Benchmarks, Voxel World Model

July 20, 2026 366/366 tests · kernel v1.0

V30, Kernel Finalization & Long-Term Stability

Final kernel milestone. Canonical worlds proven stable across replay, snapshot/restore, and replication — identical authoritative checksums.

Tests366/366

Kernelv1.0

Canonical worlds5/5 ALL OK

Replication health1.0

InvariantsPASS

Sig defaultHMAC

Kernel v1.0 panel with canonical regression ALL OK. — **Live panel.** TinyRoom · Hallway · Apartment · Office · Campus — replay ✓ snapshot ✓ replication ✓.

Definition-of-success proofs

Claim	Verified
Kernel v1.0 declared	`kernel_version: 1.0` on `/api/kernel/health`
Observation-driven identity	raw observations only; confidence recomputed; no object state on wire
Ed25519 public-key	sign/verify; tampered/wrong-key rejected; HMAC remains fast default
Canonical regression	5 worlds · replay + snapshot + replication checksum-identical
Governance / API contract	GOVERNANCE.md · ADR-0001…0005 · KERNEL_API.md · KERNEL_V1.0.md
Last kernel milestone	post-v1.0 tracks version independently (no V31/V32)

Honest limits: pure-Python Ed25519 is slow (HMAC default); anti-entropy is throttled tick; policy migrations registered by apps; regression golden masters are geometry-authoritative.

Read full V30 narrative →

July 20, 2026 355/355 tests · ops not conclusions

V29, Distributed Kernel Hardening & Trust

Replicas exchange authoritative operations; derived confidence/hypotheses/facts are recomputed locally. Trust, signatures, provenance, and versioned policies are kernel-enforced.

Tests355/355

Ops applied (live)3

Recomputes1

Viewer denied5

Forged packageREJECTED

demo_trustconverged

Distributed Trust (V29) panel with CONVERGED via operations and forged REJECTED. — **Live panel.** CONVERGED via operations · recomputes · viewer denied · forged REJECTED.

Definition-of-success proofs

Claim	Verified
Operations, not conclusions	evidence/facts replicated; confidence recomputed locally
Trust enforcement	viewer denied semantic ops; only admin may verify
Byzantine safety	forged/tampered packages rejected before mutation
Policy incompatibility	differing confidence policy reported, not merged
Compaction / partial / anti-entropy	replay preserved · region sync · detect+repair
Provenance invariant	evidence retains origin after sync

Honest limits: HMAC (not ed25519 yet); policy migration detect-only; anti-entropy on-demand; identity ops by uuid+recompute (semantic+reasoning fully op-streamed).

Read full V29 narrative →

July 20, 2026 341/341 tests · distributed world

V28, Distributed World Replication & Synchronization

Many replicas, one logical world — converge through validated transactions, vector clocks, explicit conflicts, and abstract transport.

Tests341/341

E2E scenarios8/8

Live demo_syncconverged

EntitiesSink · Sofa

Transportsmem + file

Invariants after syncpreserved

Distributed World (V28) panel with CONVERGED sync and matching checksums. — **Live panel.** a→b / b→a 1 tx · CONVERGED · matching checksums · Sink · Sofa.

Definition-of-success proofs

Claim	Verified
Sync via validated transactions only	peer changes → local `world.transaction(source="sync")`; no memory copy
Vector-clock causality	before/after/concurrent/equal; conflict only if concurrent + overlapping state
Explicit conflicts	geometry (manual) · semantic (preserve-alternatives) · schema (reject)
Offline / multi-replica	offline robot · 3-replica multi-robot · file transfer · checkpoint recovery
V26 invariants after sync	`test_replication_never_violates_invariants`
Honest live demo	fresh transactional replicas (hallway snapshot is not in the replication log)

Honest limits: mem+file transports only; state merge by uuid (not per-tx semantic deltas); discovery is abstract; HTTP/WebSocket not built.

Read full V28 narrative →

July 19, 2026 329/329 tests · world as service

V27, World Services & Query Engine

Applications use world.query only — spatial, temporal, semantic, identity, reasoning, and graph families, plus cross-layer explain.

Tests329/329

Query families6

Explain (live)~0.2 ms

Graph19n / 25e

Index entities4

Layers in explainall

World Query (V27) panel with cross-layer explanation. — **Live panel.** chairs → Chair1 · geometry + semantic + evidence + reasoning in one WorldExplanation.

Definition-of-success proofs

Claim	Verified
One app interface	apps never import semantic/identity/reasoning internals
Cross-layer explain	one call joins geometry + class/hyps + identity + evidence + facts + causal
World as graph	Entity/Identity/Fact/Evidence/Event nodes with traversal
Spatial / temporal	SVO LoS · free-space BFS reachable · changed(since=…)
Persistence / concurrency	snapshot→query · replay→query · concurrent query during capture
Honest ask()	keyword parser, not NLP; same entrypoint for future LLM

Honest limits: ask() is not NLP; planner is index-vs-scan; large-world path planning out of scope.

Read full V27 narrative →

July 19, 2026 311/311 tests · kernel proof

V26, Integrated World Kernel

One labeled observation drives the whole stack via events — transactional commit, semantics, identity, reasoning, snapshot restore with identical authoritative checksum.

Tests311/311

Tx commit~1.4 ms

Invariant audit<1 ms

Hallway snapshot~120 ms

Invariants7 PASS

Checksumidentical

World Kernel (V26) panel with live correlation trace. — **Live panel.** ObservationReceived → … → WorldFactCreated · invariants PASS · 0 event failures.

