feat: add physics-constrained pose refinement

2026-08-26 02:04:55 +00:00 · 2026-08-15 14:29:07 -04:00
parent de27336fa1
commit 0370d49e4a
69 changed files with 9352 additions and 138 deletions
--- a/docs/adr/ADR-323-native-rust-physics-constrained-pose-refinement.md
+++ b/docs/adr/ADR-323-native-rust-physics-constrained-pose-refinement.md
@@ -0,0 +1,364 @@
+# ADR-323: Native Rust physics-constrained pose refinement
+
+- **Status**: Proposed
+- **Date**: 2026-08-15
+- **Deciders**: ruv
+- **Owners**: RuView perception and edge runtime maintainers
+- **Tags**: pose, physics, rust, uncertainty, provenance, abstention, edge
+- **Numbering note**: ADR-323 is the next free number in the authoring checkout. Re-run the ADR index/collision check immediately before merge and rename if needed.
+- **Extends**: ADR-020, ADR-027, ADR-079, ADR-101, ADR-135, ADR-145, ADR-150, ADR-273, ADR-279, ADR-282, ADR-295, ADR-296, ADR-297, ADR-298, ADR-302, ADR-303, ADR-304, ADR-305, ADR-306
+- **Supersedes**: None
+
+## Executive decision
+
+RuView will add a clean-room native Rust boundary between RF pose inference and
+semantic publication. It will preserve the immutable RF observation, publish a
+physics assessment, optionally produce a bounded corrected candidate, and
+abstain when required evidence is absent. It must never increase observational
+confidence merely because a pose is physically plausible.
+
+Three independently gated layers are adopted:
+
+1. A deterministic kinematic auditor and bounded covariance-weighted projector
+   using Rust and `nalgebra`.
+2. An optional articulated-body dynamics auditor using `rapier3d`.
+3. A later optional supervised residual model using Burn.
+
+The first production milestone is deterministic audit. It is not a GRIP port,
+not PPO, and not evidence that the current pose observer is production-ready.
+
+## Context
+
+ADR-101's committed Cog emits 17 COCO keypoints as normalized 2D coordinates.
+Its model has no per-joint uncertainty head and publishes a constant confidence.
+The sensing server also contains renderer-oriented EMA and bone clamping. These
+surfaces cannot establish metric 3D physics and can make weak evidence look
+more convincing.
+
+Pose output can violate bone length, floor, velocity, acceleration, and temporal
+continuity constraints. Downstream consumers also cannot reliably distinguish
+observed coordinates from derived correction. The rejected premise is:
+"physically plausible means more likely correct." Plausibility is only a prior;
+many incorrect poses are plausible.
+
+GRIP is architectural inspiration for an observer/controller split, but it
+observes four wearable IMUs and pressure insoles and drives a simulator. RuView
+observes RF, so GRIP weights are not input-compatible. External code, weights,
+simulators, and datasets require independent license review and never enter the
+runtime dependency graph by implication.
+
+## Outcome and actors
+
+For every accepted person track/timestamp, the engine returns exactly one
+`PoseRefinementV1`, including off, timeout, rejection, and abstention paths:
+
+- immutable `PoseObservationV2` content hash;
+- constraint residuals and quality disposition;
+- an optional bounded candidate and an explicit `selected` bit;
+- a typed reason when correction is unavailable;
+- model, calibration, configuration, and optional learned-artifact provenance.
+
+The RF observer owns observations and calibrated uncertainty; tracking owns
+identity stability; physics owns assessment/correction only; the sensing server
+owns deadlines, modes, publication, and rollback; the evidence engine owns
+release evaluation; clients choose raw/both/refined without silent fallback.
+
+## Input and coordinate contract
+
+Metric correction requires a monotonic nanosecond timestamp, session-scoped
+track ID, sequence and sensor epoch, 17 ordered COCO joints in metric X/Y/Z,
+per-joint positive-semidefinite covariance calibrated on held-out data, a
+versioned right-handed Z-up room frame, a normalized upward floor plane, model
+and calibration hashes, ADR-302 trust state, and authenticated/replay-protected
+source provenance.
+
+`Image2d` observations may be audited for image-plane ratios and continuity but
+must never enter 3D projection/dynamics or be called physically corrected.
+Unknown trust, missing calibration, missing uncertainty, stale/non-monotonic
+input, non-finite values, invalid covariance, excessive tracks, and room-bound
+violations fail to raw output with a typed reason.
