Home › Use Cases › Check each data store's sensitivity label against what it holds › What the AI sees

UC0524 · store-sensitivity · the deployment question, answered with real sample data

The model never needs to know the number.
Only that there is one.

The objection is right: no bank hands customer identifiers to a model. But this job does not need the value — it needs to know a value of that type is present. That difference is what makes the whole thing deployable. Below is one real store from this kit's corpus, its nine sampled documents, and exactly what leaves the building under each of the three ways you can run it.

The store, and the disagreement

DS-10000 — an object storage bucket owned by Claims Operations. Everything below is read from the kit's own corpus; every name and number in it is fabricated.

Register saysCONFIDENTIALrecorded 2024-09-16 by the governance desk
Evidence saysRESTRICTEDC-1 and C-2 are both present
VerdictUNDER-LABELLEDthe dangerous direction
Sample9 / 2,5360.35% — a claim about the sample, never the store
Two years and eight days between the label being written and this sample being taken. Nothing re-derived it in between, and nothing was ever going to: every system that reads this label is contractually required to trust it.

Where the masking happens

The order matters. Masking is deterministic code and it runs before anything is sent — not a model deciding what to hide.

STEP 1
Sample the store
9 of 2,536 objects. Runs inside your perimeter.
STEP 2 · PURE CODE
Replace values with type tokens
Pattern match, substitute, never transmit. No model involved, $0.00, same result every run.
STEP 3 · THE ONLY MODEL CALL
Judge each category
Receives the masked text. Answers holds / does not hold, and names its evidence.
STEP 4 · PURE CODE
Derive tier, compare to label
Highest category present wins. Contradiction by subtraction. The model never sees the label.
Nothing the model returns is trusted with a number. It returns a verdict and a document id. The tier and the contradiction are computed afterwards, in code, from those verdicts — which is why a masked run and an unmasked run produce the same final answer whenever the verdicts agree.

Pick how you run it

Then read the nine documents below. The right-hand pane changes to show exactly what leaves your network.

The nine sampled documents

Every document above is real text from data/corpus/SS-0001.txt in this kit. The corpus is synthetic and generated from a fixed seed: every name, identifier, card number, address and email in it is fabricated and belongs to no real person or company. The schema DS-CLASS-2026 is invented for this kit and is not a regulator's rule.

The measured figures quoted on the lane cards — 41 of 64 and 62 of 64 contradictions, 60.5% and 96.2% precision — are from the kit's own run records over all 64 stores, not from this one store.