Glossary

Every term Broks Forge uses, defined once. These names are used consistently across the product, this documentation and the API — the same word always means the same thing.

Agent — an AI system registered for evaluation, identified by its HTTP endpoint. Framework-agnostic.

AI Engineering Operating System — the category Broks Forge belongs to: a system that models the engineering act behind AI systems and reasons over it. See the category page.

AI Git — version control for engineering reasoning: revisions, promotions, rollbacks and the rationale behind each change. Not source control. See AI Git.

Artifact — anything the platform reasons about: agent, prompt, dataset, provider, model or evaluation.

Benchmark — a comparison of variants over the same dataset.

Blast radius — everything downstream that a change to an artifact would affect. See Evolution.

Brief — a standing, derived reading of the record. Eight exist: Daily, Deployment, Incident, Prompt, Evaluation, Dataset, Knowledge, Architecture. Each follows what happened → why → evidence → impact → recommendation → next action.

Brok — the Engineering Partner. Answers engineering questions from the record, deterministically. See Brok.

Claim — an assertion about an artifact, supported by evidence. A derived object.

Confidence — a three-step verbal ladder: consistent with, likely, near-certain. Never a percentage.

Contradiction — a claim that sits uneasily beside the evidence, for example a settled canonical revision alongside failing evaluations of the same artifact. Reported as an inference.

Decision — an engineering choice with a reason: promoting a revision, deprecating an artifact. A derived object, carrying the rationale recorded at the time.

Derived — computed on read from stored records; never persisted. Also an epistemic status meaning "read directly from real records".

Engineering Intelligence — the reasoning layer: observations, claims, decisions, evidence and knowledge derived from real work. See Engineering Intelligence.

Engineering Memory — the answers to "why is this the way it is?", derived from decisions and recalled verbatim. See Engineering Memory.

Engineering record — the whole body of stored artifacts plus everything derived from them. What Brok and the Root Cause Explorer reason over.

Epistemic status — how a statement is known: derived, inferred, suggested or unknown. Exactly one per statement.

Evaluation — a reproducible measurement of an artifact against a dataset, with its configuration pinned at creation. See Evaluations & Metrics.

Evidence — an evaluation framed as support for a claim or decision. A derived object.

Evolution — an artifact's lineage, dependents and blast radius. See Evolution.

Execution Graph — the runtime path of one evaluation run, reconstructed from its own telemetry. See Execution Graph.

Failure Graph — the Execution Graph entered already narrowed to the broken links. The same surface, not a second one.

Forge Graph — the living map of the whole system: artifacts, their real relationships, and optionally the reasoning layered on top. See Forge Graph.

Forge Kernel — the invisible foundation: identity, tenancy, persistence, execution. Layer 1.

Four pillars — the navigation model: Build, Evaluate, Understand, Evolve. Every screen belongs to exactly one.

Ground — the artifacts an evaluation ran against: its agent, prompt and dataset. A precedent must share ground with the failure in question.

Inferred — an epistemic status meaning a causal reading of the evidence that could be wrong.

Investigation — an assembled analysis of a failure: chronology, four causal layers, evidence, precedents, story. Produced by the Root Cause Explorer.

Knowledge — a durable engineering fact that emerged from a decision plus evidence. Never authored. See Knowledge.

Observation — a measured fact: the outcome of an evaluation against an artifact. A derived object.

Precedent — an earlier failure sharing an agent, prompt or dataset with the current one. What makes "has this happened before?" answerable.

Promotion — activating a revision so it becomes the canonical one. A distinct act from creating it, which is why it derives a Decision.

Prompt — versioned instruction text. Version notes become Engineering Memory.

Provider — a model provider and its credentials, encrypted at rest.

Registry — one catalog of every artifact and every derived knowledge object. Layer 2. See Registry.

Revision — one immutable version of an artifact, with its snapshot, rationale, active state and rollback readiness.

Rollback — production running an older revision than the newest. Displayed explicitly, never implied.

Root Cause Explorer — the Engineering Investigation Workspace. See Root Cause Explorer.

Run — one dataset item executed within an evaluation, with its output, latency, tokens, cost, HTTP status and metric results.

Substrate — the structural identity of an artifact type: its icon and its cool-hued colour. Never mixed with the verdict palette.

Suggested — an epistemic status marking a recommendation rather than a finding.

Unknown — both a verdict state (not measured — deliberately distinct from healthy) and an epistemic status (the record cannot answer this).

Unsupported decision — a promotion with no evaluation behind it. A position carried on faith.

Verdict — the evaluative state of anything: healthy, attention, risk, failed, unknown. The product's only evaluative vocabulary.