+
+## Public contracts
+
+`wifi-densepose-core` owns `PoseObservationV2` and `PoseRefinementV1`; no
+duplicate server/Cog contract is permitted. Public output remains COCO17. The
+engine derives pelvis and thorax virtually and never labels them observed.
+
+The raw content hash is deterministic and excludes its own hash field. The
+idempotency key is `(sensor_epoch, sequence, track_id, raw_hash, config_hash)`.
+An exact duplicate returns the cached result; same sequence with different
+content is a replay rejection.
+
+Contact is `hypothesis` unless a measured sensor and its provenance say
+otherwise. Raw, derived, hypothesis, and unknown labels must survive every
+projection.
+
+## Confidence invariant
+
+For upstream calibrated confidence `c_obs`, normalized residual `r`, and
+normalized intervention `i`:
+
+```text
+c_physics = exp(-(beta_r * r + beta_i * i))
+c_effective = min(c_obs, c_obs * c_physics)
+0 <= c_effective <= c_obs <= 1
+```
+
+Only a separately witnessed multimodal fusion contract may increase fused
+confidence.
+
+## Deterministic projector
+
+The default `kinematic` feature has no Rapier, Burn, ONNX, libtorch, Python,
+CUDA, or network dependency. Per bounded iteration it:
+
+1. projects observed parent/child distances toward anonymous track-scoped
+   bone-length posteriors;
+2. applies broad joint/trunk validity checks without an upright prior;
+3. bounds temporal motion and resets derivatives after gaps;
+4. resolves floor penetration only, allowing seated, kneeling, prone, child-
+   scale, mobility-aid, and genuine-fall poses;
+5. recomputes residuals and stops below epsilon.
+
+Initial operator-owned caps are four iterations (hard maximum eight), 0.20 m
+single-joint correction, 0.10 m root correction, 250 ms derivative gap, 500 ms
+track reset, ten known joints, a 100 m metric room bound, a separate 16,384
+image-coordinate audit bound, and a 5 ms one-track Pi 5 p95 gate. Keeping image
+and metric bounds separate prevents legitimate pixel observations from
+weakening the physical room bound. A candidate over either correction cap is
+discarded in full.
+
+Bone posteriors are initialized only from high-confidence frames, anonymous,
+memory-only, track-scoped, and deleted on expiry. Persistent personalization is
+outside this ADR and requires consent/retention/deletion governance.
+
+## Optional dynamics and learned layers
+
+`dynamics` adds a process-owned Rapier humanoid and begins audit-only. Network
+input may never provide Rapier snapshots, bodies, constraints, solver limits,
+or arbitrary geometry. Dynamics approval is independent of kinematic approval.
+
+`learned` uses first-party Burn 0.21 core/NN components without `burn-tch`
+because this workspace already has a different native libtorch link.
+`learned-cpu` adds the ndarray backend. The implemented two-layer GRU uses a
+20-frame history and width 128 to predict bounded residuals, uncertainty,
+foot-contact hypotheses, and abstention. Verified model records can be loaded
+from bytes and executed natively; no trained artifact is shipped or approved.
+The resolved Burn/CubeCL graph declares Rust 1.92, while the workspace file
+pins Rust 1.89 and the authoring host provides Rust 1.91.1.
+`--ignore-rust-version` is diagnostic evidence only: learned activation remains
+blocked until an approved Rust 1.92 release-toolchain change builds it without
+that override. Residuals are hard-clipped to deterministic caps and cannot
+bypass validation or confidence monotonicity. PPO is deferred until measured
+evidence identifies a failure supervised residual learning cannot address.
+
+## Feature boundary
+
+```text
+default = kinematic
+dynamics = rapier3d
+learned = burn-core + burn-nn
+learned-cpu = learned + burn-ndarray
+learned-train = learned + burn-train
+learned-wgpu = learned-train + burn-wgpu
+learned-cuda = learned-train + burn-cuda
+deterministic = rapier3d?/enhanced-determinism
+```
+
+The lockfile is release authority. The learned feature currently requires the
+toolchain supported by Burn/CubeCL's resolved graph; this does not change the
+default edge build.
+
+## Runtime modes and API
+
+Rollout is `OFF -> AUDIT -> SHADOW_CORRECT -> OPT_IN_CORRECT -> DEFAULT_CORRECT`.
+Evidence permits forward transitions; any regression returns immediately to
+audit/off. Correct selection additionally requires authenticated sensor
+identity and replay protection from ADR-305. High model confidence cannot
+override missing source authentication.
