Bump the harness and the umbrella (which pins it), point README and the
user guide at 0.9.1, sync the stale lockfile header (0.5.1, old bin name)
with package.json, and add a plain-language changelog entry for the
live sensing views in Claude Code (ADR-377, ADR-378).
Co-Authored-By: claude-flow <ruv@ruv.net>
- Stale frames shown as live: spectrumWith now keeps only nodes in the
latest capture, so a node that stops streaming (or a failed capture)
leaves the waterfall and the "no CSI frames" card returns.
- Multi-antenna ESP32 nodes never reached the waterfall: with --spectrum
the parser adds chain-0 amplitudes (ADR-018 antenna-major) in a separate
field the SpectrumCollector reads; the kernel path is unchanged.
- Frames from different CSI layouts could be joined: spectrum results
carry their shape, and frames merge only when shape and bins match.
- Animation ran on assumed time: frames read the real clock (the one the
full render uses), so pulses keep the reported rates and the replay
keeps up when a frame fires late; replay lag is pruned with the nodes.
Tests: regression tests for each; npm test 193 pass, test:security 91
pass, engine tests 4 pass, validate passes.
Co-Authored-By: claude-flow <ruv@ruv.net>
The /ruview pane gains three views (keys 1/2/3) drawn in terminal Rasters
and animated in place with $.ui.blit (~12 fps, no redraws):
- CSI waterfall: per-subcarrier amplitude over time from the new
`esp32 --spectrum` capture, two frames per row via the half block,
5-95th percentile colour limits, replayed at the frames' arrival rate,
labelled MEASURED or SYNTHETIC from the packets' own flag; n cycles nodes.
- Radar: a 120-degree range fan with rings, a fading distance trail and a
sonar ping to the measured range; braille heart/breathing/distance charts
scaled to the data; a heart and breathing gauge that pulse at the
device-reported rates ("a metronome, not a waveform").
- A shimmer rule under the title; surfaces without Raster get a note.
Harness: captureEsp32 / ruview_esp32_capture take spectrum,
spectrum_bins (8-128) and spectrum_frames (8-256): a bounded ring of
binned amplitude frames per node, dominant shape, rate and synthetic flag.
The mod is split into model/raster/anim/views/register (own-file imports
only). Tests: 11 showcase + 11 mod Node tests, a --spectrum loopback test,
4 engine tests (waterfall tab sends --spectrum, draws the Raster, blits);
npm test 189 pass, test:security 90 pass, manifest verified, validate passes.
Evidence: MEASURED on an ESP32-C6 (COM11, release c6-adr110 v0.7.0, node 6)
streaming 7-18 Hz CSI to this host; a 1x256 capture renders the occupied
band against guard/null tones. Unpacked budget 384 -> 448 KiB (+~37 KB).
Co-Authored-By: claude-flow <ruv@ruv.net>
- mod/: a Claude Code mod (function hooks). /ruview opens a pane beside
the transcript with CSI nodes (rate, loss, RSSI, shape, SYNTHETIC) and an
optional ESPHome radar (presence, distance, device-reported vitals),
alerts (heartbeat-only boards + buffer-starvation hint), Refresh/Close
buttons and a one-line status; refreshed on $.clock.every.
- Reads only through the harness CLI beside it (fixed read-only argv,
--json) located via $.plugin.root (mods may import nothing else).
- Packaged in @ruvnet/ruview (mod/.claude-plugin, mod/hooks), listed in the
repo marketplace as ruview-live@ruview; `ruview mod` prints how to load it.
- @ruvnet/ruview 0.9.0; umbrella ruview 0.9.0 (still unpublished).
Verified: claude plugin validate (mod + marketplace); claude plugin test
2/2 x3 runs; harness 172 tests + 89 security; real Claude Code headless
`/ruview refresh` from an installed tarball and with the live radar kit.
Co-Authored-By: claude-flow <ruv@ruv.net>
- harness/ruview-umbrella -> npm `ruview` 0.8.0: one CLI and one MCP server
for the operator harness, Homecore and the vitals kernel. `ruview homecore
...`, `ruview kernel <doctor|info|selftest|parity|synth|analyze|bench>`,
`ruview capabilities`; everything else delegates to @ruvnet/ruview. MCP
serves ruview_* + homecore_* tools and the ui:// console (stdio + HTTP).
- @ruvnet/ruview 0.8.0: bin renamed `ruview-harness` (the umbrella owns
`ruview`; a clean tarball install showed the bin collision). New exports
(cli, cli-ui, mcp, mcp-http, ui, esphome, kernel); injectable MCP handler;
kernel as optional peer + setKernelImporter so the kernel resolves from
the umbrella's tree under linked/pnpm installs.
- mmwave --source esphome: dependency-free read-only ESPHome native API
client (private hosts only; encrypted/password APIs reported). Live Seeed
MR60BHA2 kit: presence, distance, device-reported HR/BR (MEASURED).
