Don't assume. Verify.
Awareness Verify continuously adversarially tests AI agents and models, compares behavior with known-good releases, and can stop a serious regression before it reaches users.
Break your AI before you ship it.
Initial cohort: up to 10 organizations. One authorized pre-production AI system per pilot.
AI behavior becomes a testable deployment condition.
Ordinary software tests can pass while an AI system changes how it uses tools, handles approval, follows poisoned context, responds under pressure, or behaves after a model or prompt update.
Adversarial testing
Replay security, behavioral, tool, MCP, authorization, and reliability scenarios against the system that is actually being shipped.
Regression evidence
Compare a candidate with an approved baseline or previous run and distinguish introduced, persistent, and resolved findings.
Release gating
Turn reviewed policy into a machine exit code so a serious new failure can stop a deployment while preserving the evidence needed to investigate it.
One real staging system. One useful assurance loop.
The pilot is deliberately narrow. The goal is to prove the workflow on a system your team already owns and understands, not to spend a month in architecture meetings.
Connect safely
Choose one authorized pre-production agent, RAG system, MCP-connected workflow, copilot, or model-backed application. Start without production side effects.
Run the first assessment
Capture JSON and human-readable evidence, review the scenario scope, and identify false positives, blind spots, and the boundaries that matter most.
Establish known-good behavior
Approve or retain a reviewed baseline, then make one controlled change that should violate a tested boundary.
Prove the gate
Verify should detect the regression, block under the reviewed policy, preserve evidence, and return to PASS after remediation.
Founder-led integration, not a self-serve handoff.
days
Private-preview access for one staging system and one primary release workflow.
software fee
No software license fee for the initial pilot. Post-pilot Team/Enterprise needs are scoped separately.
onboarding
Direct help selecting the first test surface, configuring the assessment, reviewing the report, and proving the controlled regression cycle.
Teams shipping AI systems that change often enough for regression to matter.
Strong first pilots
AI product teams, internal AI platforms, agents with tools, RAG applications, MCP-connected systems, regulated or security-sensitive AI workflows, and teams already doing manual red-team or evaluation work.
Not the first place to start
Production-only systems with no safe test target, systems you are not authorized to adversarially test, or engagements that require real destructive side effects just to prove the workflow.
PASS means the reviewed tests passed. Nothing more.
Awareness Verify is built to make evidence stronger and uncertainty more visible, not to turn a limited test into a universal safety claim.
Observed behavior first
Prefer wrapped tools, MCP calls, authorization evidence, and traces over asking the model to describe what it did.
Semantic uncertainty is explicit
Optional semantic evaluators can repeat judgments, measure disagreement, route uncertainty to review, and be empirically calibrated on labeled sets.
Capability can escalate controls
Versioned Frontier profiles can require stronger safeguards and evaluation integrity as measured capability and uncertainty approach reviewed thresholds.
Awareness Verify v0.96.0 adds a separately signed offline handoff for the complete evaluation-acceptance quorum: the plan, envelope, every planned decision and key, underlying evidence packages, and retained verification result. The handoff requires all planned approver decisions, preserves rejection vetoes, checks the exact file inventory and bytes, and replays the captured evidence locally. It supports auditor and design-partner review without a hosted dependency.
Validation passed 58 focused tests and 633 full-suite tests with 8 expected PostgreSQL-only skips. All nine CI jobs and the guarded release workflow passed. The normal Release contains nine assets, including wheel, source distribution, and SHA-256 manifest. Tag v0.96.0 matched validated main commit 13e8a67b1a71c426b7fdb25a4235f49f812042fb at publication. See the published v0.96.0 release evidence in the private repository.
This authenticates supplied evidence continuity, not signer identity, authority, independence, semantic correctness, complete disclosure, trusted time, durable retention, model safety, or deployment fitness. The repository remains private pending legal/IP/trademark/contributor-rights, reachable-history, third-party/bundled-content licensing, and public-repository control reviews. Prepared protection templates and a skipped private CodeQL job are not active controls or completed scanning.
Useful locally. Stronger when teams need coordination.
Community direction
The Apache-2.0 Community candidate retains the tested local and self-hosted core: CLI and SDK, built-in and custom scenarios, calibrated semantic evaluation, regression and CI gates, reports and evidence, Frontier evaluation-integrity replay, SQLite and PostgreSQL, and self-hosted Team Web with RBAC/OIDC and audit evidence. V0.96.0 adds complete acceptance-quorum handoff and a concrete open-core launch runbook. Community is intended to permit commercial use, modification, and redistribution under Apache-2.0. The repository remains private until the four documented reviews and controls are completed.
Team / Enterprise direction
Separate commercial products and services are reserved outside the Community repository: managed multi-organization operations, coordinated hosted/private runner fleets, additional enterprise capabilities, maintained Awareness Intel packs, customer integrations, deployment support, onboarding, and SLAs. These are commercial directions, not claims that every planned feature is implemented. Existing Community self-hosting, PostgreSQL, Team Web, RBAC/OIDC, and audit functionality stays within the documented Apache-2.0 boundary.
Put one real AI release through the loop.
Tell us what the system does, whether a staging target exists, what stack it uses, and the behavior you most want to stop from regressing.