+
+Existing pose fields stay unchanged and raw remains the migration default:
+
+```text
+GET /api/v1/pose/current?view=raw
+GET /api/v1/pose/current?view=both
+GET /api/v1/pose/current?view=refined
+```
+
+Refined-only returns HTTP 409 with `pose_refined_unavailable` when no selected
+candidate exists. It never silently returns raw labeled refined.
+
+## Security, privacy, and availability
+
+All frames, model output, geometry, and pre-verification artifacts are
+untrusted. Calibration/config/model artifacts become trusted only after signed,
+hash-addressed verification and atomic activation. Runtime inference performs
+no model retrieval or other network access.
+
+Fixed arrays/caps, bounded iterations, a maximum track count, room geometry
+limits, deadlines, and track expiry constrain denial of service. Timeout drops
+partial refinement, never raw publication. Backpressure retains the newest raw
+frame per track, drops intermediate refinement work, resets derivatives after
+250 ms, and never extrapolates beyond 500 ms.
+
+Metrics contain only allowlisted aggregate scalars: mode/disposition/reason,
+stage latency, iterations, maximum correction, residuals, confidence delta,
+track resets, invalid input, timeout, and raw/refined divergence. They exclude
+joint arrays, body dimensions, room coordinates, CSI, and persistent person
+identifiers. Bone/gait state is memory-only and excluded from logs.
+
+Refined output is not a sole medical, emergency, industrial-safety, or
+autonomous-control source. A real fall is valid state and must never be made
+upright to stabilize a simulator.
+
+## Threat model summary
+
+| Threat | Primary control | Residual risk |
+|---|---|---|
+| Spoofed/replayed sensor | ADR-305 identity, MAC, sequence and replay window; correction gate | Compromised legitimate sensor |
+| Altered model/floor/config | Signed hashes, authenticated configuration, atomic activation | Authorized unsafe configuration |
+| Poisoned data/splits | Immutable manifests, strict split validator, witnessed benchmarks | Subtle label poisoning |
+| Operator repudiation | Append-only witnessed transition with actor/old/new hash/reason | Compromised signer |
+| Biometric/log leakage | Track-local retention and fixed metric allowlist | Aggregate inference |
+| Track/geometry CPU flood | Authentication, cardinality/geometry/allocation/deadline caps | Valid dense-scene overload |
+| Remote mode escalation | Capability-scoped local control plane, deny by default | Compromised operator capability |
+| Derived output relabeled observed | Required schema/provenance and signed event envelope | Malicious downstream stripping |
+
+The implementation review records commit, lockfile hash, Rust toolchain,
+scanner versions, and advisory-feed timestamp.
+
+## Evidence protocol
+
+Evidence levels are L0 deterministic synthetic, L1 public measured replay, L2
+controlled RuView RF plus optical truth, L3 subject/room/hardware/session-
+disjoint RuView, L4 privacy-safe shadow fleet aggregates, and L5 independent
+vertical validation outside this ADR.
+
+No sequence, contiguous take, subject, room, or calibration session may cross
+train/test for the generalization gate. Preprocessing, body priors, and
+uncertainty calibration fit training data only. Reports include raw observer,
+renderer smoothing, audit, deterministic correction, dynamics audit, and
+learned residual on identical observations, plus empty-room, prone/fall,
+missing-joint, and OOD subsets.
+
+Primary metrics are 3D MPJPE, declared-threshold PCK, per-joint error, foot
+slide, floor penetration, jerk, uncertainty calibration, abstention coverage,
+and selective risk. Learned runs use at least five fixed seeds and report mean,
+median, standard deviation, and 95% bootstrap intervals. All frames count;
+selective metrics report risk and coverage.
+
+## Acceptance gates
+
+- **G0 contract**: real metric 3D/covariance output, round-trip raw hash,
+  versioned frame/floor, 2D compatibility, non-stub observer, ADR-298 artifact
+  sanity, and the ADR-079 PCK@20 >=35% gate or adopted successor. The current
+  committed Cog does not pass G0, so correction remains unavailable.
+- **G1 deterministic audit**: property/fuzz tests, deterministic hashes per
+  platform class, 24-hour accelerated replay without panic/growth, Pi 5 p95
+  <=5 ms, and universal confidence monotonicity.
+- **G2 shadow correction**: strict-disjoint measured median MPJPE improvement
+  >=10% with positive 95% CI lower bound; foot slide >=30% and jerk >=25%
+  better; no joint median >5 mm worse; fall/prone sensitivity change <=2 pp;
+  >=95% corrections below 0.10 m; every correction above 0.20 m abstains.
