (bug, feature_request, usage_question, visual, confusing, other)
severity (low / medium / high / critical) with calibration — judges score this monthly and feed corrections into the prompt-A/B framework.
component (resolved against your project’s component taxonomy if uploaded).
bug_ontology_tags (cross-customer Bug Ontology — see Whitepaper §2.6).
confidence ∈ [0, 1].
summary (one-sentence triage headline).

Vision air-gap

Stage 2 (“look at the screenshot”) runs in a separate inference call with no shared system prompt. This blocks the canonical OWASP LLM01:2025 prompt-injection-via-image class of attacks: a screenshot containing “ignore prior instructions” can no longer corrupt Stage 1’s classification because Stage 1 already finished before the screenshot was loaded.

BYOK

Per-project keys (Anthropic and/or OpenAI) override the cluster default. Failover order:

Project’s BYOK Anthropic key (if set)
Cluster Anthropic env var (if set)
Project’s BYOK OpenAI key
Cluster OpenAI env var

Every invocation logs which key source was used so finance can attribute spend without inspecting the prompt.