Skip to content

Current-state audit

Audit date: 2026-10-21. This records the repository before the review-hardening and narrative-site pass. Existing implementations and historical evidence are retained.

Three scientific stages

Stage Implemented modules Boundary Actual status
Discovery & access discovery, adapters, real_data, stream_sources, stream_rasters, neon, network, storage ScientificRequest → SourceResource / bounded stream Offline fixture transport and live Sentinel STAC/COG access; MESA explicitly unavailable; authenticated NEON normalization pending
Harmonization & review agents, model_agents, consensus, harmonization, harmonized_rasters Sources → mappings / recipe → canonical scalar or raster outputs Allow-listed execution, metadata dialogue, deterministic QA, exact-recipe human gate; existing free-text specialist review leaks jurisdiction
Representation representations, mosaiks, embedding_dataset Validated harmonized outputs → vectors / lineage Scalar reference encoders; empirical MOSAIKS with generated held-out diagnostics; real raster execution blocked upstream

orchestration, mosaiks_pipeline and cli compose these stages. science provides a descriptive synthetic comparison, not a food/water-security model. Contracts are frozen Pydantic models, plus an in-memory raster dataclass whose content is hashed.

Contracts, recipes and provenance

contracts.py defines source, request, discovery, mappings, recipes, observations, representations, QA, human decisions and track configuration. stream_contracts.py defines metadata plans, votes, messages, encoder configuration and consensus. Recipes are versioned JSON, scoped to conservative core or an explicit track. Source IDs and targets are code-owned; models cannot supply arbitrary executable code.

Scalar provenance nests source records, source metadata, recipe, executor, QA, human decisions and core/track parents. Streaming output substitutes compact IDs, window hashes and support for persisted raw/harmonized values. MOSAIKS stores frozen filters, full training context, separate labels, numeric feature shards and checksums. Metadata hashes are not raster-byte hashes. Discarded input windows cannot be reconstructed solely from an output hash.

Observed Sentinel failures

Inspected the complete available recipe, QA, consensus, dialogue and trace artifacts in run/sentinel-stream-consensus-bounded/ and run/mosaiks-real-consensus/. The former completed five reviews and requested revision; final QA rejected execution. Semantic, units, support and rights roles repeated negative-reflectance, licensing, mask and checksum concerns. Rights asserted potential usage-policy breaches from numeric values without supporting terms. Metadata-only checksums were repeatedly conflated with raster integrity. Several claims contradicted declared executor behavior. The MOSAIKS run retained a valid draft and three revision requests, then rejected a rights self-citation. No approved real harmonized MOSAIKS dataset was produced.

The exact native B08 coefficients were scale 0.0001 and offset -0.1. Unit 1 is valid dimensionless notation. Negative calibrated values were never automatically clipped. Uncertainty scaling exists, while calibration-parameter covariance does not. Human approval was never fabricated. OpenClaw traces expose normalized plugin JSON, not raw provider text, token usage or model-weight digests.

Concrete debt and minimal response

  • Review findings are free text; topics and evidence provenance need typed boundaries.
  • Integrity has no independent reviewer domain; checksum claims leak into other roles.
  • agree/revise/reject/abstain conflates missing evidence with demonstrated failure.
  • Repeated concerns need stable finding identities and references, not another vote.
  • Quality selection is performed in source adapters, separately from recipe calibration, but historical output labels do not reliably distinguish calibrated and filtered data.
  • Legacy snapshot/fixture paths materialize data; new streaming paths must remain distinct.
  • Scalar and raster executors duplicate a small allowlist; parity tests are preferable to a speculative universal execution framework at this stage.
  • The flat website mixes introductory narrative with implementation reference. Preserve MkDocs, Markdown URLs, Material overrides, search and Pages deployment; add a clear narrative/documentation split rather than another frontend.

Tests, tracks and deployment

The audited suite has 110 offline tests covering contracts, discovery, calibration, review gates, streaming, MOSAIKS, provenance, cache integrity and independent track configuration. Food/water share immutable sources/core results; the third track is disabled. No new source or track is warranted by these findings.

The existing Pages workflow builds MkDocs strictly. Template synchronization and container publishing are separate manual historical workflows, not required for site deployment. No configured linter/type checker exists; do not invent a claim that one ran. Tests and strict documentation builds remain required checks.

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.