RuView/tools/bfi_learning/RESULTS.md

Measured research checkpoint — September 29, 2026

The pipeline now imports public CSI and real captured BFI, trains bounded LSTM and GRU baselines, preserves a final holdout, and records reproducible numeric checkpoints. This is research tooling. KIT-style BFI person identification and state-of-the-art performance have not been validated.

Public CSI experiment

NTU-Fi HumanID, 14 published subject classes; custom split of 238 training, 56 validation and 546 sealed holdout recordings. Physical acquisition-session separation is unknown. This differs from the author's folder mapping; see the comparison contract. Each of the six architecture-comparison runs completed 200 epochs on the RTX 5080, using seeds 1729, 1730 and 1731. Validation loss selected each checkpoint. The final holdout has not been scored.

Method	Parameters	Validation recording accuracy	Interpretation
LSTM, hidden 64, window 128	105,358	56/56 for each seed	Sequence baseline
GRU, hidden 64, window 128	79,246	56/56 for each seed	24.8% fewer parameters; same recording accuracy on this validation split
Nearest centroid on mean amplitudes	—	46/56	No temporal order
Fixed-alpha ridge on mean amplitudes	—	56/56	No temporal order

The static ridge control matches the sequence models. These results do not establish that the sequence models learned gait or motion. The cause could be body information, placement, room, device or acquisition context. Separate sessions and deployments are needed to distinguish these explanations.

Peak PyTorch allocation was 192,955,392 bytes for LSTM and 198,683,136 bytes for GRU. Fewer GRU parameters did not reduce measured peak allocation. GPU context and other processes are excluded. Trainer receipt formats changed between the LSTM and GRU runs, so their elapsed times do not isolate an architecture speedup. No paid GPU was needed; task Vast.ai spend remains USD 0 against the authorized cumulative USD 200 cap.

Temporal controls: completed matched comparison

Twelve further GRU runs completed the predeclared control protocol: four conditions, three seeds, 50 epochs each, with the same files and model settings. These diagnostic runs are excluded from final-model selection.

Input	Time order	Correct validation recordings, three seeds	Mean window accuracy	Mean validation cross entropy
Raw	Original	56, 56, 56 out of 56	99.531%	0.01855
Raw	Shuffled	56, 56, 56 out of 56	99.671%	0.01028
Recording mean removed	Original	56, 56, 56 out of 56	96.099%	0.14544
Recording mean removed	Shuffled	52, 55, 56 out of 56	96.112%	0.12256

Chronological motion is unnecessary for perfect recording accuracy on this validation split. Removing each recording's mean also preserves high accuracy; other distributional information remains predictive. These controls cannot identify its physical cause or establish generalization to new sessions. Whole-recording demeaning is an offline operation. Windows overlap, and all seeds share the same validation recordings.

The updated trainer's raw/original seed-1729 checkpoint tensors exactly match the prior GRU checkpoint selected at the same epoch. This real-data parity check supports unchanged default numerics after the input-validation changes. Every diagnostic receipt has test: null; the final identity holdout is sealed. Exact results and the plot are reports/controls-comparison-v1.json and .png on ruvultra. src/controls-v2/summarize_controls.py verifies receipts and checkpoint hashes and reproduces the summary.

Real BFI import and prediction

The offline importer reproduced 66 and 96 unique reports from the existing captures, separating 20/40/80 MHz feedback into 192/366/750 circular-angle features. Report counts match the independently checked decoder results. The full observation intervals retain 23.291- and 16.393-second quiet periods. Relative timestamps preserve nanosecond ordering and retain the integer origin; real feature arrays and sequence values were unchanged by that timestamp fix.

A small CPU prediction pilot used one 65-frame, 80 MHz segment, with purged chronological partitions and four-frame input windows. It completed 20 epochs; five validation windows remained. Raw feature MSE was 0.004081, versus 0.002811 for repeating the last observation. This pilot did not beat persistence. Its test partition remains unscored. It proves the offline capture-to-training path executes, not movement, identity or generalization.

