Controls for teams that train, tune and supply data to models.
Governance of training data, retrieval and memory — and a briefing for the people who use the system.
The data you give a model is part of the control environment.
Context is the practice for teams that train, fine-tune or retrieve into large language models. The usual failures are not missing patches. They are over-broad corpora, memory that retains the wrong thing, shared credentials on a tool, and staff who cannot tell a model instruction from ordinary text.
Work can be a written review of how data is selected and retained, a design for retrieval and memory controls, or a briefing for staff and executives. It sits beside AI assurance when the question is the running system, and beside training when the question is the people.
Client material provided for a review is never used to train or tune any model. That is a term of the engagement.
Training and fine-tuning data
What is in the corpus, who approved it, whether personal or privileged material can be reconstructed, and who can add to it later.
Retrieval, context and memory
What is pulled into a prompt, what is stored as memory, who can see transcripts, and whether sensitive material can leave the intended boundary.
Staff and executive briefing
What the system is allowed to do, what to escalate, and how an instruction hidden in a file or an email differs from ordinary content.
Written practices, then the people.
Most clients start with a short review of the data path, then a briefing. We do not treat a vendor questionnaire as proof that your retrieval store is bounded. We look at the deployment you actually run.
The same terms apply internationally: written scope, named analyst, remote by default.
What we will not claim
- That a model is certified safe
- That staff training replaces a technical review
- That your material will be used to improve our systems
Request a written proposal.
Describe the system, vendor or AI deployment under consideration. We respond with the proposed scope, timeline and fee.