AVA has 173 tools. Every one of them can take real action on real data.
Before any of those tools reach a customer environment, they go through a structured certification process — one that verifies not just that a tool works, but that it works correctly, refuses to work incorrectly, and cannot be directed to do things it should not do.
What we are testing and why
Software that can read, restore, and modify organizational data across many platforms has to be right — not mostly right, not right under normal conditions, but right across the full range of things a real customer might ask of it.
Our certification is designed to find the edges: the questions that are asked ambiguously, the operations that could affect the wrong record, the confirmation prompts that might be skipped, the data that might leak across organizational boundaries.
We find those edges before you encounter them.
Isolated test environments — no real customer data
All certification runs happen in dedicated test environments that are structurally identical to production but isolated from all customer accounts. Test data is purpose-built for each scenario — it never includes real customer records, real customer credentials, or real customer platform connections.
The environments are configured to mirror the conditions a customer would encounter: representative data volumes, representative platform configurations, realistic field structures. But the data is ours, created for testing, and discarded when the test is complete.
Real customer accounts cannot be accessed, queried, or modified during a certification run. This is not a policy — it is enforced at the system level. A certification run has no pathway to a live customer's data.
What certification covers
Each certification cycle exercises every AVA tool in a structured sequence:
Identity and discovery — Does AVA correctly identify which organization she is operating in? Does she correctly scope all tool calls to that organization and only that organization?
Data reads — Do queries return accurate results? Do aggregations produce correct counts? Does AVA correctly represent what she does and does not know?
Creates and modifications — Do write operations create what they should, where they should, and nothing else? Do field values populate correctly? Do relationships get created correctly?
Confirmation gates — Does AVA surface the correct confirmation prompt before every write action? Does she refuse to proceed without explicit confirmation? Can she be talked out of the confirmation step? (She cannot.)
Destructive actions — Does every delete, remove, and deactivate operation require a two-step confirmation? Does a single instruction to "delete everything" get refused and escalated rather than executed?
Organizational isolation — Can a session operating in one organizational context access data from a different organization? Can it be instructed to do so? (It cannot, and the attempt is logged.)
Self-reporting — Does AVA accurately describe her own capabilities, limitations, and what she does not know? Does she give honest answers when asked about features that are not yet available?
What a certification score means
After each certification cycle, every tool receives an outcome: passed, partially passed, or failed. Partially passed and failed tools go back to engineering. They do not go to production until they pass.
This is not a one-time gate. Certification runs happen continuously — when a tool changes, when the underlying model changes, when platform APIs change, and on a regular schedule regardless. A tool that passed six months ago gets re-certified.
The certification history for each tool is maintained and used to identify regression patterns — cases where a tool that was previously working starts failing in a new way.
What this means for you
When AVA takes an action in your account, that action has been exercised against realistic conditions in an isolated environment, verified to produce the correct outcome, and confirmed to refuse the wrong instructions before reaching you.
You are not beta testing AVA's basic operation. By the time a tool is available in your account, its core behaviors have already been rigorously verified.
This does not mean AVA is infallible. Edge cases exist. Novel combinations of requests can produce unexpected results. If you encounter something AVA does that seems wrong, report it — the feedback directly informs the next certification cycle.
Related: How AVA Confirms Before Acting — Action Gates · Introducing AVA — Your Organizational Brain