Definition-of-success proofs

Claim	Verified
No manual downstream calls	one observation → automatic event chain (correlation + causation linked)
Transactional authority	single-writer TransactionManager; ConflictError / NestedTransactionError; full rollback
Save → restore equivalence	hallway 10,763 voxels · authoritative checksum identical
Replay determinism	two identical runs → identical authoritative checksums
Invariant audit	7/7 PASS live; modes audit/dev/test/production
Product e2e fixed	`test_vwm_product_e2e.py` 5/5 pass (reconstruction shadowing fixed)

Honest limits: no incremental snapshots; cross-layer query package deferred; learned perception not added; no many-worker stress suite.

Read full V26 narrative →

July 19, 2026 289/289 tests · stack complete

V25, World Reasoning Engine

Fourth layer of the stack: evidence-backed inference over the world. Classical rules, causal graphs, competing hypotheses, explanations — never mutates geometry, semantics, or identity.

Tests289/289

Live facts18

Inferred8

Rules6

Unresolved hyps1

E2E scenarios5

World Reasoning (V25) panel with causal chain for Light1 is off. — **Live panel.** Light1 is off 0.83 · Breaker1 tripped ⇒ Outlet1 no_power ⇒ Light1 off.

Definition-of-success proofs

Claim	Verified
Inference chain enforced	observation → proposal → evidence → confidence → WorldFact; nothing skips
Classical rules without AI	door→accessible, breaker→lights, chair→occupancy, disappearance→3 hyps
Causal chain reconstruction	Breaker1 tripped → Outlet1 no_power → Light1 is off (live)
Competing hypotheses	Statue disappearance keeps 3 unresolved alternatives
Never mutates the world	voxels / semantics / identities byte-identical after infer/predict
Predictions are hypothetical	`predicted=True` facts never overwrite current state

Honest limits: classical rules only; disappearance priors labelled as such; prediction is persistence projection; AI reasoner is a plugin interface.

Read full V25 narrative →

July 19, 2026 276/276 tests · time-travel ready

V24, Persistent World Identity

Objects keep UUIDs across observations. Probabilistic multi-signal matching, accumulated confidence, immutable history, and replay without mutating the present.

Tests276/276

Match signals7

Lifecycle states8

Live objects2

Chair versions5

E2E scenarios5

Persistent Identity (V24) panel with objects and change timeline. — **Live panel.** Chair Occluded · conf ~1.0 · v5 · added/moved/recolored timeline.

Definition-of-success proofs

Claim	Verified
Moved object keeps identity	chair Office→Conference same UUID; spatial weight deliberately low
Attribute change ≠ new object	wall repainted; identity persists, recolored event recorded
Occlusion → re-match	lifecycle Tracked→Occluded; UUID preserved on reappear
New object stays distinct	size-dominated geometry; cup ≠ chair; ambiguity → competing hypotheses
Replay without mutation	`state_at`/`replay`/`compare` from immutable log; versions stamp detection time
Honest matcher	weighted multi-signal heuristic; `set_matcher`/`register_embedder` hooks only

Honest limits: matcher is transparent heuristic not learned embeddings; auto merge/split is explicit API only; capture→identity auto-wiring is next integration.

Read full V24 narrative →

July 19, 2026 260/260 tests · meaning decoupled

V23, Semantic World Engine

Entities own meaning (hypotheses, attributes, relationships) and only reference authoritative voxels — one evidence path; AI/heuristics propose, humans verify.

Tests260/260

Ontology22 classes

Geometry mutations0

Review states3

Live entities3

AI proposalDoor 0.82

Semantic World (V23) panel with entities and 3D highlight. — **Live panel.** Entity list · competing hypotheses · highlight box over Wall region.

Definition-of-success proofs

Claim	Verified
Geometry never modified	every semantic op leaves `world.coords` byte-identical (14 tests)
Evidence-only mutation	noisy-OR confidence; human verify/override authoritative
Competing hypotheses	Wall entity: Door 0.82 vs Room 0.50 after AI + heuristic proposals
Relationships + inverse	contains↔inside auto-inverse; persistent roundtrip
Highlight → world	real bounds + voxel refs from authoritative coords
Honest proposer	HeuristicProposer shape-based; no bundled perception model

Honest limits: search() is lexical; HeuristicProposer is geometric not deep AI; real detectors wire via propose(source="ai") through V22 registry later.

Read full V23 narrative →

July 19, 2026 246/246 tests · real CUDA

V22, Neural Enhancement Backends

Neural models as interchangeable evolution backends — never the renderer. Real-ESRGAN RRDBNet x4 runs on the RTX 2060 SUPER; registry-driven; world untouched.

Tests246/246

Live devicecuda

detail_gain0.75

Views / tiles3 / 3

Failures0

Backends5

Neural Backends (V22) panel with realesrgan CUDA diagnostics. — **Live panel.** Preferred realesrgan · VRAM · last cuda optimize.

Definition-of-success proofs

Claim	Verified
Real GPU inference	RealESRGAN on RTX 2060 SUPER; 0 missing weight keys; live `/api/neural/control` optimize
Registry-only	evolution calls `backend.optimize(scene)`; engine names no model
New cached revision	hallway pass produces new evolution cache entry; world untouched
VRAM / fallback	tiling within budget; CPU fallback on CUDA error; OOM → classical
Honest stubs	FLUX/SwinIR/HiDream report missing weights; never fake

Honest limits: scene update is neural-gated colour-detail transfer (detail_gain), not full splat inverse-rendering; quality_score can stay flat on colour-only changes.