- monitor detects Ameba CSI buffer starvation; radar views in terminal UI
and console widget.
- homecore 0.1.0 and @ruvnet/ruview-kernel 0.1.0 gain ./cli exports.
Gates: harness 166 + 89 security, homecore 27 + 23, kernel 21 + 12,
umbrella 4; clean-dir tarball install smoke incl. live radar over HTTP MCP.
Local publish is a one-off exception to ADR-265 D2, recorded in ADR-376.
Co-Authored-By: claude-flow <ruv@ruv.net>
Live RTL8721Dx capture (node 3, COM10) reported 97% loss: one stray
sequence value (+52,835, then straight back) was counted as 52,834 lost
frames, and every late UDP packet re-counted its gap. The tracker now
treats a single out-of-window value as a stray, needs two agreeing frames
to accept a counter reset, and lets a late packet fill its own gap
(seqReordered / seqStrays / seqResyncs in the summary). Against ground
truth on the live board (unique sequence numbers over the span): 4,783 of
4,783 present, loss 0, 9 strays; the tracker matched both.
ADR-373 Amendment 2 records the live Realtek run: devices -> realtek,
7,404/7,404 packets decoded, monitor at 1500000 baud without a reset, and
12,549 live frames analyzed by the WASM kernel.
Co-Authored-By: claude-flow <ruv@ruv.net>
The flash flow's boot-log capture decoded serial lines and wrote text to a
stdout that the scrubbed child env leaves on cp1252. The ESP32-C6 release
firmware logs U+2192, so the capture died mid-boot (bootLog.captured=false);
an earlier abort would have hidden the CSI evidence and reported a good
flash as unvalidated. Write raw bytes instead. Regression test forces
PYTHONIOENCODING=cp1252 through a stand-in serial module (fails without
the fix). ADR-373 Amendment 2 records the real C6 flash evidence.
Co-Authored-By: claude-flow <ruv@ruv.net>
Findings from a hardware run of this branch against an RTL8721Dx board
(PL2303GC) and an ESP32-C6 (CP210x) on Windows 11:
- esp32 capture decodes Realtek RAC1 (ADR-323, CRC-verified, 8/16-bit
tones) and RHB1 heartbeats, ADR-110 sync and ADR-081 mesh packets, and
reports every CSI shape a node sends.
- Capture fails (no_decodable_packets / heartbeat_only) instead of
returning ok:true + MEASURED when nothing decodes; names
heartbeat-only senders.
- --analyze runs live single-antenna CSI through @ruvnet/ruview-kernel at
the measured frame rate with a bounded buffer.
- monitor opens the port with DTR/RTS deasserted (no node reboot), takes
--baud, and explains the C6/S3 USB-Serial/JTAG console.
- devices classifies Prolific 067B:23A3 as a Realtek board.
- Windows: owner-only ACL on the hosts file (fails closed); symlink test
skips only where Windows refuses symlinks.
- kernel build names uninitialized submodules when cargo fails.
- Parser ~3.2x faster (MEASURED, 0.42-0.52M -> 1.33-1.67M packets/s).
Co-Authored-By: claude-flow <ruv@ruv.net>
CI fixes (both failures on PR #2062 were in this branch):
- harness/ruview: remove `2.11.0` literals from training remedies; the
ADR-265 D3 version-literal gate rejects them.
- ruview-kernel workflow: install wasm32 for the toolchain pinned by
v2/rust-toolchain (the action only added it to `stable`); build.mjs now
fails fast with the exact `rustup target add` remedy.
Kernel optimization (ADR-368 amendment):
- New `analyze_flat` / `session_push_flat` ops via `call_f64`: WASM
`rvk_call_f64` (alignment-agnostic bytes) and napi `callF64(Float64Array)`.
`analyze()`/`session.push()` pack uniform frames automatically; irregular
input keeps the JSON path and its per-frame errors. `analyzeFlat()` is the
zero-copy entry. Shared `Session::step` makes both paths byte-identical on
identical f64 inputs (Rust test), and the binary path avoids serde_json's
non-correctly-rounded float parsing.
- MEASURED (`ruview-kernel bench --seconds 120`, one Linux x64 host):
2.82x (wasm) and 2.45x (napi) faster than the JSON transport.
- Profiled remaining cost (examples/profile_stages.rs): heart-rate ACF in
wifi-densepose-vitals dominates; left for a separate, output-reviewed
change. Rejected WASM +simd128 and opt-level=s after three measured runs
(bench/wasm-variants.mjs).
Device e2e (ADR-373 amendment):
- test/e2e/devices.e2e.mjs + `npm run test:e2e:devices` + new workflow
ruview-device-e2e.yml: pty-emulated MR60BHA2 and RPLIDAR through real
pyserial, a UDP ESP32 node, and the real iPhone relay, via the CLI.
- mmWave auto-detect keeps probe frames and reads only the remainder;
iPhone LiDAR gains max_frames and returns immediately when refused.
