Glossary - Akhara

Short definitions for engineers reading across the Akhara AI suite. Each product also has deeper concept pages; use those when you need mechanics, not just vocabulary. For Evaluation metric definitions, see the AI / ML metrics glossary.

Suite

Term	Definition
Verification layer	What Akhara AI is as a whole: measurable checks before ship, runtime policy after, and evidence you can audit. Not a single model, harness, or chat proxy. See Akhara AI.
Quality gate	A threshold check (usually in CI via Evaluation) that fails the pipeline when scores regress. Design-time and pre-deploy measurability.
Runtime policy	Control-plane enforcement at consequential agent steps (PEP/PDP verdicts, latches, permits). Different job from quality gates: gates score and block deploys; runtime policy blocks or allows live actions. See Control plane.
Runtime verification gateway	Where agent steps pass through verification and a decision (Allow, Warn, Block, Escalate) before the next step continues. Observability reconstructs later; Akhara verifies at the runtime boundary. See Akhara AI.

Control plane

Runtime policy for AI agents. Start with concepts overview.

Term	Definition
PDP	Policy Decision Point. The service that decides (`https://api.akhara.dev`). It evaluates policies and returns a verdict. It never executes your tools. See Overview.
PEP	Policy Enforcement Point. The SDK you embed in the agent runtime. It enforces: for each risky step it calls the PDP and applies the verdict before the side effect. See Overview.
Latch	A control-plane policy that activates only on the consequential action it governs (for example, renew a prescription), not on earlier chat. Do not confuse with Environments confirmation latches. See Latching.
Permit / permitId	One-time authorization id the PDP mints on `ALLOW` for an `action` stage. Downstream services should refuse to execute without a valid permit. See Overview.
Verdict	PDP decision for a step: `ALLOW`, `WARN`, `BLOCK`, or `ESCALATE`. See Verdicts.
ALLOW	Proceed; for `action` stages the PDP may mint a `permitId`.
WARN	Proceed with a recorded warning (policy-dependent handling in your runtime).
BLOCK	Do not execute the step.
ESCALATE	Hold for human or higher-authority review before proceeding.
Fail-closed	If the PDP is unavailable or the decision path is undecided, the PEP does not silently permit. See Fail-closed.
Evidence / signed evidence trail	Sealed decision records you can sign and later verify for audit. See Sign a record and Verify a record.

Environments

Environment-as-a-Service for scoring agents on hosted apps. Start with Environments.

Term	Definition
Environment / env_id	One deployable unit: app + backend + tasks + verifiers, addressed by `env_id`. See Environment.
Eval API	Public evaluation API at `https://agi.akhara.ai` (for example `POST /evaluations/run`). See Evaluations API.
Admin	Admin UI and APIs at `https://admin-agi.akhara.ai`. See Dashboard API.
Orchestrator / env API	Private env protocol (`/v1/env/*`, `/tasks`). Not publicly routed; do not invent a public hostname for it. See Env API.
VU-*	Session success-criteria verifiers (names often prefixed `VU-`). They score whether the episode met task goals. See Verifiers.
Terminal reward	Episode outcome score restricted to `+1`, `0`, or `-1`. See Scores and stats.
Confirmation latch	Sticky `confirmation_violation` when the agent skips a required `needs_confirmation` before an irreversible action. Environments-only; not a control-plane latch. See Confirmation latches.

Evaluation

Domain-agnostic scoring of models and agents. Start with Introduction.

Term	Definition
Rubric	Grading criteria or scale for human or automated review. A domain term, not the company name (company is Akhara AI). See Rubrics.
Dataset	Curated collection of samples used as the unit of eval input. See Datasets.
Sample	One evaluable item inside a dataset (prompt, context, expected output, metadata). See Samples.
Evaluation	A scored run of a model or agent against a dataset (and rubrics / reviewers as configured). See Evaluations.
CI gate	Scripted or hosted threshold check that fails CI when evaluation scores regress. See CI/CD.
Quality gate	Same idea as CI gate in the Evaluation product: measurable pass/fail before promote. Distinct from control-plane runtime policy.

Expert Review

Enterprise-invited expert audit and review: specialists review agent runs, scores, failures, task quality, and policy-sensitive cases. Docs path talent/. Start with Expert Review introduction.

Term	Definition
BYO / enterprise-invited experts	Your organization’s specialists, invited to audit runs, review hard cases, and improve tasks and rubrics. Portal: `talent.akhara.ai`.
Expert audit / review	Domain specialists reviewing trajectories, scores, failures, and task or rubric quality when automated scores are not enough. Primary Expert Review role in the verification stack. Distinct from Evaluation’s human-review scoring features.
RLHF evaluations	Secondary capability: preference, ranking, safety-flag, and rationale work that feeds Evaluation. See RLHF evaluations.
Domain writing for RL environments	Secondary capability: scenario and success-criteria authoring in the expert’s specialty, feeding Environments task banks. See Domain writing.