Results dashboard

Every chart below is generated at build time directly from CSV artifacts in results/baselines/ and results/drift/ — see results/baselines/fixed_origin_summary.csv and results/drift/global_drift_by_batch.csv.

Accuracy by chronological test batch (FIXED_ORIGIN)

Macro-F1 by chronological test batch (FIXED_ORIGIN)

Drift degradation: accuracy drop, Batch 2 → Batch 10

Every evaluated model loses accuracy from Batch 2 to Batch 10 under FIXED_ORIGIN — see claim C003.

IID vs. chronological generalization gap

The stratified-random IID diagnostic accuracy exceeds mean chronological future-batch accuracy for every model — quantifying how much a random split overstates performance (claim C005, diagnostic only, never a primary result).

Host-serialized model size

Serialized (joblib, host CPU) size — not an embedded or quantized footprint. Flash/SRAM figures remain NOT EXECUTED until stage 16.

Global feature drift vs. Batch 1

Generated figures and tables

15 figures and 5 publication tables are generated directly from evidence — browse the full index on the Paper page.

Not yet executed

Per-class F1, confusion matrices (see figure gallery on the Paper page for the chronological confusion-matrix evolution figure — that one already exists), explanation fidelity, explanation stability, latency, Flash/SRAM, and energy remain below.