Skip to content
emergent engineering

QTI 3 ASI as the canonical assessment model

Status: accepted

We adopt QTI 3.0.1 ASI (the normalized JavaScript form from @conform-ed/contracts) as the canonical internal representation for assessment items and tests — a superset: QTI core plus namespaced Native Extensions. Anything expressible in QTI round-trips losslessly to/from XML; extensions (e.g. pronunciation scoring, spaced repetition) are emergent-only and explicitly outside the interop guarantee.

This reverses the earlier native-first direction (packages/domain/native-assessment.ts and the qti.ts/qti-adapter.ts envelope), which a future reader would otherwise assume is current. The assessment_definition.authoring_standard enum (native | qti3) lets legacy rows coexist; no big-bang migration.

Why

Section titled “Why”
  • Interop and standards-alignment are core goals (offline/in-person teaching support and a research programme), not afterthoughts. A superset model gives lossless QTI interchange by construction, deleting the native↔QTI mapping layer where lossiness and drift would otherwise live.
  • We already own the model: conform-ed’s v3_0_1 contracts are MIT and deliberately a normalized JS representation, so the GPLv3 web-component libraries (e.g. Citolab) are irrelevant — we build a closed renderer on our own MIT contracts.
  • The research stack (Results Reporting, Usage Data / Item Statistics, CAT) is already contracted in conform-ed and keys off QTI identities.

Considered and rejected

Section titled “Considered and rejected”
  • Native model + thin QTI export adapter. Smaller now, but two models plus a bidirectional map to keep correct forever; defers, not avoids, the work.
  • Full-spec rendering as a launch gate. Conflates “QTI as the model” (a clean commitment) with “render the entire spec first” (a scope trap with no forcing function). We take the model commitment; rendering accrues incrementally per interaction, driven by language-learning need.

Consequences

Section titled “Consequences”
  • Embedded HTML in item bodies is currently modeled as a generic node, so schema validation does not sanitize. The renderer must apply an allowlist tree-walk (never dangerouslySetInnerHTML); contracts are tightened later to subsume it.
  • Superset discipline requires a CI round-trip gate: every core-only item must survive import→export with no loss, tested against the conform-ed/tmp/qti-examples corpora.