Skip to content

Repository files navigation

EvalKit Playground

EvalKit Playground is a browser-based AI evaluation playground for validating rubric JSON, scoring JSONL model responses, and inspecting criterion-level evidence without provisioning an evaluation backend. It is the searchable, interactive entry point for browser evaluation, LLM rubric scoring, EvalKit experiments, and local JSONL evaluation.

For: evaluation engineers, researchers, prompt/application developers, and reviewers who need to understand a rubric or reproduce a small scoring example before integrating a runner or CI workflow.

Differentiator: rubric validation, response import, scoring evidence, and shareable state are visible in one browser workspace. No AuraOne account or API key is required.

Evaluation Workflow

  1. Load the included example or import rubric JSON.
  2. Paste or import JSONL response records.
  3. Validate structure and inspect actionable input errors.
  4. Run the in-browser scoring path.
  5. Review exact criterion evidence instead of only an aggregate score.
  6. Create a permalink when the input is safe to disclose.

Use the hosted browser surface at playground.auraone.ai.

Run From Source

The supported source path is the public GitHub repository:

git clone https://github.com/auraoneai/evalkit-playground.git
cd evalkit-playground
npm ci
npm run dev

Vite prints the local URL. Node.js 20 or newer is the documented toolchain. There is no supported npm-library install for the application package.

Runtime, Data, And Network Boundary

  • Rubric and response inputs are processed in browser memory. The playground has no AuraOne application backend, account session, or server-side project store.
  • Permalinks base64url-encode the full rubric and response state in the URL fragment. A fragment is not encryption. Anyone receiving the link can decode its contents, and browser history, clipboard history, screenshots, or chat logs may retain it.
  • Do not paste real customer data, secrets, regulated records, or confidential evaluation inputs into the hosted playground or a permalink.
  • The primary execution path downloads Pyodide 0.26.4 from jsDelivr and asks micropip to install auraone-evalkit. First use therefore requires those public network resources unless they are already cached.
  • If Pyodide or package loading fails, the UI uses a clearly warned synthetic local fallback so the workspace remains demonstrable. That fallback is not a substitute for EvalKit execution and must not be treated as evaluation proof.
  • Normal product and repository links can navigate away from the app, but input data is not posted to an AuraOne backend by the playground.

Font Boundary

The public source and build contain no private licensed font binary and no remote font import. Named UI and monospace families resolve through installed system fallbacks. An authorized branded host may supply licensed typography only from a host-owned stylesheet on an approved same-origin path; if it is absent or blocked, the public fallback remains the supported rendering. Private font files must not be added to source archives, npm packages, or deployment bundles.

Proof

npx playwright install chromium
npm run test:unit
npm run test:ui
npm run build

The unit suite covers permalink round trips, validation, imports, and workspace scoring state. The Chromium suite covers local evaluation evidence, keyboard navigation, axe accessibility, reduced motion, forced colors, long content, and responsive layouts from 320 px through 1440 px.

Release Truth

Status verified on July 13, 2026:

  • The latest public GitHub release is v0.1.1.
  • The hosted browser surface is publicly reachable.
  • This checkout identifies itself as 0.2.0; it is a source candidate, not a published 0.2.0 release.
  • evalkit-playground is not published on the public npm registry. The supported paths are the hosted app and the GitHub source checkout.

Next Action

Start with the included non-sensitive example, confirm the evidence rows, and then use the source tests when integrating. Release owners should tag and publish 0.2.0 only after the deployed build, source commit, and public release record all resolve to the same verified version.

About

Import rubric JSON and response JSONL, validate them, run deterministic scoring, and inspect exact evidence in the browser.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages