Skip to content

Real data and model-backed agents

The first replacement of synthetic components is live Sentinel-2 STAC discovery, actual windowed COG pixel reads, and four model-backed roles using an existing local Ollama installation. Synthetic fixtures remain for offline tests. No live run silently falls back to fixtures or deterministic agent responses.

Verified live on 2026-10-21

  • Earth Search returned actual Sentinel-2 L2A items near the example NIWO vicinity.
  • NIR band 8 and scene-classification (SCL) pixels were read through rasterio/GDAL range requests. The initial three accepted pixels had DN values 1318, 1365, and 1280, with SCL=4 (vegetation). These are actual source-native values, not reflectance.
  • The public CLI independently acquired a one-pixel snapshot and called discovery, semantic, harmonization, and QA roles through Ollama's documented /api/chat.
  • The installed deepseek-r1:8b model digest, complete prompt/output transcripts, schema constraints, prompt version, timestamps, and token counts were recorded. Proposal evidence links prompt/response content hashes to the saved transcripts.
  • The latest model QA rejected the proposal over its interpretation of calibration context. Deterministic QA honors that veto. No scientist approval or transformed scientific product is claimed from this proposal. A model rejection is evidence to inspect, not a definitive scientific conclusion.

An earlier live call rewrote a source identifier and was rejected by validation. The boundary now keeps identifiers/targets under deterministic code control and asks the model only for its scientific judgment. Earlier failed transcripts remain under ignored run outputs; no repair or fixture substitution is hidden.

Acquire a small real sample

python -m pip install -e '.[dev,real]'
prototype acquire-sentinel examples/real/sentinel_request.json \
  --longitude -105.586 --latitude 40.0329 \
  --max-cloud 30 --max-items 3 --output run/my-sentinel
prototype propose run/my-sentinel --model deepseek-r1:8b \
  --output run/my-agent-proposal

Use a model already installed on your Ollama server. No model is downloaded and no hosted API key is needed. The optional --endpoint defaults to loopback; unencrypted remote endpoints are rejected. This is real model inference, not an OpenClaw worker invocation. The optional OpenClaw/Verde adapter now uses the installed JSON-only task plugin in a separate container.

The example point is a site vicinity, not a verified NEON sensor location. Request and date selection live in example configuration, not core architecture. The public CLI caps candidates at ten, filters scene cloud percentage, then accepts only SCL 4, 5, or 6 and nonzero NIR DN. A scene-level cloud threshold alone does not prove a pixel is clear. SCL is a quality indicator, not a guarantee of suitability. Failed/rejected candidates retain evidence and never become fabricated observations. GDAL fetches the compressed blocks needed for individual pixels, not whole scenes.

What remains unchanged scientifically

Ingestion preserves raw digital numbers. Native STAC scale/offset, platform, sensor, processing baseline, acquisition timestamp, footprint, native CRS/indices, SCL, and complete STAC Item metadata are retained. Physical reflectance calibration is not applied behind the recipe engine. Field-to-pixel equivalence, temporal aggregation, and ecological inference are not authorized by this example.

Instantaneous satellite timestamps are represented by equal start/end times; no artificial one-second or daily interval is invented. Core validation accepts zero duration only when source support explicitly says instantaneous. The 10 m native-CRS pixel footprint is retained as a WGS84 bounding box; it is not replaced with a site centroid or declared equal to field-observation support.

A source checksum with scope extracted_window covers the saved pixel-evidence JSON, not the full remote COG. Snapshot content digests separately protect the normalized resource and record descriptions. SnapshotClient verifies all evidence files and supports offline reuse through the existing discovery/access contract. Large native rasters and live outputs stay out of git.

License evidence

Earth Search's Sentinel collection labels its license proprietary and links the specific Copernicus Sentinel legal notice. The adapter recognizes only that verified notice link for this collection, retains the original STAC label, records the legal-notice identity and attribution, and preserves the native collection evidence. Unknown/custom licenses remain uncertain and require review. Do not infer reuse permission from a satellite name alone.

Agent boundary and human review

The four roles receive bounded scientific metadata and return typed JSON. They have no tools, filesystem handles, shell, network tools, or data-write capability. Full native STAC metadata stays in source evidence; model prompts contain the normalized scientific subset, explicitly recorded in transcripts.

Model-generated mappings always require human review, regardless of confidence. The model cannot rewrite source/target identity or mark a recipe approved. A model QA rejection or revision request becomes a recipe veto enforced by deterministic QA. Its content digest excludes only approval state and the review itself; editing scientific content invalidates that model review. Approval-state edits alone cannot bypass the veto. Malformed/truncated output fails and is retained for inspection. Model calls can be nondeterministic even with seed/temperature settings; exact transcripts are the evidence, not an inference-reproducibility guarantee.

Inspect discovery_proposal.json, semantic_mappings.json, draft_recipe.json, proposed_recipe.json, model_qa.json, qa_report.json, review.json, agent_traces.json, sources.json, and records.json. On failure, completed stages, transcripts, and a failure report are retained. No scientific transformation runs inside prototype propose.

Revise scientifically rejected/unknown proposals and obtain renewed model QA before human approval. For a valid proposal, save the intended approved recipe, record an explicit human decision with prototype review, then use the existing harmonize and encode stage commands. See the review workflow. An expert can disagree with model QA, but this prototype requires a recorded revised proposal/review rather than a silent override.

NEON: implemented access boundary, token required

The official NEON API authentication documentation specifies the X-API-Token header. The tested NIWO July 2023 /data request returned 403 without a token. Public product metadata was retrieved successfully. The current API documentation specifies CC BY 4.0, not the old CC0 assumption.

Configure NEON_API_TOKEN in your process environment; do not paste it in chat, commit it, or place it in an example file. The CLI reads it without logging it.

prototype neon-files --product DP1.00002.001 --site NIWO \
  --month 2023-07 --release RELEASE-2025 --output run/neon-files.json
# Select one exact small CSV name from the provider response, then:
prototype neon-files --product DP1.00002.001 --site NIWO \
  --month 2023-07 --release RELEASE-2025 \
  --filename EXACT_PROVIDER_FILENAME.csv --output run/neon-native

The adapter requires an explicit release and exact provider filename, checks the file-size limit, retains native metadata/CSV/checksum and license/citation requirements, and sends tokens only to the NEON host. It does not guess sensor coordinates or height, normalize uncertain variable semantics, or claim live CSV acquisition before a token-backed download has actually succeeded. Sensor positions, variable metadata, quality flags, uncertainty, and interval definitions must be verified before this native CSV becomes harmonized observations.

Next scientific step

Complete authenticated NEON acquisition and sensor metadata normalization, then have an expert specify a defensible field/sensor-to-pixel question and support relationship. NIR digital numbers are not air temperature, vegetation height, or food/water indicators. Keep that scientific decision distinct from making live adapters and models run.

Metadata consensus and direct streaming

The streaming harmonization module adds metadata-only specialist dialogue, digest-bound unanimous review, documented Sentinel-2 scale/offset calibration, and bounded window-to-encoder execution without raw or harmonized data files. The existing fixture/snapshot commands retain their historical materialization behavior. Current streamed vectors are reference numeric features; model consensus still requires an actual human decision before execution.