Suite wall time 19.6 s -> 5.6 s (MEASURED, dev container).
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_015H55jmprfz4oHWXSFbX5rf
@ruvnet/ruview 0.7.0 gives CLI, MCP, and SDK access to RuView sensing
hardware on the machine it is attached to, and to other machines over SSH.
Devices (ADR-373, src/devices/):
- ruview_devices_scan: pyserial port list classified by USB VID:PID
(Espressif native USB, CP210x, CH34x, FTDI) into ESP32 / mmWave / RPLIDAR
roles with the command that confirms each.
- ruview_esp32_capture: receive-only UDP listener parsing ADR-018 CSI
frames, vitals, and the other 0xC511000x packets per node: rate,
sequence loss, RSSI, subcarriers, frequency, device vitals.
- ruview_mmwave_read: MR60BHA2 (60 GHz) and LD2410 (24 GHz) parsers ported
byte-for-byte from firmware mmwave_sensor.c; baud/frame auto-detect.
- ruview_lidar_read: RPLIDAR SCAN over serial (start/stop, DTR best-effort)
and the iPhone LiDAR relay as a WebSocket client (token only via
RUVIEW_LIDAR_TOKEN; statistics only, never raw depth).
- Dependency-free serial via a fixed pyserial pump (argv-only, bounded,
hex output immune to secret redaction). Schema-level port pattern
validation added to the policy validator.
- All device tools require the new `device-access` MCP grant.
Remote hosts (ADR-374, src/remote.js):
- ~/.config/ruview/hosts.json (0600), CLI-only `hosts add`; strict
name/target validation against ssh option and shell injection.
- `ssh -o BatchMode=yes -o StrictHostKeyChecking=yes -- <target>` running a
version-pinned `ruview call <tool> --read-only`, every token single-quoted;
read-only enforced locally and remotely; no mutations, credentialed reads,
or host hopping. MCP needs `remote-host` plus the remote tool's grant.
Also: device-access skill, doctor `devices`/`remote` groups, SDK
`devices.*`/`hosts.*` with types, generic `ruview call` verb, guidance
capability, 320 KiB package budget, CI smoke expects 22 tools.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_015H55jmprfz4oHWXSFbX5rf
@ruvnet/ruview 0.6.0 turns the harness into a complete npm operator
surface: one policy-checked tool registry behind the CLI, the MCP server,
and a new typed SDK.
- SDK (ADR-369): `@ruvnet/ruview/sdk` createRuView() with doctor,
firmware.{ports,plan,flash,monitor}, training.{plan,run,gate,calibrate},
kernel.{selfTest,load}, guidance/claimCheck/verify/spaces; strict mode
throws RuViewError; TypeScript declarations in src/sdk.d.ts.
- Firmware (ADR-370): cross-platform `python -m esptool` flashing replaces
the Windows-only stub. SHA256SUMS verification (mismatch always fatal),
bundle containment, variant/partition checks, argv-only command, chip_id
preflight that refuses a mismatched chip, boot-log capture;
hardwareValidated only when CSI callbacks appear without a panic.
Read-only ruview_firmware_plan / ruview_firmware_ports.
- Training (ADR-371): ruview_train / ruview_train_plan run pose-smoke,
pose (MM-Fi), or room training from the trusted checkout with in-repo
paths only, logs/checkpoints under v2/target/ruview-train, and libtorch
remedies. ruview_train_gate enforces mean-pose baseline, leakage-free
splits, subject/time disjointness, and evidence tagging.
- Doctor (ADR-372): structured pass/warn/fail/skip checks with remedies
across runtime, integrity, hosts, repo, Rust, Python/esptool, firmware
bundles, serial, sensing server, and the compute kernel; --json/--group;
hardware probe and network checks are CLI/SDK-only.
- Plan tools cannot take `confirm`; acting tools keep grant+confirm over
MCP. Policy files and TOOL_POLICY are drift-tested.
- Unpacked budget 160 KiB → 224 KiB in both npm workflows; CI tarball smoke
now checks the SDK and the training gate.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_015H55jmprfz4oHWXSFbX5rf
Adds a portable RuView compute kernel for JavaScript, built from the Rust
source of truth and wired into the @ruvnet/ruview metaharness.
Rust
- v2/crates/ruview-kernel: bounded single-function JSON ABI
(call(op, json) -> envelope) over the ADR-021 wifi-densepose-vitals
pipeline: streaming sessions, batch analyze, deterministic SplitMix64
synthetic CSI (SYNTHETIC), bounded session registry, strict input
validation (deny_unknown_fields, width/finite/magnitude checks, size and
count limits). wasm32-only zero-import C ABI (rvk_alloc/free/call).
- v2/crates/ruview-kernel-napi: napi-rs transport exposing only
abiVersion()/call(); standalone workspace (excluded from v2) so napi never
enters `cargo test --workspace`; panic=unwind + catch_unwind.
npm (harness/ruview-kernel, zero runtime deps)
- loadKernel({backend: wasm|napi|auto}); WASM default (no host authority),
napi never silently falls back, auto reports fallbackReason.
