Glossary

Words that mean something specific here, in alphabetical order. Each entry links to the page that explains it properly.

Batch. Two separate things wear this name. On a call, batch is a flag saying the request went through a provider’s Batch API, which is priced at that model’s batch rate wherever the catalog carries one. In the ingest API, a batch is the unit of delivery: the SDK posts rows in groups, and a successful POST /v1/ingest names the batch it accepted them under. See Model catalog and pricing and Ingest API.

Campaign. The evaluation layer’s term for a run of replays over captured traffic. The evaluations capability still gates the /v1/campaign prefix, but no campaign route is registered today. Analysis runs are what replaced it. See How analysis works.

Canary. A route pinned to a candidate model for a deterministic share of sessions. model_for() exists in the SDK and is safe to call anywhere, but it is a no-op today: there is no write path to the canaries table, so the config the SDK polls is always empty and model_for() returns the default you passed it. See Python SDK.

Capability. A named product surface that a deployment either has or does not. Gating is by deployment profile, never by plan, and an absent capability answers 404 capability_unavailable rather than pretending to work. See Limits and allowances.

Deployment profile. Which of managed, byoc-core or local a server is running as. It decides the capability set. The local profile gets the full managed set; only byoc-core narrows it. See Choosing a deployment.

Detector. One of six deterministic checks that run over your routes: runaway-spend, cost-drift, failure-spike, retry-loop, latency-drift and template-change. A detector that fires raises an alert. See Set up alerts.

Environment. A free-form label saying where a call ran, from METERGRAPH_ENV or the SDK’s environment argument. Calls with no environment are shown as untagged rather than hidden. See Configuration.

Equivalence grade. The verdict an analysis gives one replay. A is an equivalent outcome. B is the same outcome with only non-impacting differences. C is an impacting difference that could change a downstream decision. F is an unusable result or an execution failure. A and B count as safe; C and F do not. See How analysis works.

Function. The code that made the call. In Python this is automatic: the SDK walks the stack to the nearest function under your app root. In TypeScript it needs track(), because bundlers and minifiers destroy the names stack attribution would rely on. See Identity model.

Ingest key. A workspace credential, mg_ followed by 48 hex characters. Keys minted in the dashboard carry the ingest scope only; read and agent:read come from the operator CLI. See Manage your workspace.

Recommendation. A concrete, costed change an analysis produced for one workload: this model instead of that one, with the equivalence grades and cost evidence behind it. Only a workload whose evaluation carried a valid opportunity becomes one; everything else stays diagnostic coverage information. See Run your first analysis.

Route. The product surface a call belongs to, independent of which function made it. It is the unit cost is compared on and the unit a cheaper model is tested against, which is why an unrouted call can be measured but not optimized. See Name a route.

Session. The end-to-end interaction a call belongs to, usually a conversation or a job. It is what a recorded outcome attaches to, and therefore what makes cost per completed task computable at all. See Record real outcomes.

Session token. A short-lived ingest credential, mgs_, which the SDK gets by exchanging your app token at POST /v1/ingest/sessions. It expires after an hour on the hosted service, and after five minutes by default on the open source server. Its job is to keep a long-lived key out of routine ingest traffic. See Ingest API.

Span. One model call: the row Metergraph stores. If you send OpenTelemetry instead of using the SDK, a span is a GenAI span, and spans that are not GenAI spans are counted as rejected. See OTLP spans.

Tag. An arbitrary key and value you attach to calls, for the slicing none of the other labels covers: customer tier, feature flag, tenant. See Identity model.

Trace. A group of spans that form one unit of work: every call a single checkout, or a single agent turn, made. See Identity model.

Unpriced. A coverage bucket, not an error. A model the catalog does not recognise produces no cost, and the call lands in unpriced so the gap stays visible instead of being silently counted as zero. See Model catalog and pricing.

Workload. A cluster of similar traffic that an analysis found inside a route, and the unit it evaluates. One run covers the three eligible workloads with the highest sample counts. See How analysis works.