+- **G3 opt-in**: >=30 subjects, 10 rooms, 3 hardware configurations, and 3
+  independent sessions/room; UNKNOWN never selected; confidence monotonic;
+  live disable; REST/WebSocket/MQTT/Home Assistant/replay compatibility.
+- **G4 default visualization only**: 30 shadow days under 0.1% timeout/internal
+  error, no open severity 1/2 incidents, and gates still valid for current
+  model/calibration.
+
+Dynamics and learned engines each repeat G2-G4; approval is not inherited.
+
+## Testing and completion evidence
+
+Unit/property/fuzz/integration/security coverage maps to requirements R1-R13:
+raw hash, confidence, modes, malformed/stale/frame/covariance input, caps and
+deadlines, provenance, dependency graph, pose diversity/fall preservation,
+strict splits, fail-to-raw faults, no network capability, and authenticated
+source/replay selection.
+
+Release commands include focused core/physics tests, default/dynamics/learned
+feature checks, format/clippy, benches, `cargo deny`, `cargo audit`, strict split
+verification, and golden replay verification. Completion also requires JSON
+schemas, measured Pi 5/x86 rows, strict manifest hashes, raw/refined metrics,
+SBOM/license report, rollback drill, and residual-risk owners. Missing measured
+or operational evidence leaves status Proposed and runtime in audit.
+
+## Rollback
+
+Rollback is an authenticated mode transition to audit/off, not a binary
+downgrade. Stop selection immediately, keep raw publication and disposition
+records, discard track state, and retain only aggregate incident metrics plus
+signed configuration history. Failed artifact activation leaves the previous
+engine atomically active. Additive schemas remain; refined-only callers receive
+the typed unavailable response.
+
+## Consequences
+
+### Positive
+
+- Explicit anti-hallucination and provenance boundary after RF inference.
+- Reusable native Rust consistency primitive with measurable abstention.
+- Python/CUDA remain absent from the production default.
+- Cross-modal teacher data remains possible without wearable runtime inputs.
+
+### Negative
+
+- Full value requires a real metric 3D observer and calibrated uncertainty.
+- Stateful tracks add latency/memory; optional backends add supply-chain surface.
+- A constrained but wrong pose can look more credible.
+- Strict data collection costs more than the software implementation.
+
+### Neutral
+
+- This ADR does not improve RF observability or current weight evidence.
+- Existing 2D consumers continue to function.
+
+## Implementation phases
+
+P0 contracts/schemas; P1 deterministic audit; P2 bounded shadow correction; P3
+server/Cog publication and evidence ledger; P4 Rapier audit; P5 Burn residual
+training/inference. Code may land ahead of evidence, but runtime authority
+advances only through the gates above.
+
+## Implementation status at proposal
+
+- P0-P3 are implemented on this branch: canonical contracts, strict schemas,
+  deterministic audit/projection, authenticated correction receipts,
+  idempotency, bounded track state, latest-frame backpressure, additive HTTP
+  and WebSocket publication, live legacy-2D audit, privacy-safe metrics, golden
+  replay, and strict-split checks.
+- P4 is implemented as an optional persistent per-track Rapier dynamics auditor
+  and remains audit-only pending independent G2-G4 evidence.
+- P5 inference architecture, artifact verification, serialization, and native
+  CPU execution are implemented. Training data, a signed trained artifact, and
+  G2-G4 accuracy/calibration evidence do not exist, so the layer has no runtime
+  selection authority. Its resolved Rust 1.92 requirement is also an explicit
+  activation blocker on the current Rust 1.91.1 release host.
+- The live Cog honestly emits `Image2d`, degraded trust, and uncalibrated
+  uncertainty. It can be audited but cannot be selected for 3D correction.
+  G0 therefore remains open until an independently released metric-3D observer
+  with calibrated covariance is integrated.
+- Local x86 latency and synthetic contract checks are recorded in the append-
+  only evidence ledger. Pi 5 measurements, 24-hour replay, 100-million-case
+  fuzzing, held-out RF/optical accuracy, fleet shadowing, and vertical safety
+  validation remain release evidence gates rather than software claims.
