Without an assurance boundary
“The build failed because of a stale lockfile. Tests pass now, so deployment is safe.”
The response merges an inference, an observation, a judgment, an authorization, and an application decision.
Itself / Open source
Itself
Knowledge itself is the question.
An open-source assurance SDK for AI agents.
Make model assertions testable before software acts on them.
The problem
An agent can sound certain before anything has been checked.
AI agents can produce fluent diagnoses, recommendations, and explanations without knowing whether they are correct (Lee et al., 2026, Appendix A.2).
The Itself SDK records the assertion, external check, observed result, evaluation, and decision as separate protocol records.
With Itself
- Inferred
- Stale lockfile hypothesis
- Model
- Observed
- 214 tests passed after regeneration
- CI worker
- Judged
- Behavioral recovery observed
- Evaluator
- Authorized
- Redeploy approved
- Release owner
- Decided
- Pipeline resumed
- Application
Itself records each item separately with scope, provenance, and a responsible actor.
How it works
The protocol records each step before an application acts.
The agent states what it believes and what it expects a check to show.
The Claim and Evidence Protocol records that formal lifecycle.
-
01 / Capture
Record the assertion.
The application records a scoped model assertion as a claim or hypothesis. Its epistemic state starts as unverified.
-
02 / Declare
State what would distinguish it.
The application records predictions and a test plan before a tool or reviewer performs the test.
-
03 / Test
Run an external check.
A sandbox, CI worker, service, calculation, source check, or person performs the check outside the model. Its observed result becomes evidence.
-
04 / Judge
Apply declared authority.
An evaluator relates evidence to the assertion. A human, organization, or software actor that applies a policy authorizes an evidence-backed status transition.
-
05 / Inspect
Replay the record.
The Itself SDK replays the ledger and generates a minimized reasoning receipt for review or another application.
Protocol rule
A model can create an assertion or recommend a test. It cannot authorize an evidence-backed transition for its own assertion.
Initial application
Add assurance to AI-driven engineering workflows.
Use Itself when an agent can influence code, CI, a release, or incident response.
The same protocol supports research, analysis, and other consequential workflows.
Gate an agent-generated merge request.
A coding agent asserts that its patch fixes an issue. CI and review policy determine whether the change may merge.
Itself preserves the assertion, check results, evaluation, and merge decision.
Check an AI review finding.
An AI reviewer identifies a possible vulnerability or regression. A tool or engineer checks the finding before it can block a merge.
Itself records whether the finding remained provisional, was supported, or was dismissed.
Diagnose and repair a failed pipeline.
An agent proposes a cause and a repair. A CI worker or sandbox runs the declared check outside the model.
Itself prevents another confident explanation from substituting for the observed result.
Test a proposed root cause.
An operations agent connects a failure to a deployment or configuration change. Telemetry, replay, rollback, or human review tests the assertion.
Itself preserves what justified the mitigation or continued investigation.
How developers use it
Add Itself between an agent assertion and a consequential action.
Itself does not replace the agent framework, CI runner, code host, or reviewer.
The Python SDK uses a local ledger. The protocol schemas and conformance fixtures are language-neutral.
- 01 Agent or human assertion
- 02 Itself ledger
- 03 External check or review
- 04 Evaluation and policy
- 05 Application decision
uv sync --extra test
uv run itself validate \
conformance/valid/proposed-claim.json
uv run itself ledger append scratch.jsonl \
conformance/valid/minimal-claim.json
uv run itself ledger replay scratch.jsonl
uv run itself receipt generate scratch.jsonl \
--output receipt.jsonWhat the SDK does
- Validates protocol records and references
- Rejects unsupported state promotion
- Appends records to an ordered local ledger
- Replays the current epistemic state
- Generates a minimized reasoning receipt
Cookbook
Put Itself into an engineering workflow.
Each recipe connects a model assertion to an external check and a controlled action.
Run one in GitHub Actions, GitLab CI, or a local shell.
01
GitHub Actions / GitLab CI
Gate an agent change.
Require repository-owned checks before an agent-generated change may merge.
02
Python / CLI
Verify an evidence bundle.
Check a portable record of the assertion, test, observations, evaluation, and decision.
03
CI artifact
Publish a reasoning receipt.
Generate a compact review artifact from the validated protocol ledger.
Reasoning receipt
Reviewers receive a compact record.
A reasoning receipt is a minimized, deterministic projection of a valid protocol ledger.
It includes public decision records and omits private chain-of-thought and sensitive evidence payloads.
A receipt supports inspection. It does not prove factual truth or authenticate external artifacts.
- Claim
- Release candidate meets the registered test policy.
- Test
- Run 214 registered tests in CI.
- Evidence
- Exit 0 · 214 tests passed
- Authority
- Release policy · release owner
- State
- Corroborated within scope
Assurance boundary
Structural assurance is not factual truth.
The SDK can establish protocol conformance, local reference integrity, deterministic state replay, and an exact match between a receipt and its ledger.
Those properties do not establish that:
- A model assertion is factually correct
- An evidence source is genuine or trustworthy
- A selected test covers every relevant context
- An evaluator, policy, or reviewer is correct
The Claim and Evidence Protocol is pre-1.0. Its formats may change before a stable v1 release.
Open source / Itself
Your agent can assert. Your system must check.
Read the protocol, inspect the reference SDK, and follow a complete assertion-to-decision workflow.