These directions are deliberately not described as current features. Each needs evidence, a clear boundary, and a reopening condition before it becomes part of the product.
| Direction | Current pain | Evidence or prerequisite | Why it is not current | Reopen when |
|---|---|---|---|---|
| Learned ETA or speed forecasts | Deterministic baselines cannot capture every traffic pattern | The forecast workspace must beat the baseline on a temporal holdout and pass the release evaluation | The current production path uses derived observations and deterministic statistics | A model shows a measured improvement without making failure handling opaque |
| Human-reviewed schematic transit map | Geographic route lines are hard to compare across a dense network | The schematic editor must preserve route identity, direction, stop order, and shared corridors through validation | Automatic octilinear layout produced plausible shapes that damaged transit meaning | A human-in-the-loop editor can export an approved artifact with topology checks |
| Loom or WASM schematic runtime | Interactive editing may need a smaller, faster geometry engine | The external Loom project and an approved artifact contract would need an integration boundary | Loom is not integrated into the current geographic MapLibre map | A production screen needs schematic editing or rendering and the contract is tested |
| Community route corrections | Jeepney data changes faster than a static review cycle | A provenance, review, rollback, and publication path is required | Unreviewed community geometry would become transit truth too easily | Corrections can be attributed, validated, and reversed |
| Reviewed transfer registry | Proximity rules alone do not express every real boarding point | The registry must be a real compiler input with identity and tests | The current compiler uses official bus-stop identity plus a 400-metre bound; there is no separate registry input | A transfer needs evidence that the current compiler cannot represent |
| Better place resolution | Street and landmark language is ambiguous | Provider comparisons and multilingual evaluation need current samples | A geocoder resolves coordinates but must not alter topology | Place failures remain measurable after current fallbacks are exhausted |
The graph-valid route postmortem, schematic editor experiment, deterministic-baseline ADR, and AI evaluation pages contain the evidence behind these boundaries.