+
+## References
+
+- [GRIP project](https://ryosukehori.github.io/grip-project/)
+- [GRIP paper (arXiv:2603.16233)](https://arxiv.org/abs/2603.16233)
+- [Rapier documentation](https://docs.rs/rapier3d/)
+- [Burn documentation](https://docs.rs/burn/0.21.0/burn/)
+- [ADR-020](./ADR-020-rust-ruvector-ai-model-migration.md)
+- [ADR-079](./ADR-079-camera-ground-truth-training.md)
+- [ADR-101](./ADR-101-pose-estimation-cog.md)
+- [ADR-150](./ADR-150-rf-foundation-encoder.md)
+- [ADR-273](./ADR-273-unified-rf-spatial-world-model.md)
+- [ADR-279](./ADR-279-native-rf-frame-contract.md)
+- [ADR-298](./ADR-298-model-release-sanity-gates.md)
+- [ADR-302](./ADR-302-out-of-distribution-detection.md)
+- [ADR-303](./ADR-303-ground-truth-synchronization.md)
+- [ADR-304](./ADR-304-evidence-engine.md)
+- [ADR-305](./ADR-305-authenticated-sensor-identity.md)
+- [ADR-306](./ADR-306-canonical-spatial-ontology.md)
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -179,6 +179,7 @@ Statuses: **Proposed** (under discussion), **Accepted** (approved and/or impleme
 | [ADR-319](ADR-319-witness-chain.md) | Witness chain — staged, signed epistemic envelope | Accepted (phase 1) |
 | [ADR-320](ADR-320-sensor-hal.md) | RuView sensor HAL — abstract all sensing hardware to one Observation type | Proposed (phase 2) |
 | [ADR-321](ADR-321-decision-policy-action-authorization.md) | Decision policy — action authorization conditioned on certificate class, freshness, uncertainty, evidence | Accepted (phase 1) |
+| [ADR-323](ADR-323-native-rust-physics-constrained-pose-refinement.md) | Native Rust physics-constrained pose refinement | Proposed |

 ---

--- a/docs/benchmarks/physics-pose-refinement.md
+++ b/docs/benchmarks/physics-pose-refinement.md
@@ -0,0 +1,71 @@
+# Physics pose refinement evidence ledger
+
+ADR-323 performance and accuracy targets are gates, not measured claims. Append
+rows; never replace prior measurements. Every row must identify the repository
+commit, lockfile hash, Rust toolchain, target, engine/features, configuration
+hash, corpus/split hash, command, sample count, and evidence label.
+
+## Runtime measurements
+
+| Date | Commit | Lock SHA-256 | Target/toolchain | Engine/config | Tracks | Samples | p50 | p95 | p99/max | RSS delta | Evidence | Reproducer |
+|---|---|---|---|---|---:|---:|---:|---:|---:|---:|---|---|
+| 2026-08-15 | `de27336` + uncommitted ADR-323 changes | `552737eab9092b59ea9dd2b2caf68389f0b0966679f0fbb33ff2b1b3d42e2668` | Windows x86_64, Intel Core Ultra 9 285H, rustc 1.91.1 | deterministic kinematic shadow, config `ef3cf581f75124c1d45a8d6bedcef32e4d1bacb39ee0dfcd4e520171fda2d8cf` | 1 | 20,000 | 0.0080 ms | 0.0097 ms | 0.0195/0.5465 ms | not measured | **MEASURED**, local host only; not Pi 5 evidence | `cargo run --release -p wifi-densepose-physics --example latency_probe -- 20000` |
+| 2026-08-15 | `de27336` + uncommitted ADR-323 changes | `552737eab9092b59ea9dd2b2caf68389f0b0966679f0fbb33ff2b1b3d42e2668` | Windows x86_64, Intel Core Ultra 9 285H, rustc 1.91.1 | deterministic kinematic shadow after final local optimization, same config | 1 | 20,000 | 0.0075 ms | 0.0084 ms | 0.0117/0.1579 ms | not measured | **MEASURED**, local host only; not Pi 5 evidence | same release probe command |
+| 2026-08-15 | `de27336` + uncommitted ADR-323 changes | `552737eab9092b59ea9dd2b2caf68389f0b0966679f0fbb33ff2b1b3d42e2668` | Windows x86_64, Intel Core Ultra 9 285H, rustc 1.91.1 | final deterministic kinematic shadow, config `a44dc696234f31eda54cd4b436bc2d2c69b9638565b729ac9f07435cedfd0dcc` | 1 | 20,000 | 0.0071 ms | 0.0084 ms | 0.0147/1.5994 ms | not measured | **MEASURED**, local host only; not Pi 5 evidence | same release probe command |
+
+The probe measures a warm, one-track `PhysicsEngine::process` call. It excludes
+transport, publication, resident-memory delta, dynamics, and learned inference.
+It is not evidence for the Pi 5 gate.