Active-traffic continuity

The new analyzer verifies the capture and traffic receipts and retains the original full-capture metrics. During the actual approximately 30-second sender intervals, the two captures contain 66 and 96 reports: 2.2 and 3.2 reports/s. Weighted occupied one-second bins cover 43.33% and 70.00% of those intervals; maximum silence remains 8.392 and 6.437 seconds. No reports occurred in the following approximately 15 seconds. The continuity target still fails. These additional rates use a different denominator; they are not an RF performance gain. See CAPTURE-CONTINUITY.md.

Public BFI activity benchmark

The research update pins a licensed public dataset and a 3.365 GB subset across three day/room setups and two device/view labels. Day and room are confounded, and the selected filenames cover three people. Default download scope is only 120 training/validation traces (929,362,200 bytes) plus the dataset license card. Held-out day-5 views require an explicit option. That default download completed in 92.9 seconds: all 120 traces and the card passed exact size/hash checks, totaling 929,368,919 bytes. No partial files or download lock remain, and no day-5 directory was created. The private receipt is data/public-bfi-c3013b3/download-receipt.json on ruvultra.

Implemented a separate VHT MU 3x1 report-body decoder. Two independent executions matched all 13,104 quantized angles across 14 public sample bodies, all control fields and signed SNR codes against tshark 4.2.2. All frame CRCs passed separate checks. The existing HE/SU capture parser remains unchanged. Packet admission, nonzero fragment semantics, frequency indices and the full activity-data importer still require work; body agreement alone is insufficient for training admission. See PUBLIC-BFI-DECODE.md.

The downloader pins URLs, revision, lengths, SHA-256 and dataset license; resumes only after rehashing completed files; retains failed partials; and bounds transfers and receipts. It does not execute downloaded code. Source and public packets stay in separate directories outside Git on ruvultra.

Validation improvements

An independent Codex review prompted two fixes: NPZ central-directory bounds are now checked before ZIP metadata allocation, and the GPU launcher verifies the expected seed, protocol, configuration and explicit absence of test metrics. The launcher requires Linux process groups for descendant cleanup. The learning suite passes; its Windows symlink-creation test was skipped locally and passed in the Linux downloader/decoder suite. No validation claim depends on that skip.

Coordination and next work

RuFlo tracks research, implementation and independent validation roles; existing Codex collaboration workers execute them. Cross-host federation is blocked by paused registration and relay membership rejection. Shared-memory writes are also blocked by the native-WAL guard. No guard was bypassed.

The hourly app automation continues tonight through 07:00 America/Toronto, then performs the selected frozen evaluation and pauses. The twelve control runs are complete and should not be repeated. Next priorities are repeatable local BFI coverage, public packet admission and a complete recording-quality census before activity modeling. See SOTA-BFI.md and BFI-INTEGRATION.md. A fair state-of-the-art comparison still requires matched data, representation, splits and independent evaluation.

Reproduce and audit

Source entry points: prepare.py, prepare_bfi.py, train.py, run_batch.py, static_baseline.py, and codex_iteration.py. Private source snapshots, exact commands, dataset/metadata hashes and checkpoint hashes are retained on ruvultra under /data/scratch/ruview-overnight-20260929.

• reports/current-comparison.json: all six neural runs and controls.
• reports/lstm-validation.png: validation curves and selected epochs.
• runs/seed-1729-recovery-verification.json: successful model receipt after fixing the supervisor's overly small results-file limit; original state kept.
• runs/static-control-01/dependency-receipt.json: static-control dependencies.
• data/bfi-1000-1-relative-time/receipt.json: updated real BFI import.
• reports/controls-comparison-v1.json: completed temporal diagnostics.
• runs/controls-matrix-v1-state.json: serial controller completion and batches.
• reports/ruview-bfi-opt-1000-*-traffic-coverage-v2.json: traffic intervals.
• Local public-bfi-oracle-v1/: frozen decoder, reproducer and oracle receipt.

Focused and existing BFI regression tests pass. Run them with:

python -B -m unittest discover -s tools/bfi_learning -v
python -B -m unittest discover -s tools/bfi -v

Raw samples, public subject data, checkpoints, logs and Codex transcripts remain outside Git. No hardware settings, existing GPU services or production models were changed during these offline experiments.