Skip to content

Current implementation status

Resource planning — 26 October 2026

The grant-period budget targets all-TACC persistent storage and versioned backup, with VERDE/Jetstream2 compute. Its retention ledger explicitly carries earlier releases forward and separates capacity growth, authorized closeout assumptions and backup history. Rates, churn and scientific production remain conditional; no remote backup is implemented. Historical OSN plans remain explicit older prototype records.

Local operational prototype — 23 October 2026

The data management workbench connects generated harmonization/MOSAIKS output to immutable typed records, SQLite lineage lookup, deduplicated files, portable JSONL/schema exports, nine workspace plans, three planned products, resource plans, measured usage, feedback and actual local snapshot/restore. The loopback UI supports filtering, pagination, evidence traversal and checked range downloads. No scientific approval or arbitrary execution API exists. Scale probes measure metadata indexing only. Remote MESA/iRODS/OSN/NDP/TACC services, awards and global scientific products are not claimed. The undefined third executable track is unchanged.

The documentation synchronization audit separates implemented metadata diagnostics from planned ontology depth, composition, accumulation, dataset potential/diversity and flagship analysis. See the ontology and dataset frameworks. No approved real Sentinel MOSAIKS release is claimed.

Semantic discovery and landscape

Module 1 now preserves typed source semantic evidence and optional concept candidates. Normalized metadata and native annotations retain distinct evidence kinds; source identifiers/definitions survive and stale bindings are rejected. Mapping and approval remain Module 2 responsibilities. MESA and external ontology-portal integration status has not changed; no universal ontology or new data source is claimed.

The Semantics resource includes the canonical 86-resource inventory, auditable provisional assessments, generated coverage comparison and linked glossary. The benchmark produces metadata-only Parquet diagnostics and review-only portfolio suggestions. Default runs use generated source metadata; explicit runs can inspect retained real source metadata without reading observations. Neither establishes scientific compatibility or approval.

Evidence-bound review and website

The new review-metadata command validates domain-scoped evidence and frozen-recipe findings from six specialists, then requests QA synthesis referencing existing IDs. Reports separate missing evidence, failure, eligibility and mandatory human review. They never authorize execution or bypass historical model vetoes. A matching report with accepted QA and no unresolved findings is now an additional required gate for real Sentinel streams, checked before source access. The legacy consensus command retains its original dialogue/vote contract. See the review guide, audit and Sentinel comparison.

The website now separates narrative research pages from technical documentation, with generated contract reference and source-maintained architecture and conceptual figures. The existing Material/Pages configuration remains the deployment mechanism.

Metadata consensus and streaming

The new streaming module separates metadata-only OpenClaw specialist dialogue from lazy Sentinel native-window access. The first target is B08 reflectance using documented scale/offset. Exact-plan consensus, deterministic QA and human approval gate direct bounded-batch encoding. No raw or harmonized data files are produced by this path. Scalar reference encoders remain available; the independent MOSAIKS raster path is implemented below. Legacy fixture/snapshot paths still materialize data explicitly.

MOSAIKS representation module

The MOSAIKS module accepts harmonized raster outputs, fits empirical random convolutional filters on training tiles only, and writes dense feature matrices with lineage and separate verification labels. Generated diagnostics run through actual harmonization and MOSAIKS computation; held-out ridge evaluation includes mean/shuffled-label baselines and spatial leakage checks. This is numerical pipeline verification, not independent ecological validation. Fresh raster-layout consensus and an actual human decision gate real execution.

Working now

Three independently callable typed modules compose into a complete offline run: question/configuration → fixture discovery/access → semantic proposal → versioned recipe → QA → approved deterministic execution → shared observations → track products → modality-specific vectors → JSON-LD/source lineage → descriptive synthetic downstream comparison. CLI, schema export, review artifacts, checksum validation, immutable content-addressed storage, and separate track outputs work.

Mocked

MockMESAClient searches local normalized fixture metadata and verifies local asset checksums. The default fixture agent roles use deterministic curated rules; the separate real workflow uses model-backed Ollama roles. Six records are synthetic field and satellite scalar measurements, not provider records, native schemas, images, or actual observations.

Real external integration

Live Earth Search STAC discovery and actual Sentinel-2 NIR/SCL COG pixel reads were exercised on 2026-10-21. Three initial actual DN observations and an independent one-pixel CLI acquisition preserve native metadata, calibration context, support, quality and licensing. Four real local Ollama role calls completed with versioned model transcripts. Ollama baseline QA rejected its proposed recipe; no scientist approval or executed model-proposed transformation is claimed.

NEON public product metadata was retrieved; the data-file endpoint returned 403 without a token. Authenticated native CSV access code is implemented and unit-tested, but token-backed acquisition/normalization has not yet been verified. Live MESA remains unimplemented. OpenClaw now has an isolated Verde-backed JSON-only task adapter using the installed 2026.5.18 plugin. See live evidence and limits.

Future MESA integration

Verify deployed MCP tool schemas/version, provide user-configured transport, normalize real discovery metadata/rights/support to SourceResource, and implement search/describe/fetch/history through verified capabilities. The current source README's pre-alpha status differs from the tool families described in overview docs. See integration inspection.

OpenClaw and Verde

An isolated gateway now uses the existing OpenClaw image and configured Verde provider/model. No existing working-group/Slack routes are changed or mounted. OpenClawBackend calls the actual JSON-only llm-task plugin over /tools/invoke. The plugin disables inference tools and validates schemas. Route checks reject provider/model substitution and there is no Ollama/mock fallback. Model weight hash and usage are unavailable from the plugin and are explicitly unreported.