Read full V22 narrative →

July 19, 2026 231/231 tests · reactive

V21, Reactive World Pipeline

Runtime becomes the backbone: immutable RuntimeEvents, bus + reactors, deterministic record/replay. One world edit cascades VoxelModified → … → GaussianFrameEnhanced with zero orchestration code.

Tests231/231

Canonical types22

Reactors4

Backpressure0

Orchestration in chain0

Replaydeterministic

Reactive Runtime (V21) panel with event stream. — **Live panel.** Colour-coded stream · named subscribers · cascade demo.

Definition-of-success proofs

Claim	Verified
Cascade	`demo_edit` → VoxelModified → ChunkInvalidated → ChunkCompiled → GaussianSceneCreated → SceneUpdated → GaussianFrameRendered → GaussianFrameEnhanced
Genuine events	LegacyBridge republishes real V11 engine events
Independence	reactors independently removable
Replay	record → identical (seq, type) sequence
World authority	`test_world_untouched_by_reactive_layer`

Honest limits: V11 mark_dirty still schedules compile internally; reactors stay lightweight; LoopbackChannel is in-process; SimulationTick/Playback* vocabulary not yet emitted from those sources.

Read full V21 narrative →

July 19, 2026 216/216 tests · unified intake

V20, Reality Capture Pipeline

Any source → ObservationPacket → Registration + Evidence → voxels. Never writes voxels directly. Closes the capture end of the E2E loop.

Tests216/216

Live voxels (via Evidence)1994

Live frames12

Direct voxel writes0

Record→replayidentical

Hardware stubsguarded

Reality Capture (V20) panel with synthetic capture stats. — **Live panel.** Synthetic capture routed through Evidence.

Definition-of-success proofs

Claim	Verified
Routing invariant	`voxels_added == world coords` from capture path
Multi-device sync	timestamp bundles within tolerance
Replay	identical packet sequence + seek/step/speed
Extensibility	new device = CaptureDevice + register
No invented depth	rgb-only → `needs_reconstruction`

Honest limits: hardware drivers are guarded stubs; distortion/extrinsic application in fusion is stored but not fully applied yet.

Read full V20 narrative →

July 19, 2026 164/164 tests · async cache

V19, Gaussian Evolution Engine

Improves Gaussian scene data (not pixels) asynchronously: compiler scene → optimizer passes → cached revisions → best version served transparently. Complements V18. Voxel world untouched; rendering never blocks.

Tests164/164

Quality0.65→1.0

Optimizers10

Splats cached73 768

World mutations0

Blocking rendernever

Gaussian Evolution (V19) panel with best quality 1.0 and cache stats. — **Live panel.** `/api/gevolution/diagnostics` — balanced pass cached; quality 1.0.

Definition-of-success proofs

Claim	Verified
Data, not pixels	8 classical optimizers transform splat arrays via `GaussianOptimizer.optimize(scene)`
Async	persistent queue + worker thread; render/sim never wait
Progressive cache	new revision per pass; hallway 0.65→1.0; history preserved
World untouched	`test_world_unchanged_by_evolution`
Browser transparency	V18 path uses `best_arrays()` when cache is better
Extensibility	brighten = class + register; engine unchanged
AI stubs	FLUX / HiDream report blocker; skipped, not fatal

Honest limits: classical splat processing is what runs without GPU weights; quality score is a splat-data heuristic (not perceptual); camera-motion predictive prefetch is scaffolded, predictor TBD.

Read full V19 narrative →

July 18, 2026 150/150 tests · packet-only

V18, Gaussian Enhancement Pipeline

Enhances the Gaussian render itself (not the full display frame). Compiler Gaussian artifact → CPU splat rasterizer → GaussianImagePacket → visible-region 32-D embedding → chainable enhancers → enhanced image. Voxel world untouched.

Tests150/150

Enhancers10

Embedding32-D

Profileslow→cloud

World mutations0

Disabledpassthrough

Live panel. /api/genh/diagnostics — enabled · medium · upscale→detail→material.

Definition-of-success proofs

Claim	Verified
Packet contract	rgb + depth + normals + motion + semantic + material + object-ids + confidence + metadata
World untouched	`test_...world_untouched` — enhancement is packet-only
Metadata conditioning	visible-region → 32-D embedding; never raw voxels / whole world
Passthrough	disabled / low profile identical to raw Gaussian render
Live image	`/api/genh/image?profile=medium` → 320×240 enhanced hallway
Extensibility	new enhancer (sepia) = class + register; engine unchanged
AI stubs	RealESRGAN / photorealistic / reflection report exact blocker; no fabricate

Honest limits: classical-CV enhancement is what runs without GPU weights; CPU reference rasterizer ≠ browser WebGL splats; display-mode wiring for the enhanced frame is a follow-up; metadata unprojection is subsampled.

Read full V18 narrative →

July 17, 2026 135/135 tests · world alive

V17, World Simulation

Simulation evolves the authoritative world through time — no new foundation. Tick → behaviors → Runtime → Evidence → World → Compiler (auto) → Visualization. Door open auto-recompiles gaussian 768→764 with zero special integration code.

Tests135/135

Door → splats768→764

Behaviors9

Artifact writes by sim0

auto_wireChunkInvalidated

Determinism✓

Simulation (V17) panel showing running tick and object/behavior counts. — **Simulation panel.** Live tick / time / speed / objects / behaviors / triggers.