- SHA256SUMS integrity verification; mismatch is always fatal; loader
rejects any WASM module that imports host functions.
- Browser-safe ./wasm entry; CLI doctor/info/selftest/parity/synth/
analyze/bench; build.mjs builds artifacts from source (no shell).
- Unit, security, and e2e tests (CLI round trip, cross-backend parity).
Metaharness
- New read-only tool ruview_kernel_selftest / `ruview kernel`: imports the
fixed optional package, fails closed with kernel_not_installed; policy,
claims, MCP policy, guidance capability, and manifest updated.
CI: .github/workflows/ruview-kernel-npm.yml builds both transports on
Linux/macOS/Windows, requires napi in tests, runs parity, verifies
artifacts, packs, and runs the cross-package e2e. No publish step.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_015H55jmprfz4oHWXSFbX5rf
* feat(firmware): add the 16MB flash profile the C6 fleet runs on
The 4MB and display layouts have 1.875MB and 2MB OTA slots, sized for 4MB
and 8MB parts. A 16MB C6 (DevKitC-1 N16) has no layout in the tree at all,
so building for one meant hand-editing sdkconfig -- and a smaller table
written over a deployed 16MB node relocates NVS and destroys its
provisioning, which is why this belongs in the repo rather than in a
workstation's untracked sdkconfig.
Two 4MB OTA slots, plus a 64K coredump partition and an 8000K FAT volume
for on-node storage. Rollback is enabled so an OTA'd image that cannot
confirm itself is reverted by the bootloader.
Layered, not inherited: sdkconfig.defaults.esp32c6 keeps its 4MB layout so
the common 4MB devkit is never silently repartitioned.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01PVWMiHQifoYXL7uL3bphrZ
* feat(firmware): persistent on-node log and coredump-to-flash
A remote fault currently leaves no post-mortem. main.c decodes the reset
reason at boot and logs it to the CONSOLE -- which is exactly what a power
cycle destroys, and a power cycle is how a wedged node gets recovered.
The flash was already provisioned for this and the features were simply off:
partitions_16mb.csv has reserved a 64K `coredump` partition and an 8000K
`storage` FAT volume since the 16MB layout was written, neither had ever been
mounted or written, and the build defaulted to CONFIG_ESP_COREDUMP_ENABLE_TO_NONE.
- node_log.c/h: a 128 B-record ring on the storage volume. Boot records
(reset reason + the previous session's final uptime + a health snapshot),
periodic health samples, and event records. Ordering is boot_id + uptime_s
because these boards have no RTC and nothing can be timestamped absolutely.
Counters are stored RAW and differenced at read time, so a cadence change
never invalidates old records. Retention, not wear, binds: 128 B every
300 s is ~208 days in 8 MB, and the write rate is bounded explicitly.
- ap_rssi_dbm, the association link from esp_wifi_sta_get_ap_info. The
adaptive controller's rssi_median_dbm is a CSI statistic, not the uplink,
so a node with a bad uplink currently shows no signal for it anywhere.
- CONFIG_ESP_COREDUMP_ENABLE_TO_FLASH, complementary rather than a
substitute: a coredump catches a CRASH, the boot record catches a HANG
followed by a watchdog reset.
node_log_init failing is non-fatal by design -- the module disables itself
and every entry point becomes a no-op. A logging volume that will not mount
must never take the fleet down with it.
csi_collector gains two gate accessors. This build selects its frame gate at
compile time, so both return CSI_GATE_NOT_CONFIGURABLE rather than a
plausible number: the log keeps a field of fixed width and meaning either
way, and an explicit "not applicable" is stored instead of something a
reader would later reason from as a real setting.
Built with ESP-IDF v5.4 for esp32c6 using
defaults + .esp32c6 + .16mb. Verified from the resulting artifacts: 16MB
table with coredump and storage present, COREDUMP_ENABLE_TO_FLASH set,
rollback enabled, image 0x10abd0 bytes in a 4M slot. NOT flashed to hardware
from this branch -- the equivalent code was exercised on a node-3 pilot, but
that is not evidence for this build.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01PVWMiHQifoYXL7uL3bphrZ
* docs(adr-362): reading the on-node log is on-demand, never streamed
The log engine lands with no consumer, so the access pattern was open. Records
the decision to pull a node's log only when a human asks -- after a hang or a
crash, or while troubleshooting that node -- and to keep nothing server-side.
The reasoning that decided it: the ring already survives a power cycle on the
node (~208 days of retention), so a server copy duplicates a durability
property that is already met; and airtime is this fleet's binding constraint,
measured at 91% consumed by two slow nodes, so a periodic drain across nine
boards would spend exactly the resource whose exhaustion you would open the
log to investigate. An instrument must not perturb what it measures.