+
+Criterion separately measured `kinematic_one_track` at
+`[11.911, 12.757, 14.069] us` across 100 samples (approximately 369,000 timed
+iterations). That benchmark includes observation construction and canonical
+hashing in the timed routine and uses fresh engine state; it is **MEASURED** on
+the same local host and is not a percentile or Pi 5 claim.
+
+## Accuracy measurements
+
+| Date | Commit | Corpus/split | Variant | Coverage | MPJPE | PCK threshold/result | Foot slide | Jerk | Fall/prone delta | Evidence |
+|---|---|---|---|---:|---:|---|---:|---:|---:|---|
+
+No measured accuracy evidence has been recorded. The deterministic tests are
+L0/SYNTHETIC contract evidence only and cannot satisfy G2.
+
+## Validation and supply-chain record
+
+- The default dependency graph is checked to exclude Burn, Rapier, Tch, and
+  ONNX Runtime. Dynamics and learned backends remain opt-in.
+- Burn CPU serialization/inference tests pass on the authoring host only with
+  Cargo's `--ignore-rust-version`; the resolved CubeCL graph requires Rust 1.92.
+  The workspace file pins Rust 1.89 and the host provides Rust 1.91.1. This is
+  diagnostic, not release approval.
+- `cargo audit 0.22.1` used RustSec database commit
+  `69f93cf294852cfa9b53751f4ca86de3283dd290` (feed timestamp 2026-08-12).
+  ADR-323 updates remove resolved advisories in `event-listener`, `rkyv`, and
+  `wasmtime`. The workspace still has five advisories in pre-existing
+  `quick-xml` and `rsa` dependency paths; the default physics graph contains
+  none of them. The optional Burn training graph includes yanked `spin 0.9.8`.
+- `cargo-deny` is not installed on the authoring host, so the required license
+  and policy gate is not claimed complete.
+- Strict Clippy passes with warnings denied for core/physics default and
+  dynamics builds, the diagnostic learned-CPU build, and the Cog itself with
+  dependency linting excluded. Focused core, physics, dynamics, learned, Cog,
+  sensing-server adapter/live-audit/HTTP, schema, golden, strict-split,
+  feature-boundary, and fuzz-build checks pass.
+- The repository-wide rustfmt gate is already red across unrelated crates. The
+  sensing-server library has existing warning debt, and unscoped Cog Clippy is
+  blocked by existing `wifi-densepose-ruvector` warnings. The prescribed
+  `cargo test --workspace --no-default-features` did not reach a terminal result
+  in either a 904-second cold or 604-second warm serial run on this Windows
+  host. None of these broader gates is represented as green.
+- The standalone fuzz lock SHA-256 is
+  `d386c4edb130bb6b2d1a4ef77334c78e25e0695e90a9d97c01284876acb8c2c6`.
+
+## Required commands
+
+```text
+cargo bench -p wifi-densepose-physics
+node scripts/pose-physics/verify-feature-boundary.mjs
+bash scripts/verify-pose-physics-splits.sh <manifest.json>
+bash scripts/replay-pose-physics-golden.sh <golden-results.jsonl>
+```
--- a/docs/schemas/pose-observation-v2.schema.json
+++ b/docs/schemas/pose-observation-v2.schema.json
@@ -0,0 +1,38 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://ruview.net/schemas/pose-observation-v2.schema.json",
+  "title": "PoseObservationV2",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "timestamp_ns", "sensor_epoch", "sequence", "track_id", "frame", "calibration_id", "floor_plane", "model", "source", "trust_state", "dimensionality", "uncertainty_calibrated", "joints", "observer_confidence", "canonical_hash"],
+  "properties": {
+    "schema_version": { "const": 2 },
+    "timestamp_ns": { "type": "integer", "minimum": 0 },
+    "sensor_epoch": { "type": "integer", "minimum": 0 },
+    "sequence": { "type": "integer", "minimum": 0 },
+    "track_id": { "$ref": "#/$defs/string_id" },
+    "frame": { "$ref": "#/$defs/frame" },
+    "calibration_id": { "$ref": "#/$defs/string_id" },
+    "floor_plane": { "oneOf": [{ "type": "null" }, { "$ref": "#/$defs/floor" }] },
+    "model": { "$ref": "#/$defs/model" },
+    "source": { "$ref": "#/$defs/source" },
+    "trust_state": { "enum": ["KNOWN", "DEGRADED", "UNKNOWN"] },