Live testing identified two compatibility requirements: fully qualify the model as verde/js2/gpt-oss-120b, and set compat.supportsTools: false so Verde receives no unsupported empty tools array. See OpenClaw/Verde setup. The live four-role OpenClaw → Verde run completed with schema-valid outputs from js2/gpt-oss-120b. Model QA requested revision and deterministic QA blocked execution. The isolated gateway is running; scientific acceptance remains pending. Full working-group scheduling and authenticated human review remain future work.

Scientific assumptions

In the default fixture demo, synthetic fields and satellite values have intentionally aligned coordinates, daily intervals, and common support. Metadata matches curated identity targets. Scalar/squared encoders illustrate reusable interfaces, not ecological feature quality. The downstream correlation illustrates software composition only. The real Sentinel path retains uncalibrated band-8 DN, actual 10 m pixel footprints, instantaneous acquisition times and SCL. Scale/offset remain metadata, not a hidden conversion to reflectance. No real field/pixel join has been asserted. No universal ontology, universal latent space, food/water model, real cross-modality support alignment, or scientifically validated prediction is claimed.

Human decisions

Novel/low-confidence/disputed/ambiguous meaning, changed temporal/spatial support, joins, and uncertain rights require exact-recipe review records. Unknown mappings must be revised explicitly; QA rejection/known rights prohibitions block execution. Default conservative fixture identity mappings pass deterministic QA. Real model mappings always require human review; model QA vetoes cannot be bypassed by changing approval state, and scientific edits invalidate the model review. Human identity/signatures and tamper-resistant external audit are not implemented.

Science tracks

Food and water run independently through shared modules using the same immutable sources/core cache. Their only demonstrated scientific configuration is the common synthetic comparison; encoders differ. Third track is disabled until scope is specified; tests prove configuration-only extensibility with an enabled synthetic copy. Default orchestration does not yet load arbitrary per-track recipe files.

Testing status

Offline unit, contract/integration, CLI, deterministic repeatability and track independence tests are implemented. An earlier MOSAIKS integration checkpoint passed 110 tests; current verification is recorded in the website review, the installed CLI ran both tracks, and the strict website build passed. Rendered navigation/readability and local links/fragments were checked. Testing describes coverage and scientific gaps. CI also builds the existing MkDocs site strictly.

Current limits

The legacy executor buffers inputs; the new streaming module limits each call to a bounded microbatch. Large partitioned joins/aggregation remain unimplemented and are refused by streaming preflight. Source/core results use integrity-checked read-through caching and atomic publication; distributed cache coordination remains future work. Linear unit/calibration uncertainty scaling is implemented; calibration-parameter uncertainty, covariance and broader propagation remain incomplete. Only WGS84 identity coordinate normalization is supported. Native Sentinel STAC/window access now exists; sensor-aware NEON normalization and real field/satellite alignment remain pending. Local JSON review is an auditable artifact, not an authenticated permissions system.

Next milestone

Complete token-backed NEON CSV acquisition and verify sensor/variable metadata. Review the actual model QA rejection and improve evidence/contracts without forcing approval. Then specify and review a defensible real NEON + satellite support and temporal alignment question. Then add a non-raster cross-network/trait vertical slice, followed by ORNL DAAC context, before making generality claims.

First live metadata consensus result

Real Earth Search metadata fed four specialist analyses, a schema-valid recipe, and five actual OpenClaw/Verde peer reviews. The final run requested revision; coefficient checks passed, while final QA enforced the model veto. No calibrated real-data embeddings were emitted. Model claims and executor limitations remain separate; see the live evidence.

MOSAIKS verification result

The generated harmonization → MOSAIKS → held-out ridge workflow produced 32 vectors and passed numerical verification, alongside 110 offline tests. A real-metadata 640×640 m raster plan was reviewed through Verde, but the run stopped on an invalid rights peer citation after three valid revision requests. No real calibrated raster/vector dataset was emitted. See the MOSAIKS result for diagnostic metrics, the concrete candidate and remaining scientific gaps.

Model trust hardening

WORKING: a ten-operation closed registry, declarative decision IR, constrained capabilities, deterministic parameter resolution, scientific validation, explicit review/failure states and no-model approved execution. The fixture demonstration is MOCKED, including its human decision. OpenClaw/VERDE and model scientific roles remain PROTOTYPE; no new live call or approved real Sentinel product is claimed. General support algorithms, ontology composition and authenticated human approval remain PLANNED. Detailed status and limits.

CubeDynamics backend

WORKING in generated offline tests: external CubeDynamics 0.1.0rc3, seven registered capabilities, source-bound JSON backend plans, lazy xarray/Dask validation, explicit bounded materialization and typed transformation lineage. The generated Sentinel-like window → CubeDynamics → MOSAIKS example is PROTOTYPE; its model/reviewer are MOCKED. No new live source/inference run or approved real Sentinel product is claimed. General support-changing representation bridges and reprojection are not exposed. Integration · API audit.

Visual evidence report

The pipeline state report provides an auditable visual snapshot, including measured local test/registry/inventory counts and explicit scientific limitations. Source, figures, machine inventory and rebuild scripts are retained in the repository. This communication artifact does not change scientific approval or external integration status.

Scientific validation update — 22 October 2026

The new scientific validation findings retain real graph measurements, canonical requirement diagnostics, scoped dataset/pair assessments and live model results. Earlier unmeasured/planned statements describe the prior architecture snapshot: depth is now measured for three local graphs, while exact scientific mappings, independent ecological validity and the weaker-model hypothesis remain unvalidated. The real Sentinel and NEON–Sentinel cases still require explicit expert decisions; no scientific execution is approved.