Pipeline + interaction proofs

Claim	Verified
Tick pipeline	`tick → behaviors+triggers → ctx.change_voxels → Runtime → Evidence → World → Compiler (auto) → Viz`
Auto recompile	button→door removes 4 voxels; gaussian chunk 768→764 splats, checksum changes
Events	DoorOpened + VoxelChanged both fire; sim never writes artifacts
Interactions	button→door · battery→light (to 0) · traffic-light cycle · proximity→open
Determinism	same world + events + seed → identical timeline
Extensibility	new Behavior (sprinkler) = class + attach; engine unchanged

Honest limits: behaviors are illustrative (architecture over realism); browser play/pause/step UI is a follow-up; world edits should bump runtime revision the same way simulation does.

Read full V17 narrative →

July 17, 2026 118/118 tests · stack complete

V16, Universal Visualization Engine

One rendering entry point. Every renderer consumes V15 compiler artifacts — never voxels. Hot-switch never recompiles. Caps the major subsystem stack.

Tests118/118

Renderers17

Gaussian via artifacts73 768

Hot-switch compiles0

World mutations0

source=compiler_artifacts

Visualization Engine (V16) panel listing 17 geometry and debug renderers. — **Visualization Engine panel.** 17 renderers · active voxel · artifacts-only invariant.

Renderer catalog + invariants

Category	Renderers
Geometry (10)	voxel · gaussian · mesh · wireframe · normals · semantic · materials · lighting · occupancy · bounding_boxes
Debug (7)	chunk_boundaries · dirty_regions · registration_graph · evidence_confidence · compiler_artifacts · component_hierarchy · runtime_events
Live verify	`/api/viz/render?renderer=gaussian` → n=73 768 · `source: "compiler_artifacts"`

Honest limits: Three.js live modes still draw via V8–V15 representation paths (browser migration remaining); stylized renderers are the extensible target; screenshots/recording descriptions are engine-side, pixels captured in-browser.

Read full V16 narrative →

July 17, 2026 105/105 tests · Part 1

V15 Part 1, World Compiler Framework

One runtime owns all derived-artifact compilation. Five built-in plugins wrap real Gaussian / SVO / mesh / semantic / lighting logic; dependency ordering, transactions, cache, and error isolation are verified.

Tests105/105

Compilers5

Bus events10

Kernel modules15

World writes0

Extensibilityregister()

World Compiler (V15) panel showing 5 compilers and cache/transaction stats. — **World Compiler panel.** Live compiler count, artifacts, transactions, failures, and cache rate.

Kernel guarantees (verified in tests)

Property	Verified behavior
Dependency order	gaussian before lighting; topological `(compiler, chunk)` work
Incremental	only dirty-chunk artifacts recompile
Transactional	invisible until commit · atomic swap · rollback on failure
Cache	same chunk/revision/params → hit, no recompile
Determinism	identical inputs → identical content checksum
Isolation	throwing compiler fails alone; world untouched
Extensibility	`implement Compiler + register()`; runtime class unchanged

Part 1 honest limits: viewer still reads V11 runtime.compilers for live Gaussian/SVO (Part 2 routes through world.compiler); single-threaded drain today; persistence is metadata + per-compiler payloads.

Read full V15 Part 1 narrative →

July 17, 2026 88/88 tests

V14, Complete Registration & Mapping

Global SE(3) pose-graph optimization, loop closure, landmarks, anchors, and incremental world corrections. Synthetic-GT benchmark deliberately surfaces the hard cases — not just the wins.

Tests88/88

Loop drift8.99→0 m

Optimize time70 ms

Chi²241→0

Rotated GT0.0°

No-overlaprejected

Registration / Mapping (V14) diagnostics panel over the voxel hallway. — **Registration / Mapping panel.** Live pose-graph stats, quality, drift, sessions, landmarks, anchors.

Synthetic GT registration benchmark (artifacts/registration/benchmark.json)

Scene	Fitness / error	Accepted?	Notes
Loop opt (12 nodes)	8.99 m → 0.0 m · chi² 241 → 0	yes	5 iters · 70 ms
Rotated 0.4 rad	fitness 1.0 · 0.0°	yes	exact recovery
Large overlap 90%	fitness 0.90 · 1.4°	yes	accepted
Scale 1.8× from identity	fitness 0.03	no	ICP needs scale-aware init
Tiny overlap 16%	fitness 0.38 · 151° wrong	yes (shouldn’t)	raise confidence gate
No overlap	fitness 0.02	rejected	correct gate

Measured on CPU synthetic GT only — real-sensor / real-model registration is not claimed. Numerical Jacobians today; analytic + sparse Cholesky for scale-up.

Read full V14 narrative →

July 17, 2026 72/72 tests

V12, Streaming Reconstruction & Persistent Gaussians

Provider-driven streaming perception into the V11 evidence engine, ICP registration that closes the V10 gap, and a chunked persistent Gaussian scene with a real elliptical splat renderer. Real LingBot inference is unsupported here (no CUDA / no checkpoint) — and that is stated, not faked.

Tests72/72

Splats live73 768

Gaussian chunks268

ICP fitness>0.9

Scale recovery2.5× exact

LingBot drift0

Gaussian Splat (v12) representation: elliptical splat reconstruction of the hallway with Runtime panel showing gaussian_scene 268. — **Gaussian Splat (v12).** 73 768 elliptical splats from 268 compiled chunks — instanced quads with gaussian falloff, not point sprites.

