Get started
Core concepts
Understand deployments, sessions, operations, governing revisions, and decisions.
A trajectory is the story of one task: what the agent was asked to do, what it proposed, what happened, and what remains unresolved. Integrity keeps this context attached to the decisions made along the way.
Intent, expectations, and evidence#
The task describes the desired result. A constitution supplies the governing instructions and constraints. A rubric turns selected expectations into criteria that a grader can assess. Evidence is the request history, available outputs, tool results, and reports used to explain a decision.
For the authorization example, the task is to implement the permission check. The constitution requires preserving authorization tests. A rubric can separately assess whether the implementation solves the task, retains meaningful coverage, and avoids unnecessary changes. A failing test is evidence to investigate, not permission to remove the constraint.
The units of work#
| Term | Meaning |
|---|---|
| Workspace and project | The organization and project scope for people, deployments, and evidence. |
| Deployment | A configured Integrity connection with credentials, capabilities, governing configuration, and budgets. |
| Session / trajectory | One task and its accumulated evidence, identified by the application’s persisted session ID. |
| Operation | One distinct model or control request, identified before dispatch by an idempotency key. |
| Governing revision | The constitution and configuration recorded for the session or an authorized transition. |
| Integrity version | The reviewer or adaptation artifact used by an available connection; separate from your target model. |
Use opaque, nonsensitive session and operation IDs. They are routing metadata, so they must not contain a person's email, a credential, or task content.
What a decision means#
Released content passed the required checks for that operation. A hold withholds release and preserves unresolved work. A correction asks the same customer model connection for a revised continuation within the supported budgets, then checks the result before release.
These decisions do not establish that a tool ran or that the entire task succeeded. Your application reports effects separately. A supported executor permit authorizes an exact action under its session, governing revision, and expiry; the executor must enforce that boundary.
Feedback and adaptation#
Feedback explains an expected decision and why it is appropriate. Grading adds an assessment against a pinned rubric. Together they help you refine a constitution and prepare representative evaluation or training data. They do not automatically edit the live contract or change model weights.
Learn the workflow in Feedback and constitutions, then follow grading, synthetic generation, and adaptation.
Coverage is explicit#
The current Base connection reviews proposed actions and final outputs. Inputs and tool results can supply context but do not receive a separate pre-inference review. Integrity cannot expose hidden model reasoning. Missing evidence, unavailable review, and unknown usage remain explicit.
Read the session and executor contract when implementing your harness, and the capability reference when selecting a workflow.