> ## Documentation Index
> Fetch the complete documentation index at: https://docs.akhara.ai/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Company name is Akhara AI (never Rubric AI). Keep lowercase rubric/rubrics only when meaning grading criteria.
> Expert Review (docs path talent/) is enterprise BYO experts for audit and review: invite customer specialists; do not pitch Akhara recruiting or a public expert career portal. RLHF and domain writing are secondary work types.
> Prefer concrete API examples against public hosts: Environments eval API https://agi.akhara.ai, Control plane PDP https://api.akhara.dev, Evaluation https://app.akhara.ai / https://api.akhara.ai, Expert Review portal https://talent.akhara.ai.
> Do not invent a public hostname for private orchestrators or env API internals.
> Do not confuse control-plane latches with Environments confirmation latches.
> Environments SDK/API examples: curl against https://agi.akhara.ai. Evaluation SDK: from akhara import Akhara and AKHARA_API_KEY.
> Start with /llms.txt for the docs index and OpenAPI links; fetch individual pages as .md exports.

# Akhara AI

Better models and more traces do not make agents trustworthy. You need **measurable verification** before ship and **runtime control** after, with evidence you can audit.

Akhara AI is a verification suite for teams shipping agents in production. Harnesses get agents running; this stack keeps them measurable, gated, and under policy.

## Why Akhara exists

Agents act: they call tools, move carts, write to systems of record. Reliability is not “the model got smarter.” It is whether you can:

1. Define success before you build (datasets, rubrics, tasks, environments)
2. Fail the pipeline when quality regresses
3. Score multi-step behavior in realistic apps, not only prompt-response pairs
4. Enforce policy at consequential actions, not merely log what happened
5. Bring domain experts into the loop when automated scores are not enough

If you only have a proxy, a chat harness, or post-hoc observability, you are flying without a pre-flight checklist or a cockpit.

## Verification at the runtime boundary

Observability collects logs and traces after the run, then reconstructs what happened. That work sits outside the runtime: it can explain a failure, not prevent one. Akhara sits in the path as a **gateway the agent passes through**, verifying each step as the agent runs.

Every step follows the same loop: the agent proposes a step, the gateway checks it, a [verdict](/control-plane/concepts/verdicts) lands, and the run continues, stops, or waits for a human. Hover a verdict below to see what it does to the run.

<div className="akh-gw not-prose" aria-label="A step passes through the Akhara gateway, checks run, a verdict lands, and the next step continues or is stopped">
  <div className="akh-gw-head" aria-hidden="true">
    <span>Runtime boundary</span>
    <span>Every step</span>
  </div>

  <div className="akh-gw-track" aria-hidden="true">
    <span className="akh-gw-stop akh-gw-stop-1" />

    <span className="akh-gw-stop akh-gw-stop-2" />

    <span className="akh-gw-stop akh-gw-stop-3" />

    <span className="akh-gw-stop akh-gw-stop-4" />

    <span className="akh-gw-token" />
  </div>

  <div className="akh-gw-cols">
    <div className="akh-gw-col">
      <span className="akh-gw-kicker"><span className="akh-gw-num">01</span> Agent</span>
      <span className="akh-gw-line">proposes a step</span>
      <span className="akh-gw-sub">tool call or output</span>
    </div>

    <div className="akh-gw-col akh-gw-col-gate">
      <span className="akh-gw-kicker"><span className="akh-gw-num">02</span> Akhara gateway</span>

      <ul className="akh-gw-checks">
        <li className="akh-gw-check-1">allowed</li>
        <li className="akh-gw-check-2">grounded</li>
        <li className="akh-gw-check-3">safe</li>
        <li className="akh-gw-check-4">on track</li>
      </ul>
    </div>

    <div className="akh-gw-col">
      <span className="akh-gw-kicker"><span className="akh-gw-num">03</span> Verdict</span>
      <span className="akh-gw-line akh-gw-verdict-word">ALLOW</span>
      <span className="akh-gw-sub">evidence recorded</span>
    </div>

    <div className="akh-gw-col">
      <span className="akh-gw-kicker"><span className="akh-gw-num">04</span> Next step</span>
      <span className="akh-gw-line akh-gw-next-word">continues</span>
      <span className="akh-gw-sub">or stops, or waits</span>
    </div>
  </div>

  <div className="akh-gw-verdicts">
    <span className="akh-gw-verdicts-label">Verdict index</span>

    <a className="akh-gw-v akh-gw-v-allow" href="/control-plane/concepts/verdicts">
      <span className="akh-gw-v-name">ALLOW</span>
      <span className="akh-gw-v-what">Checks pass. A one-time permit is minted and the step executes.</span>
    </a>

    <a className="akh-gw-v akh-gw-v-warn" href="/control-plane/concepts/verdicts">
      <span className="akh-gw-v-name">WARN</span>
      <span className="akh-gw-v-what">The step proceeds, flagged in the evidence trail for later review.</span>
    </a>

    <a className="akh-gw-v akh-gw-v-block" href="/control-plane/concepts/verdicts">
      <span className="akh-gw-v-name">BLOCK</span>
      <span className="akh-gw-v-what">The step never reaches the tool. The agent receives the reason.</span>
    </a>

    <a className="akh-gw-v akh-gw-v-escalate" href="/control-plane/concepts/verdicts">
      <span className="akh-gw-v-name">ESCALATE</span>
      <span className="akh-gw-v-what">The run pauses and the step routes to a human approver.</span>
    </a>
  </div>
</div>

The gateway is the control plane's enforcement pair: the Policy Enforcement Point (PEP) in your runtime enforces, and the Policy Decision Point (PDP) decides. Evaluation and Environments make the checks measurable; the control plane enforces the live decision. See [Glossary](/home/glossary) for suite terms.