Measured (docs/V12_BENCHMARKS.md) — synthetic GT + real algorithms

Metric	Value	Basis
ICP known-transform recovery	fitness >0.9 · rmse <0.05 · ~0° rot err	synthetic GT
Umeyama scale recovery	exact (2.5×)	synthetic GT
Non-overlap registration	rejected → pending alignment	synthetic
Gaussian full compile (hallway)	268 chunks / 73 768 splats / ~7.3 s	measured
Gaussian incremental (1 new voxel)	1 chunk recompiled · 23 unchanged	measured
Gaussian serialization	SHA-256 checksummed · corruption detected	measured
Real LingBot throughput / VRAM / traj error	NOT MEASURED	needs GPU + checkpoint

Unavailable hardware was never converted into a passing benchmark. The synthetic provider proves plumbing and math; LingBot’s perception quality is not claimed.

Read full V12 narrative →

July 17, 2026 47/47 tests

V11, Persistent World Runtime & Evidence Engine

Every modification flows through one runtime: evidence accumulation, six-layer dirty propagation, prioritized incremental recompiles, and a pub/sub event bus. Derived systems consume events; nothing polls, nothing rebuilds globally.

Tests47/47

Events / upload12 028

Dirty chunks31

Dirty layers6

Event types10

Evidence voxels79 439

Runtime Inspector panel with live stats, confidence histogram, and streaming ChunkInvalidated event feed. — **Runtime Inspector, live.** Evidence stats, dirty state, scheduler queue, event totals, confidence histogram, and the streaming event feed.

One GLB upload through the runtime, measured live

Event	Count	What it proves
`RevisionCreated`	1	The upload became a world revision
`CaptureImported`	1	Provenance recorded as a Capture
`ObservationAdded`	5 997	Every voxel backed by an observation
`ConfidenceUpdated`	5 997	Evidence engine ran per observed voxel
`ChunkInvalidated`	31	Only affected chunks marked dirty — never a global rebuild

Honest notes carried from the report: confidence is uniform 1.0 on a fresh single-capture world (the mechanism only diverges with re-observation, conflict, and decay); per-voxel heatmap modes are partial; runtime state is in-memory this build; per-voxel evidence dicts need vectorization for very large worlds.

Read full V11 narrative →

July 17, 2026 33/33 tests

V10, Universal World Engine

The world-OS milestone: universal upload pipeline, persistent components, SVO-routed queries, revision timeline, and full provenance — trace(point) walks voxel → SVO node → Component → observations → Capture → original file, verified live.

Tests33/33

Raycast speedup4.3×

Formats ingested9

Honest-error routes6

GLB upload5 671 vox

World rev1

VWM viewer with the World OS panel: drag-and-drop upload, query search, and live stats line. — **World OS panel.** Upload, query search, and the live stats line: voxels 79 439 · svo 237 697 nodes · components 3 · captures 1 · world rev 1.

SVO-routed queries vs linear baseline (wos_query_benchmark.json, 73 768 voxels)

Query	Linear baseline	Via SVO
Raycast (carries to picking, probes, LOS)	1 654 q/s	7 076 q/s ▲ 4.3×
nearest_occupied	564 q/s (numpy vectorized)	83 q/s ▼ honest loss

Reported plainly: SVO nearest-neighbour currently loses to a vectorized numpy scan at this scale; a Morton neighbour-search optimization is future work. Raycast-shaped queries are where the octree's empty-space skipping pays.

Universal upload pipeline — format coverage

Status	Formats
End-to-end	glTF / GLB · OBJ · PLY mesh · PLY/XYZ point cloud · Polycam zip · VWM voxel npz · image heightmap
Routed, honest error	LAS/LAZ · video (offline SfM job) · USD/USDZ · FBX · Gaussian splats · COLMAP — specific reasons, never fake success

Every successful upload creates a Capture + world revision + incremental SVO update, via an async job queue with progress, ETA, and retry.

Read full V10 narrative →

July 17, 2026 19/19 tests

Sparse Voxel Octree Integration

A derived octree acceleration structure built from the unchanged authoritative VoxelWorld — algorithms audited and adapted from MIT-licensed otaku690/SparseVoxelOctree, no upstream code compiled in. Exact occupancy on all scenes; openings verified open.

Tests19/19

Incremental edit19 ms

Collision95–148k q/s

Sparse topology7.6×

Thin recall1.00

Compression6.7–12×

SVO Depth mode: nested wireframe octree boxes colored by level over the scanned hallway. — **SVO Depth mode.** Node boxes coloured by octree level — huge cells over empty space, fine cells only at the scanned surface (99 121 nodes served, depth 8).

Four benchmark scenes (artifacts/svo/benchmark.json)

Scene	Voxels	Incremental vs full	Collision q/s	Ray q/s	Topology vs dense grid
Hallway (real scan)	73 768	19.6 ms / 1 223 ms	101 607	6 631	0.25× (grid wins)
Thin features	4 519	6.3 ms / 40 ms	148 629	10 597	~1×
Large sparse	19 998	9.1 ms / 1 550 ms	94 939	1 509	7.6× smaller
Dense indoor	13 256	16.1 ms / 123 ms	109 738	4 963	0.12× (grid wins)

Honest limitation, stated in the report: on a small dense surface shell a 1-byte dense grid is more compact in RAM. The SVO wins on sparse scaling, empty-space skipping, hierarchical aggregates, and incremental maintenance — its advantage grows with sparsity. Accuracy verified on every scene: 0 false-occupied, 0 false-empty, thin recall 1.00, doorways open.

SVO Leaves mode: solid green occupied leaf shell of the hallway. — **Leaves.** The occupied leaf shell the collision suite queries.

SVO Interior mode: gold wireframe hierarchy nodes above the leaves. — **Interior.** Aggregate nodes that let traversal skip empty space.

