Complete research architecture
System Map The end-to-end pipeline from physical sample to evidence ledger, grouped by domain and linked by dependency.
Every status below is a research/evidence status, not a software-implementation status. A stage appearing on this map means it is documented here — it does not mean the stage itself has been implemented or executed. Statuses are read from the same evidence as /pipeline , /xai , and /hardware .
Open 3D digital twin → Open device view →
Physical Layer 1 Sample Executed ↓ 2 Sensor System Executed depends on 1 ↓ 3 Signal Acquisition Executed depends on 1 Data / ML Layer 4 Preprocessing Executed depends on 1 ↓ 5 Feature Extraction Executed depends on 1 ↓ 6 Drift Analysis Executed depends on 1 ↓ 7 Machine Learning Executed depends on 1 ↓ 8 Calibration Planned depends on 1 ↓ 9 Explainability Executed depends on 1 Edge Layer 10 Compression / Quantization Not executed depends on 1 ↓ 11 TinyML Artifact Not executed depends on 1 ↓ 12 nRF52840 Deployment Hardware blocked depends on 1 ↓ 13 Edge Prediction Blocked depends on 1 Operations Layer 14 Gateway / Telemetry Planned depends on 1 ↓ 15 Monitoring Planned depends on 1 ↓ 16 Adaptation / Retraining Planned depends on 1 Research Layer 17 Experiment Registry Executed ↓ 18 Evidence Ledger Executed depends on 1 Select a stage in the map — or in the list below — to inspect its role, dependencies, and evidence.
All stages (text list) 1. Sample Executed The gas sample presented to the sensor array. 2. Sensor System Executed 16-channel metal-oxide gas sensor array. 3. Signal Acquisition Executed Acquisition and digitization of the raw sensor response (128 features: 16 sensors × 8 response characteristics). 4. Preprocessing Executed Training-only standardization and chronological split preparation. 5. Feature Extraction Executed 128-feature ontology: 16 sensors × 8 response-characteristic features (steady-state, normalized, transient EMA). 6. Drift Analysis Executed Per-feature and global drift measured against the Batch-1 reference distribution. 7. Machine Learning Executed Four chronologically-evaluated classical models under FIXED_ORIGIN and EXPANDING_WINDOW protocols. 8. Calibration Planned Causal confidence calibration (ECE, Brier, NLL, risk-coverage) for selective prediction. 9. Explainability Executed Resource-aware explanations (permutation importance, intrinsic coefficients/impurity, single-feature ablation) — deliberately not SHAP or LIME. 10. Compression / Quantization Not executed INT8 post-training quantization planned; not yet executed at any tier. 11. TinyML Artifact Not executed Deployable, quantized, cross-compiled model artifact for the nRF52840. 12. nRF52840 Deployment Hardware blocked Physical flashing and execution on a Nordic nRF52840 (Cortex-M4F) development kit. 13. Edge Prediction Blocked On-device inference output, produced entirely on the MCU. 14. Gateway / Telemetry Planned Planned USB / Serial / BLE bridge forwarding device telemetry to a host gateway. 15. Monitoring Planned Live dashboarding of deployed-device drift, confidence, and health. 16. Adaptation / Retraining Planned A live production retraining loop triggered by detected drift. 17. Experiment Registry Executed Append-only registry of every executed experiment, with dataset/config/model hashes and git commit. 18. Evidence Ledger Executed Claim → experiment → dataset/config hash → git commit → result artifact mapping.