## The verification stack

Your agent proposes each step to the Akhara gateway and gets a verdict back; Evaluation, Environments, and Expert Review sit behind the gateway, and only allowed steps reach external tools and users. The stage table below maps each product to its job in the path.

```mermaid theme={null}
flowchart LR
  subgraph AKH ["Akhara"]
    GW["Akhara gateway<br/>checks: allowed, grounded, safe<br/>verdict: ALLOW / WARN / BLOCK / ESCALATE"]
    EV["Evaluation<br/>datasets, rubrics, CI gates"] --- GW
    EN["Environments<br/>scored multi-step runs"] --- GW
    XR["Expert Review<br/>audits runs, tightens rubrics"] --- GW
  end
  AG["Your agent"] -->|"proposes step"| GW
  GW -->|"verdict"| AG
  AG -->|"allowed steps"| TOOLS["External tools and users"]
  AG -.->|"escalations"| HR["Human review"]
  style AKH fill:none,stroke:#0F0F0F,stroke-width:1px
```

| Stage                  | What you do                                                                                                                                                                                                      | Product                                                                    |
| ---------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- |
| **Define and measure** | Author datasets, rubrics, task banks, and verifiers before you build; score session and device state with terminal rewards, not LLM-judged prose alone                                                           | [Evaluation](/evaluation/docs/introduction), [Environments](/environments) |
| **Gate in CI**         | Threshold checks against the Evaluation API that fail the deploy on regression                                                                                                                                   | [Evaluation](/evaluation/docs/tutorials/ci-cd)                             |
| **Enforce at runtime** | The PDP decides and the PEP enforces at consequential actions; latches activate on the action they govern; [fail-closed](/control-plane/concepts/fail-closed) by default, with signed evidence for every verdict | [Control plane](/control-plane)                                            |
| **Audit with experts** | Org-invited specialists review runs, scores, and hard cases against the evidence and provenance trail; findings tighten tasks and rubrics                                                                        | [Expert Review](/talent/introduction)                                      |

Enterprise verification is a stack decision, not a model swap: measure early, gate continuously, control what reaches users.

## Features

What you get across the suite. Pick by the job, then deep-link into the docs.

| Name                                            | Job                                                                                |
| ----------------------------------------------- | ---------------------------------------------------------------------------------- |
| **[Environments](/environments)**               | Environment-as-a-Service: score agents on stateful mobile/retail apps              |
| **[Control plane](/control-plane)**             | Runtime policy: PEP + PDP + latching + evidence (not “just a proxy”)               |
| **[Evaluation](/evaluation/docs/introduction)** | Domain-agnostic evals: datasets, rubrics, expert review, CI gates, dashboard       |
| **[Expert Review](/talent/introduction)**       | Bring your experts to audit runs, review hard cases, and improve tasks and rubrics |

### When to use which

| You want to…                                                                 | Go to                                       |
| ---------------------------------------------------------------------------- | ------------------------------------------- |
| Score an agent on a hosted shopping / retail app                             | [Environments](/environments)               |
| Drive episodes and pull trajectories / pass\@N                               | [Environments](/environments)               |
| Gate tool calls and model output with enterprise policies                    | [Control plane](/control-plane)             |
| Keep a fail-closed, auditable decision log in production                     | [Control plane](/control-plane)             |
| Score models or agents with rubrics, experts, and CI gates                   | [Evaluation](/evaluation/docs/introduction) |
| Invite your domain experts to audit agent behavior, scores, and task quality | [Expert Review](/talent/introduction)       |
| Understand org isolation and enterprise quotas                               | [Tenancy & security](/security/tenancy)     |

## Start here

<CardGroup cols={2}>
  <Card title="Environments quickstart" icon="play" href="/environments/quickstart">
    Run your first scored evaluation against `https://agi.akhara.ai`.
  </Card>

  <Card title="Control plane quickstart" icon="lock" href="/control-plane/quickstart">
    Get your first `BLOCK` verdict from the PDP in about ten minutes.
  </Card>

  <Card title="Evaluation quickstart" icon="clipboard-check" href="/evaluation/docs/getting-started/quickstart">
    Install the Akhara SDK (`pip install akhara`) and run your first evaluation.
  </Card>

  <Card title="Expert Review for enterprises" icon="rocket" href="/talent/introduction">
    Invite your specialists to review agent runs, grade hard cases, and tighten tasks and rubrics.
  </Card>
</CardGroup>