Leaves the decode-where question open on purpose, with the argument for a raw
proxy stated: decoding server-side defines the record layout in both C and
Rust, and the pilot already paid once for a host/firmware disagreement about a
field's meaning.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01PVWMiHQifoYXL7uL3bphrZ
---------
Co-authored-by: Joe <joe@GitDev.hedgehog>
Co-authored-by: claude-flow <ruv@ruv.net>
A-B-A capture from the overnight run, ten 30 s windows per arm: pairing
fraction rises 0.5668 -> 0.7287 (Welch t=13.26, p<0.0001) when the fleet is
filtered to one transmitter, but paired observations/s falls 44% because
tx/s falls 56%. Fraction was the wrong figure of merit; not adopted.
Firmware changes (persistent on-node log, live filter_mac, and the node-3
pilot fixes) that produced this capture are held back pending a port against
main's restructured firmware/main/ tree (config_api.c and ota_update.c are
gone there).
Claude-Session: https://claude.ai/code/session_01PVWMiHQifoYXL7uL3bphrZ
Co-authored-by: Joe <joe@GitDev.hedgehog>
Co-authored-by: claude-flow <ruv@ruv.net>
scripts/train-ruvllm.js ran two separate computations during contrastive
pretraining: ContrastiveTrainer.train() (a baseline pass that computes loss
but never updates the encoder) and a hand-written loop that actually applies
gradient updates to encoder.w2. The exported training-metrics.json took
lossHistory and contrastive.durationMs from the first (no-op) pass instead
of the second (real) one, so the published curve was flat/meaningless
(~0.1352 for all 20 epochs) and the duration was ~30x too short.
This did not affect finalLoss/improvement or the held-out temporal-triplet
accuracy metric published for this model on HuggingFace
(ruvnet/wifi-densepose-pretrained) -- those already came from the real loop.
It only affected the diagnostic curve/duration anyone auditing convergence
from the published JSON would see.
Also: config.json's training.steps field (epochs * triplets, technically
consistent but reads like an optimizer-step count) is now reported as
separate epochs/tripletUpdatesPerEpoch fields, and the exported version
bumped 1.0.0 -> 2.0.1 to match the "v2" naming already used elsewhere for
this file set.
Verified by re-running against the real capture
(data/recordings/overnight-1775217646.csi.jsonl): lossHistory now shows a
genuine decreasing curve (0.1037 -> 0.0654 across 20 epochs, matching the
already-published finalLoss exactly), durationMs reports the real ~40s
gradient-loop time instead of ~12s.
Model card updated with the same disclosure (docs/huggingface/MODEL_CARD.md)
plus a real, previously-undisclosed EWC forgetting-rate figure (32.97%,
present in training-metrics.json but not surfaced in the human-readable
card) and a corrected repro-script pointer.
Records why device-free position kept failing here, measured rather than
argued, and the join key built in response. The implementing PRs land without
this and are harder to read for it.
**Phase is unusable on this hardware.** The bistatic tier needed a signed
per-link radial velocity, which requires the common-mode component of CSI
phase. Measured across three ESP32-C6 nodes over ~40,000 frames, the
frame-to-frame change in common-mode phase is statistically indistinguishable
from uniform random: resultant length R = 0.011-0.032 against a 0.000
reference, sd 1.78-1.81 against pi/sqrt(3) = 1.814, no lag-1 autocorrelation.
A second, unwrap-free estimator agreed.
The mechanism is packet-detection timing quantization. Common-mode phase is
-2*pi*f_c*tau, and one ADC sample of jitter at 20 MHz is about 758 rad -- some
120 full wraps, re-randomized every packet. Not calibratable. The bistatic tier
is default-off as a result.
**Amplitude works but had no parallax.** The three AP->node links leave a single
distant transmitter inside a 36.5 degree fan; scalar per-node energy weighting
closed with a clean negative result -- magnitude responds to proximity,
direction does not.
What follows from both: a position tier needs several transmitters seen from
several receivers, and fusing those requires knowing that two nodes observed
the SAME transmission. Arrival timestamps cannot say so -- the sink sees network
jitter, not capture time -- and the per-node sequence counter is private to each
node. Hence (tx_mac, rx_seq): the 802.11 sequence control field is assigned by
the transmitter and is identical at every receiver, so it names one
transmission without any coordination between receivers, no shared clock and no
negotiation.
Status is Proposed, deliberately. The measurement layer is implemented and
validated on hardware; localization is NOT built, and the ADR says so along
with the rest of its known limitations -- peer links are ~10x weaker than data
frames, three links is below the threshold for tomographic localization
regardless of quality, and no position tier consumes the link table.
References only ADR-018, ADR-060 and ADR-110, all already present here.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BHNNWbmKvxXrVaEDwXEBNo
The mesh gate uses mesh-time buckets only while c6_sync_espnow_is_valid(); an
unsynced node falls back to the elapsed gate. Three of nine nodes never
synchronised during these runs, so for those three the mesh and mesh+seq arms
silently ran the elapsed gate.