+    "dimensionality": { "enum": ["image2d", "metric3d"] },
+    "uncertainty_calibrated": { "type": "boolean" },
+    "joints": { "type": "array", "minItems": 17, "maxItems": 17, "items": { "$ref": "#/$defs/joint" } },
+    "observer_confidence": { "$ref": "#/$defs/probability" },
+    "canonical_hash": { "$ref": "#/$defs/hash" }
+  },
+  "$defs": {
+    "probability": { "type": "number", "minimum": 0, "maximum": 1 },
+    "hash": { "type": "array", "minItems": 32, "maxItems": 32, "items": { "type": "integer", "minimum": 0, "maximum": 255 } },
+    "string_id": { "type": "string", "minLength": 1, "maxLength": 128 },
+    "vec3": { "type": "array", "minItems": 3, "maxItems": 3, "items": { "type": "number" } },
+    "frame": { "type": "object", "additionalProperties": false, "required": ["name", "version", "metric", "right_handed", "z_up"], "properties": { "name": { "type": "string", "minLength": 1, "maxLength": 128 }, "version": { "type": "integer", "minimum": 1 }, "metric": { "type": "boolean" }, "right_handed": { "type": "boolean" }, "z_up": { "type": "boolean" } } },
+    "floor": { "type": "object", "additionalProperties": false, "required": ["normal", "offset_m"], "properties": { "normal": { "$ref": "#/$defs/vec3" }, "offset_m": { "type": "number" } } },
+    "model": { "type": "object", "additionalProperties": false, "required": ["id", "artifact_hash"], "properties": { "id": { "type": "string", "minLength": 1, "maxLength": 128 }, "artifact_hash": { "$ref": "#/$defs/hash" } } },
+    "source": { "type": "object", "additionalProperties": false, "required": ["sensor_id", "authenticated", "replay_protected"], "properties": { "sensor_id": { "type": "string", "minLength": 1, "maxLength": 128 }, "authenticated": { "type": "boolean" }, "replay_protected": { "type": "boolean" } } },
+    "covariance": { "type": "object", "additionalProperties": false, "required": ["xx", "xy", "xz", "yy", "yz", "zz"], "properties": { "xx": { "type": "number", "minimum": 0 }, "xy": { "type": "number" }, "xz": { "type": "number" }, "yy": { "type": "number", "minimum": 0 }, "yz": { "type": "number" }, "zz": { "type": "number", "minimum": 0 } } },
+    "joint": { "type": "object", "additionalProperties": false, "required": ["kind", "position_m", "covariance_m2", "confidence", "visibility"], "properties": { "kind": { "enum": ["nose", "left_eye", "right_eye", "left_ear", "right_ear", "left_shoulder", "right_shoulder", "left_elbow", "right_elbow", "left_wrist", "right_wrist", "left_hip", "right_hip", "left_knee", "right_knee", "left_ankle", "right_ankle"] }, "position_m": { "$ref": "#/$defs/vec3" }, "covariance_m2": { "$ref": "#/$defs/covariance" }, "confidence": { "$ref": "#/$defs/probability" }, "visibility": { "enum": ["visible", "occluded", "unknown"] } } }
+  }
+}
--- a/docs/schemas/pose-refinement-v1.schema.json
+++ b/docs/schemas/pose-refinement-v1.schema.json
@@ -0,0 +1,36 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://ruview.net/schemas/pose-refinement-v1.schema.json",
+  "title": "PoseRefinementV1",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema_version", "raw_observation_hash", "mode", "disposition", "selected", "refined_joints_m", "physics_confidence", "effective_confidence", "intervention", "residuals", "refined_residuals", "contact_hypotheses", "dynamics", "provenance", "reason", "canonical_hash"],
+  "properties": {
+    "schema_version": { "const": 1 },
+    "raw_observation_hash": { "$ref": "#/$defs/hash" },
+    "mode": { "enum": ["off", "audit", "shadow_correct", "opt_in_correct", "default_correct"] },
+    "disposition": { "enum": ["bypassed", "audited2d", "audited", "shadowed", "corrected", "abstained", "rejected"] },
+    "selected": { "type": "boolean" },
+    "refined_joints_m": { "oneOf": [{ "type": "null" }, { "type": "array", "minItems": 17, "maxItems": 17, "items": { "$ref": "#/$defs/vec3" } }] },
+    "physics_confidence": { "$ref": "#/$defs/probability" },
+    "effective_confidence": { "$ref": "#/$defs/probability" },
+    "intervention": { "$ref": "#/$defs/intervention" },
+    "residuals": { "$ref": "#/$defs/residuals" },
