Sets up docs/research/sota-2026-05-22/ as the autonomous-research
output dir, with PROGRESS.md as the canonical 15-vector research
agenda spanning spatial intelligence, RF features, RSSI-only, and
exotic/long-horizon verticals. Cron d6e5c473 (*/10 * * * *) picks
threads from this file and self-terminates at 2026-05-22 08:00 ET.
First concrete contribution this tick — R5 subcarrier saliency:
* examples/research-sota/r5_subcarrier_saliency.py: pure-numpy port
of the count cog's Conv1d encoder + count head, computes per-
subcarrier input×gradient saliency via central-difference. 128
samples × 56 subcarriers × 2 forward passes/subcarrier ≈ ~3 s on
CPU, no GPU or framework dependency.
* docs/research/sota-2026-05-22/R5-subcarrier-saliency.md: research
note with motivation, method, novelty argument, and the first
measured ranking. Top-8 subcarriers for cog-person-count v0.0.2:
[41, 52, 30, 31, 10, 35, 2, 38]. Max/mean ratio 2.85x.
* v2/crates/cog-person-count/cog/artifacts/saliency.json: machine-
readable per-subcarrier saliency + top-K lists, so future-tick
experiments (retrain at K=8/16/32) consume it without re-running.
Key insight from the first measurement: top-8 saliency is *band-
spread* (indices span 2-52), not concentrated. This directly raises
R8's (RSSI-only) feasibility ceiling, because RSSI is a band-
aggregate — it retains the integral of a band-spread signal. First-
order estimate: RSSI-only should hit ~60% of full-CSI accuracy for
the count task. R7 (adversarial defence) inherits a concrete defender-
priority list: corroborate these 8 subcarriers across nodes.
This commit is the first of many short, focused contributions over
the next ~12 hours. PROGRESS.md is the canonical pointer for the
next tick to pick up the next thread.
* chore: stage v0.0.2 artifacts + temperature scalar for build pipeline
Stages count_v1.{safetensors,onnx,temperature,train_results.json}
ahead of the build/sign/upload step. This commit is a momentary
side-effect — the next commit will refresh the per-arch manifests
with the new binary SHAs once ruvultra finishes the cross-build.
The .temperature file holds the calibration scalar from LBFGS over the
held-out conf logits. The Rust cog will read it post-load and divide
conf_logits by it before sigmoid, exactly matching the Python eval.
* feat(cog-person-count): v0.0.2 — K-fold validated, label smoothing + early stop + temp scale
The v0.0.1 "65.1% but class-1=0%" result was an unlucky temporal split
that let a degenerate "always predict 0" classifier hit eval acc =
class-0 fraction. 5-fold stratified random CV proved the architecture
actually learns ~57.1% class-1 accuracy under fair splits — a real,
modestly useful signal.
v0.0.2 ships a retrained model that:
* **Splits randomly (seed=42) 80/20** instead of temporally — eliminates
the trailing-window-class-imbalance cheat.
* **Class-balanced sampler** (multinomial with replacement, weighted by
inverse class frequency) — per-batch expected counts are equal
regardless of dataset distribution.
* **Label smoothing 0.1** on the cross-entropy — reduces confidence
saturation that drove v0.0.1's all-or-nothing predictions.
* **Early stopping** with patience=20 — stops at epoch 29 instead of
overfitting through 400.
* **Temperature scaling** of the conf head — LBFGS fits a scalar T on
held-out conf logits; ships as a count_v1.temperature sidecar so the
Rust cog can divide conf_logits by T before sigmoid.
Numbers on the same data:
| Metric | v0.0.1 | v0.0.2 | K-fold (5x100) |
|------------------|--------|--------|----------------|
| Overall acc | 65.1% | 62.3% | 62.2% ± 1.9% |
| Class 0 acc | 100% | 86.2% | 67.4% |
| Class 1 acc | 0% | 34.3% | 57.1% ✓ |
| MAE | 0.349 | 0.377 | 0.378 |
| Spearman | 0.023 | 0.013 | 0.160 |
Class-1 accuracy 0 → 34.3% is the headline win. Net acc moves slightly
because we stopped cheating on class 0. K-fold's 57% says there's
headroom remaining; reaching it needs more independent splits (== more
data), not more training tricks.
Confidence calibration didn't move. Temperature scaling alone can't fix
a confidence head trained against a noisy argmax==truth indicator over
a 62%-accurate classifier — the head's training signal is the issue,
not its post-hoc transform. The honest fix is multi-room data (#645),
not another calibration knob.
Live on cognitum-v0 at /var/lib/cognitum/apps/person-count/ — health
reports candle-cpu backend, count = 1 (was 0 in v0.0.1) on synthetic
zero input.
Files changed:
* scripts/train-count.py — adds --k-fold (no sklearn dep, hand-rolled
stratified splits with deterministic shuffle) and --v2 paths.
* v2/.../cog/artifacts/count_v1.safetensors (392 KB, new sha
32996433…) + count_v1.onnx (16 KB) + count_v1.temperature (0.9262
scalar) + count_train_results.json (full epoch trace).
* v2/.../cog/artifacts/manifests/{arm,x86_64}/manifest.json bumped to
version 0.0.2 with the new weights_sha256 + caveats.
* docs/benchmarks/person-count-cog.md — appends a v0.0.2 section
with the K-fold diagnostic table and honest-read paragraph.
GCS:
gs://cognitum-apps/cogs/arm/cog-person-count-count_v1.safetensors
refreshed (binaries unchanged — load weights via mmap at runtime).