+0.5 pp is therefore not a clean fleet-wide measurement of window alignment.
Reported as observed rather than reconstructed: the arms were assigned before
we understood the fallback, and a post-hoc per-node reanalysis would not be the
experiment we ran.
Names the cause as environmental rather than a defect. Sync is leader-elected
and hub-and-spoke, so whether every node hears the leader is a property of the
building -- and a house that fragments into sync domains degrades any
shared-time option quietly, for exactly the nodes whose data you would most
want.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01BHNNWbmKvxXrVaEDwXEBNo
How a node rate-limits CSI callbacks decides whether cross-node fusion works at
all, because it decides whether two nodes keep the SAME frames or merely the
same NUMBER of frames. We measured three gates on a nine-node ESP32-C6 fleet as
a 2x2 factorial, three Latin-square blocks, every arm verified live on every
node.
Selecting frames by 802.11 rx_seq roughly doubled cross-node pairing over the
elapsed-time gate: 0.5605 against 0.2434, +31.7 pp, reproduced in all three
blocks at ~70x the measured drift floor.
Two results we did not want and are reporting anyway:
- Mesh-time alignment -- a gate we built and expected to be the answer -- bought
+0.5 pp, about one standard deviation of noise. Indistinguishable from doing
nothing.
- We predicted rx_seq selection would win most on node pairs spanning a broken
sync domain. Per-pair yield says the opposite: never-synced nodes gained
LEAST. So it is not a workaround for broken time sync, and repairing our mesh
split would probably not have closed the gap.
The document leads with the method rather than the numbers, because the numbers
are one house. It records the three mistakes that made an earlier attempt
unreadable -- comparing different node subsets, not verifying the gate was
actually live, and using frame rate as the check that a gate is working when a
bucket gate makes frame rate nearly insensitive to it -- and the null run that
establishes the noise floor without which none of the contrasts can be judged.
run_arms.py is included so a reader can measure their own environment instead
of trusting ours. That is the point of the contribution; our figures are the
worked example.
Stated as a trade, not a win: pairing is bought with raw frame rate. On-device
edge processing was unaffected here (separately rate-limited), so the cost fell
on the uplink -- but whether that is acceptable depends on what consumes the
stream.
Co-Authored-By: claude-flow <ruv@ruv.net>
Points v2/crates/ruforecast at ruvnet/RuForecast@a802067f, the tip of
fix/governance-and-spend-enforcement (ruvnet/RuForecast PR #2), which closes
14 code-review findings against the training stack merged as #1766 -- most
critically, ADR-349's governance-verification and spend-approval-record
requirements were specified but never actually enforced by the fal.ai
hosted path; they are now.
Adds ADR-354 recording this as a RuView-level integration decision and
pointing to the submodule's own ADR-350 for the full technical record,
since that repository is now the source of truth for the forecast
subsystem's decision history.
Depends on ruvnet/RuForecast#2 merging with green CI first; see that PR for
status before merging this one.
Trains a real fal.ai LoRA fine-tune (fal-ai/flux-lora-fast-training) on
frames extracted from the real fal.ai-generated room-visualization MP4s
already captured this session, so display-layer generation has a specific
trained style target instead of prompt engineering against a stock model.
Architecture boundary: this model only ever produces supplementary
synthetic visual/pixel output. It has no path into measured/derived RF,
pose, vitals, identity, or confidence values, which remain local and
authoritative.
Real run: request 01a060b2-c940-7972-abf9-ea48847eb2ae, 1000 steps, 27
training images, ~$2.04 spend, validated with a real inference call
against the trained LoRA (assets/fal-visual-teacher-sample.png).
Chronos-Bolt (base/small), Chronos-T5-small, and TimesFM-2.5-200M run
zero-shot (no fine-tuning) on the same BIDMC cross-entity test windows,
baselines, and WQL/MAE metrics as the prior from-scratch RuForecast
BIDMC result. Unlike every from-scratch RuForecast configuration tried
this session, all four pretrained checkpoints beat both last-value and
seasonal-naive baselines on WQL, on both independent patient-partition
judges (12-42% margin), consistent with the massive-pretraining-corpus
hypothesis from the prior literature pass.
* feat(forecast): add independent Rust training stack
Signed-off-by: Codex <codex@openai.com>
* docs(forecast): record informal HPO exploration note (unaccepted)
Records an exploratory Darwin Mode numeric-genome hyperparameter search
against the tiny_ci synthetic accuracy protocol: WQL improved 0.257 -> 0.161
-> 0.153 over two search rounds, staying ahead of both baselines throughout.
Explicitly scoped as informal/exploratory (not a frozen leakage-free report,
not eligible for the evidence ledger, not validated for large_linux or any
real dataset) per this doc's own evidence-tagging discipline.
