12,000 simulated episodes • 9 fault types • 0 unsafe commits under the tested conditions
Context
WWAIOS™ is WhiteWater.ai's governed decision operating system. Its agents read records, decide, and write back to systems like the CRM. The goal is simple: an AI acts only on information that is still true when it acts.
What happened
The review, recorded on September 17, 2026, traced one write path. The system confirmed that a user still held access to a workspace when the request came in. The decision record was committed later, in its own database transaction, which never re-read that access. If access was revoked in between, the write would still go through with a clean check on file. Moving the recheck earlier in the request would not close the gap.
Why it happened
Verification was treated as a step, not a condition. The check proved the context was good at check time. It proved nothing about commit time.
What we changed
Every gated action now carries a contract with three terms: how fresh each piece of information must be, where it must come from, and how confident the system must be. The terms are checked at commit. A check-to-commit rule pins the versions seen at check time to the write, so the target refuses a write if anything moved. Every evaluation, permitted or rejected, writes a hash-chained decision record.
Results (simulated)
In a seeded, 12,000-episode fault-injection simulation across nine fault types, CVDC with target-enforced conditional commit recorded zero unsafe commits under the tested conditions, and rejected no valid actions.
The same run reproduced a limitation: re-reading context just before the write still missed changes that landed after that last read and before the action committed. Closing that window takes a target that refuses the write when the checked version has moved.
Baseline approaches, such as asking the model to check its own context, missed most injected faults; their detection rates were modeled assumptions, not measured benchmarks of any AI product.
CVDC runs in shadow mode on WWAIOS™ today. It is not enforced on live actions, and there are no customer results to report.
Lessons for others
- A check that happened before the action is not a check at the action. Bind them.
- A model judging its own context cannot see what changed after it answered. The check belongs outside the model.
- Write the decision's terms down before the agent acts, by risk tier.
- Record the denials too. A log of only what happened hides what was stopped.
