Documentation

What Heliast is

Heliast is the layer between your people and the AI models you allow them to use. This page says what it does, what it is made of, and what it leaves to you.

What it does

Heliast runs in your own environment. Every request to a model passes through it, and it answers four questions before a model is used and one after.

  • Is this model allowed? Admission checks a public model against rules your organisation can read: its licence, where the weights came from, whether the files match what the publisher released. The answer is a verdict: admitted, held or rejected.
  • Is this person allowed this model? Your own policy says which groups may use which models, and within what limits.
  • Is it any good for our work? Compare runs your own tasks across models and scores them the way you decide.
  • What happened? Every request leaves a row in a tamper-evident ledger. Rows carry who, which model, the decision, the token counts and a hash of the prompt. They never carry the prompt or the answer.
  • What can we show? The Compliance section collects evidence about your estate for your own assessments.

The parts you will see

PartIn the consoleWhat it is for
AdmissionAdmissionThe list of model verdicts, and a page for each model.
CompareCompareYour tasks, your scorers, several models.
LedgerLedgerThe audit query, policy denials, and a check that the chain is intact.
EntitlementEntitlementVirtual keys, budgets and consumption.
ComplianceComplianceThe data-flow document, the AI register, the supply chain register, framework mappings, the live-release check and the signed evidence pack.

What Heliast does not do

  • It does not host model weights. They come from the model's publisher. Your operator downloads them and Heliast verifies the files against what the gate admitted. See using an admitted model.
  • It does not read your prompts or answers. The ledger holds hashes and counts.
  • It does not make compliance claims. Evidence in the Compliance section is a record of what is declared and what was observed. Whether your estate meets a standard is your assessment, and ours does not replace it.
  • It does not say a model is good. Admission says a model is allowed. Quality on your tasks is what Compare measures.

Where to start

  1. Read reading a verdict to understand what the console tells you about a model.
  2. Check hardware to see what your estate needs.
  3. Follow deploying to put it in place.
  4. Set policy and entitlement so people can use what you have admitted.

Limits

Heliast is early. Most of its parts have been exercised in tests and on a single development machine, and none has run on production-sized hardware. The pages say so where it matters, for example in hardware.