0 Fork 0

mirror of https://github.com/ruvnet/RuView.git synced 2026-10-18 21:00:22 -05:00

Files

T

Claudeandclaude-flow 888222277e feat(npm): RuView operator surface — SDK, cross-platform flashing, training, doctor (ADR-369..372)

@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

2026-09-30 16:47:30 +00:00

3.6 KiB

Raw Permalink Blame History

ADR-371: Harness training runner and mean-pose evidence gate

Status: proposed
Date: 2026-09-30
Deciders: RuView maintainers; acceptance pending
Tags: training, evaluation, evidence, pose, honesty
Related: ADR-079 (camera-supervised pose), ADR-151 (room calibration), ADR-181 (baseline-relative PCK), ADR-369

Context

RuView has a retracted pose result (a CLAIMED 92.9%/100% that did not survive baseline and leakage review). The project rule is that pose PCK is quoted only with the mean-pose baseline, on a leakage-free held-out split, with an evidence tag. The train-pose skill says this in prose, but nothing enforces it. Running the trainers also needs repository knowledge: the tch-backend feature, a libtorch version match, and where checkpoints go.

Decision

Gate (`ruview_train_gate`, pure, read-only)

The input is a small report: metric, model and baseline score, split, train/test subjects, train end and test start, n_test, reproducer, and data source.

Hard failures:

no baseline;
a random-frame split, or any split outside chronological / blocked-gap / grouped-subject / grouped-session;
subject overlap in a grouped-subject split;
missing or overlapping time bounds for chronological or blocked-gap splits;
no improvement over the baseline.

Warnings:

subjects not reported;
subjects overlapping in non-subject splits;
n_test < 100;
a near-perfect score: at or above the implausible_score threshold in src/training.js.

Evidence tag: SYNTHETIC when data: synthetic, MEASURED only with a reproducer, otherwise CLAIMED.

On PASS the gate returns the one acceptable sentence: Held-out <metric> X% vs Y% mean-pose baseline = +Z pp (<tag>, …, <split> split).

Runner (`ruview_train`, `ruview_train_plan`)

Modes:

pose-smoke: wifi-densepose-train train --dry-run on the synthetic dataset.
pose: --data-dir, MM-Fi recordings.
room: wifi-densepose train-room, or cargo run -p wifi-densepose-cli when the binary is absent.

Rules:

All caller paths (config, data, checkpoints, enrollment, output) must resolve inside the trusted checkout. Realpaths are checked for inputs that already exist.
Pose runs use cargo run --manifest-path v2/Cargo.toml with the working directory set to the checkpoint directory, default v2/target/ruview-train/<mode>. The trainer's relative logs/ therefore never lands in the source tree.
Without confirm it returns a plan. MCP requires workspace-write plus confirm.
Failures are classified with remedies:
- libtorch absent;
- libtorch version mismatch (tch 0.24 expects torch 2.11.0);
- link failure.

Evidence

SYNTHETIC (npx @ruvnet/ruview train --mode pose-smoke --samples 16 --confirm, CPU PyTorch 2.11.0, LIBTORCH_USE_PYTORCH=1, Linux x64): the real trainer compiled, ran 20 epochs with early stopping, and wrote safetensors checkpoints and logs/training_log.csv under v2/target/ruview-train/pose-smoke. The trainer labels its own smoke metrics "DO NOT REPORT". None are reported here.
With PyTorch 2.14.1 the tch 0.24 C++ shim failed to compile. The runner classified this as a version mismatch rather than a missing install.
The gate is covered by test/training.test.mjs, including the leakage patterns behind the retraction.

Consequences

A number can pass through the harness only with its baseline and split attached. The gate does not verify the arithmetic that produced the input; reviewers must still re-run the reproducer.
Training needs the Rust toolchain and a matching libtorch. doctor covers Rust; libtorch is detected at run time.

3.6 KiB Raw Permalink Blame History