Discovery and Access

discovery.run(request, client) is independently callable. A request specifies a question, WGS84 extent, timezone-aware interval, variables, modalities, and optional track. Discovery filters by bounding-box intersection, overlapping intervals, modality, and exact declared variable identifiers. It does not equate similar names. The orchestrator additionally requires every configured variable to be present.

A SourceResource preserves dataset, source, asset/version, URI, checksum, retrieval time, variable definitions/units/method/entity/ontology, spatial and temporal support, modality, sensor/platform metadata, provenance, and rights. Rights distinguish license, attribution, derivative permission, redistribution, and access restrictions. No blanket license is inferred from a source name.

MockMESAClient reads data/fixtures/catalog.json, validates metadata, verifies asset bytes before parsing, and exposes search, describe, fetch, and get_provenance. Fixture paths cannot escape the fixture root. Registration is intentionally unavailable for the immutable fixture catalog. Duplicate asset IDs, conflicting results, missing metadata, unavailable assets, and checksum changes fail explicitly. Empty searches return a warning.

from pathlib import Path
from dataset_prototype.adapters import MockMESAClient
from dataset_prototype.contracts import ScientificRequest
from dataset_prototype.discovery import run

client = MockMESAClient(Path("data/fixtures"))
request = ScientificRequest.model_validate_json(Path("run/food_security/request.json").read_text())
result = run(request, client)

prototype discover run/food_security/request.json --output /tmp/discovery.json

MESA, local files, and STAC transports must normalize their results into this same source contract. Discovery filtering establishes candidacy only; it does not establish semantic compatibility or observational support equivalence.

The real workflow adds actual Sentinel STAC/windowed COG acquisition and a checksum-verifying SnapshotClient. NEONClient implements authenticated native file discovery/download but normalization needs verified sensor and variable metadata. The tested unauthenticated NEON data route returns 403; there is no fixture fallback.

Discover semantic evidence alongside assets

discovery.run(request, client, semantic_registry=None) now returns sources with optional typed SourceSemantics: evidence records, concept candidates, extractor identity and a source-metadata binding. Existing source records remain immutable. Source definitions, identifiers, methods and declared support survive in provenance. Absent semantics is omitted from serialization for legacy identity compatibility.

Normalized adapter metadata is marked deterministically derived; native Sentinel band metadata retains explicit locations and source-documented classification. A ConceptCandidate has a retrieval basis, evidence and confidence but no approved relationship. Labels are insufficient to establish equivalence. Module 2 retains all mapping and harmonization authority; source access and MOSAIKS remain separate.

The local SemanticResourceRegistry indexes resource records and explicit fixture concepts behind a portal-independent protocol. External ontology resolvers remain planned/configurable adapters; no ontology graph or source observations are fetched by semantic enrichment. See pipeline contracts and Sentinel regression, landscape and human review.

Finding terms is not approving mappings

An ontology, vocabulary and standard serve different purposes. An ontology is also distinct from a knowledge graph. Discovering an identifier does not establish equivalence.