+    "refined_residuals": { "oneOf": [{ "type": "null" }, { "$ref": "#/$defs/residuals" }] },
+    "contact_hypotheses": { "type": "array", "minItems": 2, "maxItems": 2, "items": { "$ref": "#/$defs/contact" } },
+    "dynamics": { "oneOf": [{ "type": "null" }, { "$ref": "#/$defs/dynamics" }] },
+    "provenance": { "$ref": "#/$defs/provenance" },
+    "reason": { "enum": [null, "mode_off", "unsupported_schema", "hash_mismatch", "invalid_number", "invalid_covariance", "stale_input", "coordinate_frame_mismatch", "calibration_unavailable", "uncertainty_uncalibrated", "ood_unknown", "source_unauthenticated", "replay_protection_unavailable", "correction_not_authorized", "replay_rejected", "non_monotonic_input", "too_few_known_joints", "correction_too_large", "deadline_exceeded", "track_capacity", "internal_error"] },
+    "canonical_hash": { "$ref": "#/$defs/hash" }
+  },
+  "$defs": {
+    "probability": { "type": "number", "minimum": 0, "maximum": 1 },
+    "hash": { "type": "array", "minItems": 32, "maxItems": 32, "items": { "type": "integer", "minimum": 0, "maximum": 255 } },
+    "vec3": { "type": "array", "minItems": 3, "maxItems": 3, "items": { "type": "number" } },
+    "intervention": { "type": "object", "additionalProperties": false, "required": ["max_joint_correction_m", "root_correction_m", "corrected_joint_count", "solver_iterations", "elapsed_us"], "properties": { "max_joint_correction_m": { "type": "number", "minimum": 0 }, "root_correction_m": { "type": "number", "minimum": 0 }, "corrected_joint_count": { "type": "integer", "minimum": 0, "maximum": 17 }, "solver_iterations": { "type": "integer", "minimum": 0, "maximum": 8 }, "elapsed_us": { "type": "integer", "minimum": 0 } } },
+    "residuals": { "type": "object", "additionalProperties": false, "required": ["bone_m", "joint_limit_rad", "velocity_mps", "acceleration_mps2", "temporal_jerk", "floor_penetration_m", "contact_m", "collision_m", "normalized_total"], "properties": { "bone_m": { "type": "number", "minimum": 0 }, "joint_limit_rad": { "type": "number", "minimum": 0 }, "velocity_mps": { "type": "number", "minimum": 0 }, "acceleration_mps2": { "type": "number", "minimum": 0 }, "temporal_jerk": { "type": "number", "minimum": 0 }, "floor_penetration_m": { "type": "number", "minimum": 0 }, "contact_m": { "type": "number", "minimum": 0 }, "collision_m": { "type": "number", "minimum": 0 }, "normalized_total": { "type": "number", "minimum": 0 } } },
+    "contact": { "type": "object", "additionalProperties": false, "required": ["state", "probability"], "properties": { "state": { "enum": ["hypothesis", "measured", "unknown"] }, "probability": { "$ref": "#/$defs/probability" } } },
+    "dynamics": { "type": "object", "additionalProperties": false, "required": ["stable", "segment_count", "joint_count", "contact_count", "substeps", "tracking_error_m", "joint_anchor_error_m", "floor_penetration_m", "control_effort"], "properties": { "stable": { "type": "boolean" }, "segment_count": { "type": "integer", "minimum": 0, "maximum": 255 }, "joint_count": { "type": "integer", "minimum": 0, "maximum": 255 }, "contact_count": { "type": "integer", "minimum": 0, "maximum": 65535 }, "substeps": { "type": "integer", "minimum": 1, "maximum": 8 }, "tracking_error_m": { "type": "number", "minimum": 0 }, "joint_anchor_error_m": { "type": "number", "minimum": 0 }, "floor_penetration_m": { "type": "number", "minimum": 0 }, "control_effort": { "type": "number", "minimum": 0 } } },
+    "provenance": { "type": "object", "additionalProperties": false, "required": ["engine", "engine_version", "config_hash", "rf_model_hash", "calibration_id", "learned_artifact_hash"], "properties": { "engine": { "type": "string", "minLength": 1, "maxLength": 128 }, "engine_version": { "type": "string", "minLength": 1, "maxLength": 64 }, "config_hash": { "$ref": "#/$defs/hash" }, "rf_model_hash": { "$ref": "#/$defs/hash" }, "calibration_id": { "type": "string", "minLength": 1, "maxLength": 128 }, "learned_artifact_hash": { "oneOf": [{ "type": "null" }, { "$ref": "#/$defs/hash" }] } } }
+  }
+}