Live classifier
One field
The engine the paper evaluated, ported line for line and drift-tested against every corpus row. Enter a field name — a description helps — and get everything Algorithm 1 emits: a candidate level or an abstention, the four-dimension impact profile, the policy citation, the PDPL flags on their own axis, the gate decision with its trigger, and the full match trace showing how conflicting rules were resolved.
A whole schema
Reading the output
Candidate level means a policy-cited rule fired and derived a level from its impact profile. Unclassified — interim Restricted means nothing fired: the field is not labelled, it carries NDMO §1.4's protective default until a Business Data Executive completes an impact assessment, and it is always gated.
Gated means a human must confirm or override. Four of Algorithm 1's five triggers can fire here: a Top Secret candidate, the winning rule's hand-assigned assurance below 0.75, the §1.4 default, or a name-based PDPL candidate. The fifth — an AI-detected category — cannot, because the optional detector is absent from this port, as it was switched off in every experiment the paper reports. The paper's sensitivity sweep shows silent under-classification stays at zero for every cut from 0.60 to 1.00 — the assurance values tune reviewer workload, not safety.
The PDPL axis is a union over every rule that fired plus a conservative name-based detector,
so credit_score_balance keeps its Credit Data candidate even though the lower-level balance
rule wins the security level. Flags annotate; they never re-level.
Reference labels in the paper are author-assigned; treat every output here as a candidate pending human confirmation, not as legal advice.