Introducao | O perfil: Argissolo Vermelho-Amarelo distrofico tipico | Diagnosticos manuais (sanity check) | Tres classificacoes paralelas | SiBCS 5a edicao | WRB 2022 | USDA Soil Taxonomy 13a edicao | Comparacao cross-system | Relatorio HTML | Cruzamento com priore espacial (opcional) | Cruzamento com SoilGrids 250m (global) | Resumo

0. O começo de 30 segundos | A. Interface Shiny (sem código) | B. Uma chamada R, uma fixture | 1. Construindo um PedonRecord do zero | 2. Classificando nos três sistemas | Atalho: classificar nos três sistemas em uma chamada | 3. Inspecionando o key trace | 4. Provenance + evidence grade | 5. Cross-system: o mesmo perfil em três taxonomias | 6. Onde ir a partir daqui | 7. Onde reportar bugs / sugerir features / pedir ajuda | Referências canônicas

1. Set the scene | 2. Spatial guide -- before any pedon data | 3. Multimodal extraction with local Gemma 4 | 4. Spectral analogy | 5. Deterministic classification | 6. HTML report | 7. GIS export | 8. The complete picture | Summary

1. Build the pedon | 2. Run the WRB key | 3. Inspect the principal qualifier resolution | 4. Inspect the supplementary qualifier resolution | 5. Compose the name | 6. Redundant and sibling qualifiers | 7. Evidence grade | 8. Render a self-contained pedologist-facing report | 9. Tier-3 strict mode for borderline pedons | Summary

1. The mock provider | 2. Schemas and prompts | 3. Walking the extraction loop with the mock | 4. Merging extraction into a PedonRecord | 5. Provenance and evidence grade | 6. Production: swap the mock for an ellmer chat | 6a. Local-first with Gemma 4 (Ollama) | 6b. Cloud providers (Anthropic, OpenAI, Google) | 6c. The one-liner: classify_from_documents()

1. Protocol overview | 2. Mini-benchmark on the canonical fixtures | 3. Confusion at the qualifier level | 4. Benchmark protocol against WoSIS (run-once, paper-grade) | 5. Reporting standards | 6. Offline canonical-fixture run (release-time sanity check) | Summary | 7. v0.9.27 -- per-page retry + graceful degradation | 8. v0.9.30 -- bundled WoSIS sample for offline / CI testing

1. The same Ferralsol through three keys | 2. Cross-system table on the canonical fixtures | 3. Where the systems diverge | 4. Recovering the qualifier-level correspondence | 5. Validating the SiBCS ↔ WRB alignment | 6. Use cases for cross-system classification

The pipeline | A minimal run | The provider argument is required | The SoilGrids depth prior | Reading the evidence grade | When to use it

0. The 30-second on-ramp | A. Zero-code GUI | B. One R call, one fixture | 1. Building a PedonRecord from scratch | 2. Canonical fixtures | 3. Calling the diagnostics directly | 4. Diagnostic matrix across the canonical fixtures | 5. RSG-derived diagnostics: argic and mollic families | 6. End-to-end WRB classification | 7. Provenance and evidence grade | 8. Interoperability with aqp | 9. Module 4 -- OSSL spectroscopy bridge (gap-filling) | 10. Module 3 -- SoilGrids / Embrapa spatial prior (sanity check) | 11. Module 2 -- Multimodal extraction via ellmer | 12. SiBCS 5ª edição (Embrapa, 2018) | 13. v0.7 scope and the v0.3.3+ roadmap

1. The KSSL + NASIS join (load_kssl_pedons_with_nasis) | 2. The four levels (benchmark_run_classification) | 3. Headline numbers (v0.9.31, n=2638, ±1.7 pp CI) | 4. Release-by-release trajectory | 5. The four levels are independently measurable | 6. v0.9.25 -- the KST canonicaliser story | 7. What's missing (roadmap) | Summary

The idea | A first run | Provenance changes the answer | The sensitivity ranking | Backward compatibility | In the Shiny app

1. Start from a partially-described pedon | 2. Module 3 -- spatial prior consistency check | 3. Module 4 -- OSSL gap-filling | 4. Re-classify with the gap-filled pedon | 5. Combining priors and posteriors | Summary

Launching the app | The eight tabs | Pedon | Classify | Photo | Spectra | Spatial | Uncertainty | Report | Settings | Notes

Packages | Data | Utility: evaluation metrics (base R) | Spectra preprocessing (reflectance → absorbance → resample → SNV → movav) | Split: calibration (datC) vs test (datV) | Baseline: PLS (train on datC, evaluate on datV) | soilVAE: supervised VAE regression (skips automatically if TF/Keras unavailable) | Prepare matrices (scale X using train stats; keep y in original units) | Fit and evaluate (only runs when TF/Keras is available) | Final comparison table (PLS vs soilVAE)

Overview | What is the SWRC? | Key features | Quick-start | Installation | Minimal example | Data preparation | Expected columns | Unit handling | Model architecture | Monotone integral layer | Inputs to the model | Physics constraints | Saving and loading models | Visualisation | SWRC curves | Predicted vs. observed | Training history | Texture classification | Session information | References