Training on production traffic is subject to the obligations that already apply to that traffic. Personal-data handling, residency, retention and the authorized purpose all carry from serving the traffic to training on it.
The records the training reads are the records Akka Verify already governs. Sanitizers redact before an event is written. Retention categories carry their own floors, and the interaction log shows what the training used.
Akka saves every interaction into tamper-evident storage, indexed, queryable and cryptographically chained. The same record feeds replay testing, conformance reconstruction, drift detection, audit evidence and the fine-tuning corpus.
One record under one set of controls keeps the question answerable. A corpus assembled by copying traffic into a second store is a second place personal data is held, with its own residency rule, retention schedule and access review.
Sanitizers run before the event is written. Redaction, masking and reshaping happen at capture, so the corpus never contains material the policy said it may not hold.
A review of a training programme starts with where the data went. Answering that means tracing an export, the store that received it, the permissions on that store and the schedule that deletes it.
When the training reads the governed record in place, the answer is the classification already on file. Akka Verify classifies the system against 190 AI regulations and 1,230 controls, and a typical system is subject to 30 to 60 of them.
A new specialized model is a declared change. The sign-off recipe engine routes it to the personas the matrix names, with the dossier attached, and the review is scoped to what the change affects.
Sanitizers redact and mask before an event is written, so the material available to training is the material policy permits. What the model can learn is bounded by what the corpus was allowed to hold.
Training runs where the data already is. Full-parity sovereign deployment keeps traffic and data in region, with data isolation, networking isolation and local support.
Retention categories and sanitizer policy decide what the record holds and what the training corpus may draw from, and both are rows in the signed control matrix.
The interaction log is non-sampled and hash-chained, and the grades and training runs are written to it. The records handed to an auditor are the records the runtime emitted.
Synthetic records are generated from the structure of the class of work and graded by the criteria that grade production records. Generation writes them into the same governed record, so the controls covering the corpus cover them.