* feat(forecast-core): add per-horizon weighted quantile loss
weighted_quantile_loss collapses the whole horizon into one aggregate
number, hiding whether error grows with lead time. Add a sibling
function using the identical per-cell pinball formula and domain
checks, just reduced per horizon step instead of globally. A unit
test reconstructs the aggregate from the per-step numerators/
denominators to prove the two functions agree exactly, not just
approximately.
* feat(forecast-model): add evaluation-only local activation
Adds activate_for_evaluation and build_eval_input, gated behind the
cpu feature. activate_for_evaluation self-signs an unsigned candidate
with a fixed, publicly-known, non-secret Ed25519 key so a CLI operator
can run inference against their own just-trained candidate without a
real release signature -- explicitly never a production trust path,
documented as such in the module doc comment. The schema-digest check
in ArtifactActivationPolicy still applies in full; only signing is
relaxed. build_eval_input constructs a single-window ModelInput for
CPU inference from raw context-major values/mask arrays, matching the
training batch builders exact time-feature/descriptor encoding.
Six unit tests cover: malformed-candidate rejection, digest-mismatch
rejection (proving the schema check has teeth even on this relaxed
path), getting past the policy gate with a matching digest, and
eval-input shape validation.
* feat(forecast-train): add evaluate and prepare-synthetic-dataset CLI
Turns this sessions throwaway HPO scripts into real, tested CLI
surfaces, replacing two example binaries that bypassed the CLI/TOML
training path entirely.
prepare-synthetic-dataset generates a larger, configurable synthetic
training shard (default 24 windows) plus a matching train-local.toml
with the given OptimizerSpec hyperparameters baked in, and a separate
held-out test.jsonl -- the same synthetic-only, local-validation-only
posture as prepare-local-example, just bigger and configurable.
evaluate scores a trained (unsigned) candidate against held-out
windows and the LastValue/SeasonalNaive baselines: overall weighted
quantile loss, a per-horizon breakdown, 80% interval coverage, and
missingness, for the model and both baselines. Explicitly lists what
docs/benchmarks/ruforecast.mds full accuracy protocol additionally
asks for that this does not cover (abstention coverage, selective
risk, site/device slices, interference regime, RuVector-retrieval
ablation) rather than silently omitting them -- each needs
infrastructure a single-entity synthetic fixture does not have.
Together the three real commands (prepare-synthetic-dataset,
train-local, evaluate) reproduce the exact WQL numbers the old
library-bypassing example scripts produced for the same
hyperparameters, confirmed by re-running the harness/ruview HPO
dry-run end to end against the new CLI path.
New integration test trains a real candidate via smoke, evaluates it
against held-out synthetic windows, and asserts the full report shape
and an empty-input rejection.
* feat(forecast-train): add prepare-synthetic-dataset --seed + Autogenous bridge
--seed (default 0, backward compatible) makes prepare-synthetic-dataset's
generator deterministic per offset: same seed -> byte-identical corpus
across runs (needed to train two genomes on the SAME corpus for a fair
comparison); different seeds -> genuinely independent corpora (needed for
honest multi-seed/multi-judge evaluation upstream, both in
harness/ruview/flywheel/ruforecast/gate.mjs's fitness function and in the
new v2/crates/ruforecast-autogenous-bridge crate added here).
ruforecast-autogenous-bridge is a LOCAL-DEV-ONLY (excluded from the v2
workspace, publish = false) bridge from this crate's real evaluate CLI
into ruvnet/autogenous's new regression-candidate promotion path (separate,
unpushed branch feat/regression-candidate-kind): runs N independent
train+evaluate judges on genuinely distinct synthetic corpora, signs
receipts, and gets a real cryptographically-verified PROMOTE/REJECT
decision -- defense in depth on top of, not a replacement for, Darwin's own
promotion gate. Path-depends on a sibling autogenous checkout that does not
exist in CI; build/run directly via `cargo build --manifest-path
crates/ruforecast-autogenous-bridge/Cargo.toml`.
cargo test -p ruview-forecast-train --no-default-features --features
cpu,cli: all green (unaffected existing tests + this backward-compatible
addition).
* docs(forecast): retract round-2 HPO result, record honest multi-seed finding
The earlier "Informal HPO exploration note" claimed a real improvement
(WQL 0.257 -> 0.153) from a 3-round hyperparameter search. That search
evaluated every candidate against one fixed synthetic corpus (seed 0) for
every round -- textbook overfitting. Independent verification against two
fresh corpora (via the new ruforecast-autogenous-bridge crate) showed the
"winner" losing to the baseline on both.
Fixed the root cause in harness/ruview/flywheel/ruforecast/gate.mjs:
candidates are now scored against three independent corpora, worst-case
across them, not one fixed corpus.
Re-ran the search under the fix. It found a new winner that genuinely beat
baseline on all three of its own search seeds -- and that winner ALSO lost
independent verification on fresh seeds. Two independent search rounds,
pre- and post-fix, both produced an illusory "winner." The honest reading:
at this dataset scale (24 synthetic windows) held-out WQL varies enormously
by which corpus is drawn, regardless of hyperparameters -- confirmed
directly by the baseline genome's own primary swinging from 0.83 to a full
regression across the three fixed search seeds with unchanged
hyperparameters. No RuForecast hyperparameter configuration has been shown
to reliably beat the trivial baselines out-of-sample at this scale.
