Topic 15 of 20 · Part IV : Data and Learning Operations

Privacy

The first and strongest privacy knob is collection: data that is not necessary should not be gathered.

The control pattern

Data asset

Sensitive user and portfolio information

Knobs

Collection, Consent, Redaction, Retention, Access, Deletion and auditability

Outcome

Useful personalisation with lower privacy risk

Measurement

See how to measure this below

What it is

Privacy treats sensitive information as a controlled asset rather than ordinary model context. The first and strongest knob is collection: data that is not necessary should not be gathered.

Additional knobs govern consent, redaction, retention, access, aggregation, deletion, and whether content may be used for product improvement. These controls should apply across logs, analytics, model prompts, caches, vector stores, exports and support workflows : deleting a visible profile while leaving derived copies elsewhere does not honour the user's intent.

Privacy is not simply a disclaimer. It is an architectural knob that determines what the system is capable of remembering and reusing.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
CollectionWhat is gathered at allThe strongest control available
ConsentWhat the user agreed toGoverns every downstream use
RedactionWhat is stripped before storageLimits exposure in logs and prompts
RetentionHow long data survivesBounds accumulated risk
AccessWho may see it, and whySeparates identity from analysis
Deletion and auditabilityWhat removal actually reachesMakes the promise verifiable

Applied: Stocks Assistant

A Stocks Assistant portfolio may reveal wealth, employer equity, trading behaviour, risk tolerance and financial goals. Separate identity data from analytical data where possible, restrict access by role and purpose, and avoid exposing one user's holdings in another user's retrieval context. Personal data should not enter a general distillation dataset by default. If examples are needed for training, use properly de-identified, aggregated or synthetic representations, and retain provenance showing what transformation occurred.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Access-policy violations
  • Deletion completion across every store
  • Retention compliance
  • Sensitive-data leakage in logs
  • Percentage of features operating with minimised data

Common mistakes

Deleting the visible profile while derived copies survive in caches, vector stores and exports.

Letting personal data flow into a general training set by default.