Read full SVO narrative →

July 17, 2026 Error bounded by construction

V9, Error-Driven Adaptive Voxelization

v8's heuristic ("is there detail?") replaced by a guarantee: a region stays one voxel iff it approximates the mesh within a configurable error budget. Appearance is decoupled from geometry via a texture-patch atlas — paintings no longer spawn tiny voxels.

Within 12 mm budget99.9%

Preview voxels8 888

Appearance patches1 534

Plane-snapped1 685

LOD stability0.939

Budget profiles6

Three top-down panels: v8 heuristic voxels, v9 error-driven voxels colored by budget status, and the v9 appearance layer. — **v8 heuristic vs v9 error-driven vs v9 appearance layer.** Middle: grey = within the 12 mm budget, red = subdivided for geometry, blue = topology. Right: coloured voxels carry texture patches that display detail without subdividing.

v8 heuristic vs v9 error-driven (benchmark_v8_v9.json)

Metric	v8 heuristic	v9 error-driven
Error bounded?	no guarantee	99.9% within 12 mm, by construction ▲
Voxels	13 542	29 027 (architectural) · 8 888 (preview, bounded at 55 mm)
Surface Hausdorff	206 mm	196 mm ▼
LOD stability	0.927	0.939 ▲
Appearance patches	0	1 534 (geometry unchanged)
Snapped planar voxels	0	1 685

Independent error terms drive subdivision: geometry (master, p95 residual to best-fit plane), normal (p90), curvature (p90 dihedral), topology (never bridge gaps), thin, silhouette — appearance is metadata only and never triggers subdivision. Robust percentiles replaced max after a single LiDAR speckle forced 67k voxels; budgets are calibrated to measured mesh noise (3.3 mm RMS flat / 9.6 mm edges at 24 cm cells).

Error-budget profiles — tolerances change, algorithm doesn't

Profile	Geometry budget	Voxels	p95 geometry error	LOD stability	Patches
preview	55 mm	8 888	16.2 mm	0.958	1 404
gaming	26 mm	15 399	8.4 mm	0.947	1 249
architectural	12 mm	18 849	6.3 mm	0.956	1 078
museum	6 mm	23 468	4.4 mm	0.967	886

Tighter budget → more voxels. Honest finding: this phone-LiDAR mesh has ~3–5 mm intrinsic surface noise, so budgets below that force fine voxels everywhere — the sensor sets the floor.

Geometry-only mode: blue coarse within-budget blocks, orange fine voxels where error demanded subdivision. — **Geometry only.** Blue = coarse, within budget; orange = subdivided by measured error.

Appearance-layer mode: large voxels displaying projected texture patches. — **Appearance layer.** Same geometry, texture patches supply the visual fidelity.

Read full v9 narrative →

July 16, 2026 Measured fidelity

V8, Feature-Preserving Adaptive Voxelization

The Polycam mesh voxelized by measured curvature: 24 cm blocks on flat walls, 3 cm voxels only at edges, corners, and trim. Every voxel anchored to a measured surface sample; zero invented geometry.

Feature voxels10 231

vs uniform13.6×

Wall RMS2.8 cm

Corner error0.28°

Trim uplift3.2×

Invented0

Three top-down panels: textured Polycam mesh, uniform 3 cm voxels, and feature voxels with blue coarse walls and red fine edges. — **Mesh vs uniform vs feature voxels, top-down.** 138 910 undifferentiated 3 cm blocks (middle) become 10 231 feature voxels (right): blue 24 cm on flat walls, red 3 cm at edges and frames.

Uniform @3 cm vs feature voxels, same surface

Metric	Uniform @3 cm	Feature voxels (v8)
Voxel count	138 910	10 231 ▼
Size levels	1 (3 cm)	4 (3 / 6 / 12 / 24 cm, mean 9.5 cm)
Level breakdown	—	2 993 @24 · 234 @12 · 521 @6 · 6 483 @3
Fine voxels on high-curvature trim	12.5% (baseline)	40% ▲
Surface-anchored voxels	100%	100% (0 invented)

Fidelity measured from the mesh + emitted voxels, nothing assumed: 15 fitted planes flat to 2.8 cm RMS; 48 wall–wall corners at 89.7° (0.28° mean error, 1.48° max) with 51% fine voxels at the seams. Tuning insight: on noisy phone-LiDAR, curv_flat ≈ 0.35 — 0.10 left 93% at the finest level, 0.55 over-coarsened the detail away.

Voxel-size mode: blue large flat-wall voxels, orange and red fine voxels at features. — **Voxel-size mode.** Blue coarse walls, red fine detail — resolution allocated where measured.

Feature-density mode: green quiet surfaces, red feature-rich regions. — **Feature-density mode.** The measured curvature signal that drives subdivision.

Read full v8 narrative →

July 16, 2026 Live in the viewer

The Polycam Scan in the VWM Viewer

The Polycam capture served standalone by the authoritative-world server — extracted surface by default, 73 768 literal voxels one click away, all render modes and per-voxel inspection working unchanged.

Voxels73 768

Mesh verts30 974

Texture8192²

Scale12.3 m

Grid106×55×246

FPS63

Fly-mode view down the voxelized Polycam hallway in real-color mode with the voxel inspector open. — **Literal Voxels, real color, fly mode.** The Polycam scan voxelized at 5 cm, with per-voxel inspection (class SURFACE, provenance `observed_surface`, metric position) at 64 FPS.

Occupancy mode: all voxels classified SURFACE, stats panel showing FREE 0. — **Occupancy mode.** All-SURFACE, FREE 0 — this direct import is surface-only; carving is the named next step.

