Skip to content

Glossary

Term Meaning
Trace One complete end-to-end execution of an AI request.
Span One operation within a trace (an LLM call, a retrieval, a tool call, and so on).
Span link A data-flow edge: span B consumed the output of span A (relation: consumes_output_of).
Component key Stable name of a pipeline step, such as rag.reranker, that stays the same across runs. Used to match expected root causes in regression runs.
Quality signal The result of one check on a span or trace: a rule, a statistic, a DecisionGate decision, an LLM judge or a user flag.
DecisionGate Interface for fast bounded classifications with a confidence score (PASS/FAIL/REVIEW, failure family, evidence sufficiency).
Jev The bounded-decision model named in the original blueprint, used as one DecisionGate backend. An LLM classifier is the fallback. Open question: what Jev is and how we access it (ADR-004).
Failure family / category Two-level taxonomy, such as RETRIEVAL / RERANKING_FAILURE. Defined once in taxonomy.py.
Root cause The earliest causal operation that introduced the significant failure.
Propagation Downstream reuse or transformation of an earlier incorrect state.
Span role A span's part in a diagnosis: healthy, root_cause, propagated, degraded, failed_output, unknown.
Evidence Stored trace data supporting a diagnostic claim (EvidenceRef).
Grounding gate Step that checks every claim in a diagnosis against stored evidence and removes unsupported claims.
Eval case A reusable input and expectation pair used to test future pipeline versions.
Regression A previously passing case that fails after a change.
Fault injection Deliberately creating a controlled failure with a known root cause in a benchmark app.
MVP / Target / Stretch Scope tiers. MVP ships by M2 (29 Jan), Target by M3 (12 Feb), Stretch only if time allows.
ADR Architecture Decision Record, stored in Confluence and docs/adr/.
UAIS Used in the blueprint headers. Open question: confirm what it stands for (possibly the club's name).