27-stage research pipeline

Captures Python/OS/package versions, git commit, and seed for every subsequent stage.

Official UCI archive downloaded and SHA-256 verified against the published archive hash.

Verified 13,910 samples, 128 features, 10 batches, 6 classes; 0 missing values.

FIXED_ORIGIN (train batch 1, test 2-10), EXPANDING_WINDOW, and IID_DIAGNOSTIC protocols frozen in config.

Standardization fitted on training data only per protocol; no leakage across chronological boundary.

Logistic Regression, Random Forest, RBF-SVM, MLP trained on Batch 1 (seed 42).

All 4 models evaluated on batches 2-10 under FIXED_ORIGIN, EXPANDING_WINDOW, and IID_DIAGNOSTIC.

Per-feature and global drift (normalized Wasserstein, standardized mean shift) vs. Batch 1.

Claim-evidence matrix (paper/claim_evidence_matrix.csv) links each candidate claim to its artifact; C006 explicitly UNSUPPORTED.

EXP-XAI-0001 (registry rows EXP-XAI-0001-C1..C4). Intrinsic coefficient explanations for MODEL-C1, intrinsic impurity for MODEL-C2, model-agnostic permutation importance (macro-F1) and single-feature-ablation local explanations for all 4 frozen FIXED_ORIGIN models. No retraining; chronological split unchanged. Full applicability matrix in results/xai/stage09_manifest.csv. See docs/experiments/STAGE09_RESOURCE_AWARE_XAI.md. This stage prepares evidence for Stages 10-12 — it does not itself report fidelity, stability, or latency conclusions.

EXP-XAI-FIDELITY-001; method-specific global, local, sensor-group, family, matched-random, and bootstrap evidence. Local ablation is labelled ABLATION_CONSISTENCY because its perturbation evaluation is partly circular.

EXP-XAI-STABILITY-001; global chronological, physical sensor/family, natural-neighbor, matched cross-sectional, relative, and fidelity-linkage evidence. No universal stability score.

EXP-XAI-LATENCY-001 measured controlled single-thread workstation cost with raw nanosecond timings and hardware-independent operation counts. Global and local scopes remain separate. Host latency is not an estimate of nRF52840 latency or energy; MCU latency, physical energy, Flash, and SRAM remain NOT_MEASURED/BLOCKED. Stage 13 did not execute.

GATE-EMBED-EXPORT-001 architecture and numerical-equivalence protocol frozen. This is not an executed model export or performance experiment. Export, quantization, firmware, compiled Flash/SRAM, MCU timing, and energy remain NOT_EXECUTED/NOT_MEASURED.

EXP-EMBED-FP32-EQUIV-001 completed with scientific outcome FAILED. C1 and C4 passed score, probability, normalization, golden decision, and boundary decision checks but failed frozen preprocessing criteria. Host-compiled validation only; quantization and MCU work remain unexecuted.

EXP-EMBED-C1-PREPROC-REPAIR-001 completed with outcome FAILED. P0 baseline reproduced; P1-P4 all failed unchanged universal preprocessing criteria. C4 remained frozen failed, and quantization, firmware, MCU measurement, and hardware resource claims remain unexecuted.

GATE-C1-FUSED-PREPROC-001 frozen. No fused implementation or equivalence output was generated. Future EXP-EMBED-C1-FUSED-EQUIV-001 is authorized separately; quantization and hardware work remain prohibited.

EXP-EMBED-C1-FUSED-EQUIV-001 PASSED prospectively frozen score, probability, normalization, deterministic-generation/execution, golden-decision, and boundary-decision criteria on the host. Fused XAI, quantization, MCU compilation, and hardware measurements remain NOT_EXECUTED/NOT_MEASURED.

EXP-EMBED-C1-FUSED-XAI-EQUIV-001 PASSED all prospectively frozen local attribution, vector, sign, Top-K, additivity, semantic, determinism, and lineage criteria on the host. Quantization, MCU execution, and hardware measurements remain unexecuted.

EXP-MCU-C1-FP32-PORT-001 is BLOCKED_HARDWARE. No physical Nordic/J-Link/CMSIS-DAP device was detected, so no target was guessed, no toolchain was changed, and no build, flash, physical correctness, linked ROM, or linked static-RAM evidence was produced.

Requires stage 15.

Requires stage 15.

Requires Nordic Power Profiler Kit II and stage 15. No raw traces exist.

Requires stages 12 and 18.

Multi-objective (accuracy, fidelity, stability, latency, Flash, SRAM, energy) frontier — needs stages 10, 11, 17, 18.

15 figures + 5 tables generated from Checkpoint 1-2 evidence, each with a source CSV under results/figures/sources/.

Claim-evidence matrix and figure/table indices exist; Results/Discussion prose intentionally left undrafted pending further evidence (see paper/RESULTS_DRAFT.md).