Wireframe mode exposing the voxel structure of the hallway. — **Wireframe mode.** Voxel structure exposed down the corridor, 65 FPS.

Direct Polycam import, measured

Metric	Value	Note
Extracted surface	30 974 verts / 49 475 faces	The real Polycam mesh with its 8192² texture, default view
Literal voxels	73 768 @ 5 cm	Grid 106×55×246, correctly metric (12.3 m hallway)
Render performance	63–65 FPS	Fly mode through the voxel hallway, all modes instant-switch
Free space	0 (surface-only)	No carving run on this import yet — Walk has nothing to stand on; next step

Read full narrative with all mode screenshots →

July 16, 2026 Multi-source, measured

Reality-Capture v7, Polycam Multi-Source Fusion

A LiDAR-metric Polycam scan of the same hallway fused into the persistent authoritative VoxelWorld as another observation source. Registration is computed, not hand-tweaked; fusion accumulates, never overwrites.

ICP median0.030 vox

Scale error4.9×

Confidence+15%

Coverage+528

Cross-confirmed2 818

Merge stability1.0

Registration: Polycam → video-built world (PCA coarse + robust ICP)

Metric	Value	Meaning
ICP median residual	0.030 voxels	Sub-voxel — two independent reconstructions are provably the same hallway
ICP mean residual	0.056 voxels	Robust alignment throughout
Implied metric scale	1 unit = 4.91 m	Monocular world was compressed ~4.9× (real hallway 12.3 m, video had ~2.5 m)

The long-standing monocular scale error — flagged honestly since v2 — is now exposed and quantified by a metric source, exactly the value of a second heterogeneous sensor.

Viewer source mode: blue video-only voxels, orange Polycam-only voxels, green voxels observed by both sources. — **Source mode.** Blue video-only (6 418), orange Polycam-only (528), green both-agree (2 818) — the green mass is cross-source validation.

Fusion: video-only world vs after Polycam merge

Metric	Video only	After Polycam
Matter voxels	10 014	11 044 ▲
Mean confidence	6.84	7.90 ▲
Mean obs-count / voxel	1.64	1.86 (max 3) ▲
Voxels strengthened	—	2 924
Geometry stability on merge	—	1.0

Polycam import: 30 974 verts / 49 475 faces, metric 12.28 × 5.28 × 2.72 m, 8192² PBR texture baked to vertex colors, 120 000 sampled observations. Not yet done (honest): applying the metric-scale correction (re-voxelization is a deliberate, reversible next step), metric-density surface re-extraction, semantic re-estimation, live Luanti walkthrough.

Read full v7 narrative →

July 16, 2026 Representation ≠ storage

Reality-Capture v5, Extracted-Surface Reconstruction

The VoxelWorld/HAV is unchanged — it stores the TSDF field. The viewer now renders the continuous surface extracted from that field (marching cubes) instead of storage cubes. Extraction runs only where the field was observed; zero geometry invented.

Fidelity (EFS)0.962

Vertices7 809

Faces14 452

Manifold edges0.924

Area12.8 m²

Gaps closed4

Same authoritative TSDF field rendered as literal voxel cubes (left) and as the smooth extracted surface mesh (right) in the VWM viewer. — **The decisive comparison.** Same authoritative TSDF field: literal voxels (left) vs the extracted marching-cubes surface (right), vertex colors from observed appearance. A pure representation change; the world is untouched.

v4 (literal voxels) vs v5 (extracted surface), same field

Signal	v4 literal voxels	v5 extracted surface
Representation rendered	11k+ storage cubes	7 809-vert / 14 452-face mesh
Manifold-edge fraction	n/a	0.924
Surface area	—	12.8 m²
Environment Fidelity Score	—	0.962
Look	blocky blob	smooth architectural surface
Authoritative world changed	—	no (derived view only)

Viewer default is now Extracted Surface with a one-click toggle to Literal Voxels, both from the same field via /api/surface. Honest limits: walls are smooth-but-wavy and corners rounded (marching cubes) — plane-snapping + dual contouring is the named next step; fine features need a capture that observed them. Nothing was faked.

Read full v5 narrative →

July 16, 2026 No invented geometry

Reality-Capture v4, Recognizable Adaptive Reconstruction

Same authoritative v3 world, made recognizable: conservative hole-filling (every fill provenance-tagged INFERRED, free air never touched) plus adaptive per-voxel detail levels. Viewer upgraded with Walk mode and three new render modes.

EIS0.743

vs v3+47%

Holes427 → 0

Free voxelsunchanged

FPS63

Walk modeyes

v3 → v4, same authoritative world

Signal	v3 hollow	v4 adaptive
Environment Identity Score	0.507	0.743 ▲
Enclosed holes	427	0 ▼
Hole-freeness	0.21	1.0 ▲
Matter voxels	10 763	11 362 (+599 INFERRED, +5.6%)
Free voxels (air)	4 850	4 850, unchanged
Detail levels (coarse / medium / fine)	—	9 067 / 1 893 / 402

Fill rule: an UNKNOWN voxel enclosed by observed matter on ≥4 of 6 faces is a gap in a real surface, not open space. Free air never qualifies, so doorways and unseen rooms stay open. Filled-voxel appearance is interpolated from neighbouring observed colour.

Side-by-side viewer screenshots: v3 with speckled reconstruction holes, v4 with continuous filled surfaces. — **v3 (left) vs v4 (right), real-color mode.** 427 enclosed holes closed; the shell reads as a continuous surface. Free space untouched.