The arm + x86_64 manifests committed in #696 referenced the binaries
built before #697 wired the `run` subcommand. Rebuilt + re-signed +
re-uploaded to GCS, and re-deployed to cognitum-v0:
arm sha 15c2fbac…7728ea5 (3,807,456 B, up from 2,168,816 — added Tokio runtime)
x86_64 sha 051614ce…cc8388b3 (4,502,960 B, up from 2,615,528)
Both re-signed Ed25519 with COGNITUM_OWNER_SIGNING_KEY. Manifests
now match the binaries published at gs://cognitum-apps/cogs/{arm,
x86_64}/cog-person-count-* and the binary installed at
/var/lib/cognitum/apps/person-count/ on cognitum-v0.
Phase 3 of ADR-103. Cross-compiled aarch64 + x86_64 on ruvultra, signed
with COGNITUM_OWNER_SIGNING_KEY (Ed25519), uploaded to GCS, and live-
installed on the cognitum-v0 Pi 5 alongside cog-pose-estimation.
Real-hardware bench on cognitum-v0:
./cog-person-count-arm health
→ backend=candle-cpu, count=0, confidence=0.49, p95=[0,7]
30 sequential health invocations: 0.276 s → 9.2 ms/invocation cold
Compares to cog-pose-estimation's 8.4 ms — count cog is ~10% slower
because the dual-head (count softmax + confidence sigmoid) does ~2x
the work after the shared encoder.
GCS release artifacts (publicly downloadable, SHA-verified):
arm/cog-person-count-arm 2,168,816 B
sha: 36bc0bb0...0d47b507b3c3
sig: R/00xdzHriyr/2r...JK+a6k71NDg== (Ed25519)
x86_64/cog-person-count-x86_64 2,615,528 B
sha: 76cdd1ec...3923 7392b01db
sig: QB+8cnGSMQmu...ZtTNIQ2rDg== (Ed25519)
arm/cog-person-count-count_v1.safetensors 392,088 B
sha: dacb0551...e6e04ff56d15c3a65a9ff
Live install at /var/lib/cognitum/apps/person-count/ on cognitum-v0
matches the layout of every other installed cog (anomaly-detect,
seizure-detect, pose-estimation): cog-person-count-arm binary,
count_v1.safetensors weights, manifest.json, config.json.
Adds:
* v2/.../cog/artifacts/manifests/{arm,x86_64}/manifest.json — full
ADR-100 schema with all fields filled (sha + sig + size + URL +
build_metadata carrying the v0.0.1 honest training caveats).
* docs/benchmarks/person-count-cog.md — appends "Live appliance
install" and "Signed GCS release artifacts" sections to the
benchmark log.
Honest v0.0.1 caveat still applies (class-1 accuracy 0% on the held-
out tail of the single-session training data) — same data-bound
limit as pose_v1. The shipped artifact is the *vehicle*; production-
quality accuracy follows from multi-room paired data per ADR-103's
v0.2.0 plan + #645.
Phase 2 of ADR-103: trained count head on the existing 1,077 paired
samples (the same data that produced pose_v1 yesterday).
Honest result: 65.1% eval accuracy / 100% within ±1 / MAE 0.349 on
the held-out time-window. Per-class: 100% on "empty room" / 0% on
"1 person". The model overfit by epoch 100 (train_acc → 1.0,
eval_loss climbed 0.67 → 7.8) and the "best" checkpoint is the
snapshot that happened to predict the eval window's class
distribution (140/215 = 65.1%, matches eval_acc exactly). Confidence
head Spearman = 0.023 ⇒ uncalibrated. Same data-bound failure mode
as pose_v1 (#645), bounded by single-session training data; same
fix path (multi-room).
What v0.0.1 still validates end-to-end:
* PyTorch → safetensors → Candle Rust loads cleanly on first try.
`cog-person-count health` reports `backend: candle-cpu` and emits
real per-frame predictions instead of the stub backend's hard-coded
{1 person, 0 confidence}. Architecture parity between train-count.py
and src/inference.rs::CountNet is bit-exact.
* ONNX export bit-clean (16 KB, opset 18, dynamic batch axis).
* Training wall time: 5.6 s for 400 epochs on RTX 5080.
* Binary size unchanged (2.36 MB stripped), model loads via mmap at
runtime.
This commit ships:
* scripts/align-ground-truth.js: extended to emit n_persons_mode +
n_persons_max per window so the training pipeline has count
labels. Backwards-compatible (additive fields).
* scripts/train-count.py: new — mirrors CountNet architecture
exactly, loads paired.jsonl, trains 400 epochs with
CE+BCE+Brier loss, exports safetensors + ONNX + per-epoch JSON.
* v2/.../cog/artifacts/{count_v1.safetensors,count_v1.onnx,
count_train_results.json}: the trained artifacts.
* v2/.../cog/README.md: Status table updated with the v0.0.1 numbers
+ an Honest Caveat section explaining the data-bound result.
* docs/benchmarks/person-count-cog.md: new — full v0.0.1 benchmark
log mirroring the format docs/benchmarks/pose-estimation-cog.md
established. Includes comparison to ADR-103 v0.1.0 acceptance
gates and per-class breakdown.
Still pending:
* `run` subcommand wiring (long-running polling loop, same as pose)
* Cross-compile + sign + GCS upload (mirror of pose cog pipeline)
* Live install on cognitum-v0
* v0.2.0: re-train on multi-room data, LoRA per-room adapters,
Stoer-Wagner min-cut clip in fusion stage
rUv