Diff: semantic and type-aware

Why a normal diff isn’t enough

Run diff on two versions of an act and you get added and removed lines. A lawyer, a journalist, or a downstream system does not care that line 412 moved. They care what kind of change it was. Did the substantive text of an article change? Did the entry into force shift? Or was it just a typo fix, a renumbering, a formatting cleanup?

Those three things have nothing in common except that the bytes are different. A line-based diff throws them all in the same pile. Ours doesn’t.

Type-aware change classification

Because the differ works on Akoma Ntoso structure instead of raw text, it knows what each node is. So it classifies a change by the element it touches and the edit it represents:

• Text amendment. The normative content of a provision changed.
• Temporal change. A date in the lifecycle moved: enactment, entry into force, repeal.
• Structural change. An article was inserted, moved, or renumbered.
• Citation change. A reference was added, removed, or re-targeted.
• Cosmetic. No normative effect at all.

The output is a changeset: a machine-readable account of what kind of change happened between two states, addressed by OLF identifier right down to the element.

The changeset model

A changeset is a first-class object, not a rendering you read once and throw away. Store it, query it, reason over it. “Every substantive amendment to this act since 2019.” “Flag temporal changes, ignore cosmetic ones.” And because it speaks in OLF identifiers, a changeset in Italy has the same shape as a changeset in France.

This is what turns the archive from a snapshot into a history you can actually interrogate.

Status

Early. The classifier and the changeset model are being built next to the first two jurisdictions in the pipeline, so the differ runs against real legislative change from the start, not toy examples.