Concepts
How Orcher thinks
Six ideas cover the whole platform. Everything in the API reference and the SDK maps back to one of these.
Directives
A directive is a plain-language instruction for AI agents: 'refund every customer affected by yesterday's outage', 'reconcile the Frankfurt wire'. The original text is stored immutably — it can never be edited, only interpreted.
Each directive lives in exactly one environment and moves through a fixed lifecycle.
Interpretations
An interpretation turns the raw text into structure: the action, the systems of record it touches, the data classes involved and a risk level. Interpretations are versioned — the AI proposes one, a human can correct it, and exactly one version is confirmed and bound to the directive.
Once bound, the risk level is frozen. That risk level drives everything downstream: who can approve, how many approvals are needed, and what the Logic Scrubber checks.
Authority and the two-person rule
The Role Identity Fabric decides who may authorize a directive. Each role (owner, admin, operator, auditor, viewer) has a maximum risk it may approve, per environment. Above a configurable risk threshold, two different people must approve — and the submitter can never approve their own directive.
When the required approvals are in, the platform issues a signed authority token with an expiry. Tool calls must present a valid token; losing the role or production access revokes outstanding tokens automatically.
Environments: sandbox and production
Every organization has two fully separated environments. The sandbox has all components switched on and uses demo fixtures by default, so teams can rehearse safely. Production starts locked: components must be licensed, connectors need real credentials, and approvers need explicit production access.
The promotion flow carries an authorized, verified sandbox directive into production with its interpretation intact — but production authority applies fresh, including the two-person rule.
The audit ledger
Every event — capture, interpretation, approval, verdict, tool call, policy change — is appended to a hash-chained ledger, one chain per organization and environment. Each entry includes the hash of its predecessor, so any tampering breaks the chain and is detected by verification.
Daily root hashes give you a compact anchor for long-term retention, and any directive can be exported as a signed evidence bundle for auditors.
Data controls
The Data Control Gateway classifies data (personal, financial, health, credentials, confidential, public) and applies an allow / redact / block rule per class, per environment. The scan runs on text entering AI models and on payloads leaving through tool calls.
Redacted values are replaced with placeholders like [REDACTED CARD] before they reach a model or an external system — the platform treats them as present but unknown.