Also fixes a real design inconsistency surfaced by this exploration: the
Autogenous regression-candidate promotion verifier
(envelope::regression::verify_regression_promotion, separate unpushed
ruvnet/autogenous branch feat/regression-candidate-kind) required all
judges to share one corpus_id, which conflicts with this kind's
intentional cross-corpus judge design. Corrected there (commit bfa4c48);
did not change either REJECT verdict, which were already driven by the
real NotBetterThanParent signal on their own.
Append-only: the original round-2 row is kept, read together with this
new amendment section, per this doc's own evidence-ledger discipline.
* refactor(forecast): extract RuForecast to ruvnet/RuForecast submodule
Mirrors the v2/crates/worldgraph pattern: RuForecast becomes its own
independent public repo/workspace (ruforecast-core/model/train), mounted
at v2/crates/ruforecast as a git submodule, with v2/Cargo.toml path-depping
into its sub-crates and excluding the submodule from the v2 workspace.
Full v2 --workspace check passes; the ruforecast CLI builds and runs
correctly from its new location. Real git history for the extracted
crates (7 commits) was preserved via git-filter-repo, not squashed.
* feat(autogenous-bridge): real-data independent verification tool
real_data_verify.rs: signs two real, temporally-independent judge
measurements (different train/test split boundaries on the same real
household vitals corpus, not synthetic seeds) through the real
Autogenous regression-candidate promotion path, and gets a genuine
cryptographically-checked PROMOTE/REJECT verdict.
Result recorded (see docs/benchmarks/ruforecast.md): REJECT. One judge
nominally beat the trivial baseline but by less than the 0.01
non-inferiority margin; the other judge lost outright. Consistent with
every synthetic-data search this session -- no configuration has yet
been shown to reliably beat trivial baselines out-of-sample, now
including a real 6390-sample household corpus.
Bumps the ruforecast submodule to pick up the evaluate real-gap fix.
* docs(forecast): record real-household-data result (REJECT, signed)
6390 real vital-signs samples, two independent temporal splits,
independently verified via the real Autogenous regression-candidate
promotion path. Signed verdict: REJECT -- neither split cleared the
non-inferiority margin. Same conclusion as every synthetic search this
session, now confirmed with real data too: more real data is the
credible next lever, not further search on this scale of fixture.
* chore(forecast): point ruforecast submodule at published main (post gap-tolerance fix)
* ci(forecast): fix ruforecast-ci.yml for the new v2/crates/ruforecast submodule layout
The RuForecast crates were extracted into a standalone submodule (v2/crates/ruforecast, mirroring the existing v2/crates/worldgraph pattern) and are no longer members of the v2 workspace. The CI workflow still referenced the old in-tree package names/paths (ruview-forecast-core/model/train under v2/), which broke every forecast job with "cannot specify features for packages outside of workspace".
Fixed every job to target the submodule: working-directory -> v2/crates/ruforecast, -p ruview-forecast-* -> -p ruforecast-*, all hardcoded Cargo.toml/source paths in the Python assertion and clean-room-scan blocks, the Swatinem/rust-cache workspaces input, the artifact-tree output paths, and the two levels of relative path (../ -> ../../) in the informational-benchmark steps whose working directory moved one level deeper.
Also collapsed the paths: trigger filters internal-crate globs (v2/crates/ruview-forecast-*/**) down to the single v2/crates/ruforecast path, since GitHub Actions path filters only ever see a submodule gitlink change as one entry in the parent tree, never its internal file paths -- the old globs could never have matched anything.
Includes a real cargo fmt fix (submodule commit d9902f9) for formatting drift left over from the earlier real-data gap-tolerance fix, which the contract job's fmt --check step would otherwise have failed on.
All four forecast CI jobs' real commands verified locally before push: cargo test/check for CPU, CUDA/WGPU compile-check, the contract job's Python assertion block, the hosted-boundaries clean-room scan, and cargo fmt --check.
* docs(forecast): record real BIDMC cross-entity holdout result (REJECT)
Two independent real-patient partitions (34/16 and 24/26 splits, 53 total
ICU patients from PhysioNet BIDMC, Open Data Commons Attribution License
v1.0) both show the model losing to trivial baselines by a decisive margin
(+69% and +601% worse WQL). First genuine cross-entity real-data test this
project has run; same conclusion as every prior synthetic and single-household
test. Full Autogenous signed verification was not run for this entry -
flagged explicitly in the note.
* chore(forecast): point ruforecast submodule at published main (example feature-gate fix)
* chore(forecast): point ruforecast submodule at published main (clippy fix)
---------
Signed-off-by: Codex <codex@openai.com>
ruv