S02.1 — Pin and implement the BANKING77 loader

Migration identity#

• Canonical identifier: S02.1 (GitHub issue type: subtask).
• Legacy identifier: C02.1 (retained as old_identifier).
• Parent epic: E02.
• Milestone: M1 — IntentGuard Weekend MVP.

Objective#

Resolve an immutable PolyAI/banking77 revision and load it through one offline data boundary.

Rationale#

Dataset identity must be verifiable before any benchmark or split claim is valid.

Parent identifier#

E02.

Source task#

T-002.

Traceability#

Primary: T-002. Secondary: FR-001, AC-001.

Prerequisites#

U01 complete; network or verified cache available; upstream revision and license verified.

Likely files#

configs/default.toml, src/intentguard/data.py, scripts/prepare_data.py, tests/unit/test_data.py.

Implementation boundary#

Load BANKING77 by explicit revision; no substitute corpus or training behavior.

MUST scope#

Revision resolution, loader configuration, offline/cache error handling, and raw schema validation.

Explicit non-goals#

Synthetic fallback as strict data, augmentation, external OOD data, or model tokenization.

Acceptance criteria#

Loader records the exact revision and source, validates required fields, and fails with actionable diagnostics when neither source nor verified cache is usable.

Validation commands#

make data && uv run pytest tests/unit/test_data.py -q.

Expected evidence#

Resolved revision, source/license metadata, successful load summary, and deterministic failure tests.

Fallback and status consequence#

Verified cache is Implemented; synthetic data is Mocked/Partial and fails strict MVP.

Stop condition#

Stop for unresolved revision, incompatible schema, license conflict, or unverified cache.

Definition of ready#

U01 is done and revision verification can be performed without test leakage.

Definition of done#

The pinned loader passes and provenance contains an immutable revision.

Labels#

type:subtask, priority:MUST, area:data

Estimate#

0.5 hour.