LOD heatmap: blue coarse, green medium, red fine detail on edges and corners. — **LOD heatmap.** Fine detail (red) concentrates on edges/corners.

Inferred-fill mode: pink voxels mark every hole-fill addition. — **Inferred-fill mode.** Every fill visible in pink — auditable provenance.

Read full v4 narrative →

July 16, 2026 Working, measured

QA Loop + Three.js VWM Viewer

Automated QA over the v3 world (3 iterations, measured fixes) plus the official browser engineering client: a Three.js viewer/editor over the single authoritative VoxelWorld, which owns no world state of its own.

FPS63

Voxels15 613

Acceptance13/15

Render modes9

Edit scope3×3×3

Global rebuildfalse

The VWM Three.js viewer rendering the reconstructed hallway in real-color mode with mode buttons, voxel inspector, and live stats panel. — **The viewer on the v3 world, real-color mode.** Live stats bottom right: 61 FPS, 15 613 voxels, SURFACE 8 047 / FREE 4 850. Bottom left: the click-to-inspect panel with one voxel's authoritative metadata.

QA loop, defects fixed with before/after measurements

Iteration	Defect	Measured result
1 → 2	Spawn out-of-bounds (12,22,−2); 146 voxels dropped; connectivity 0.0	Unified shift → 0 dropped, connectivity 0.992, spawn in-bounds ▲
3	Spawn had no floor (only 9% of free voxels have observed floor below)	`find_standable_spawn` → spawn on observed floor with head clearance ▲

Remaining defects are capture-data limitations (floor coverage, absolute scale), not code bugs. Full log: v3/qa/ENGINEERING_LOG.md.

Viewer measurements

Metric	Value	Note
Render performance	63 FPS	15 613 instanced voxels, single draw call, no culling needed
Mode switching	instant	9 modes, in-place instance recolour, no geometry rebuild
Edit round-trip	local 3×3×3	POST → authoritative mutation → revision++ → dirty region; `global_rebuild=false`
Camera movement	0 server calls	Pure client render — the zero-regeneration invariant at the viewer level
World state in client	none	All state from `/api/world\|meta\|stats\|edit`; Three.js r128 vendored, no CDN
Luanti edit-reflection	partial	Edits hit the shared world Luanti reads; Luanti not connected this session

Read full viewer narrative →

July 16, 2026 Proven, not asserted

Reality-Capture v3, Hollow Interior + Live Import

Same canonical video, pipeline extended with an occupancy stage (FREE / SURFACE / OCCUPIED / UNKNOWN), a free-space connectivity graph, and a working live VoxelWorld/HAV import. The rule a room is free space bounded by voxelized matter is measured, not asserted.

VerdictHOLLOW

Centreline free1.0

Flood-fill reach99.2%

False-occ air0

HAV import9.9 s

Free spawnyes

v1 → v2 → v3, same canonical video

Metric	v1 point-splat	v2 TSDF	v3 hollow
Reprojection L1	0.0757	0.0543	0.0543 (same surfaces)
Surface voxels	137 685	9 612	8 047 ▼
FREE / OCC / UNKNOWN	none	4 416 / 9 612 / 24 283	4 850 / 10 763 / 29 387
Hallway interior	solid shells	unvalidated	HOLLOW, centreline free 1.0
False-occupied air	—	—	0
Free connectivity	—	—	0.992 (1 region)
Live VoxelWorld import	stalled	not attempted	10 763 vox / 15 177 leaves / 9.9 s ▲
Spawn in free space	no	no	yes ▲

The HAV stall fix was structural, not a tweak: TSDF yields a compact shell over a ~31³ span versus v1's 137k splats over a 1 000-unit span, so live import now runs at 1 090 fills/s. Wall thickness (2 716 inferred voxels) extrudes ≤2 voxels behind surfaces, only into UNKNOWN, never into FREE, and is provenance-tagged as inferred.

Occupancy slice: green flood-filled free interior bounded by tan matter voxels. — **Flood-fill proof.** Green FREE interior (99.2% reached from the camera, one region) bounded by tan matter.

Side occupancy slice showing the thin matter shell and unobserved space left unknown. — **Side slice.** Thin matter shell with bounded inferred thickness; unseen space stays UNKNOWN.

Not run this pass (honest)

Item	Status	Note
Live Luanti client walkthrough	not run	Needs interactive human session; world imported with free-space spawn, ready
Restart persistence round-trip	not run	Export-import round-trip deferred
Held-out appearance / exposure norm	not this pass	Iterative refinement, not blocked by the model
Portals / multi-room segmentation	not emitted	Free-space graph exists; this capture is a single hallway region
Absolute metric scale	best-effort (~2 m)	Compressed monocular cloud; richer capture (orbit / AR poses) is the lever

Read full v3 narrative →

July 16, 2026 v1 vs v2, measured

Reality-Capture v2, TSDF Fusion

Same canonical phone video as the walkthrough below; fusion upgraded from point-splat to true TSDF volumetric fusion with free-space carving, plane-based semantics, and a 42″-door metric anchor. Deterministic run, same authoritative pipeline order.

Reprojection L10.0543

vs v1−28%

Obs / voxel15.2

Watertight0.99

Surface voxels9.6k

Scale0.0098 m/u

v1 (point-splat) vs v2 (TSDF), same video

Metric	v1 point-splat	v2 TSDF
Reprojection photometric L1	0.0757	0.0543 ▼
Observations accumulated / voxel	1.0	15.2 ▲
Free / occupied / unknown voxels	not classified	4 416 / 9 612 / 24 283
Watertightness (backed fraction)	fan shells	0.99
Surface voxels	137 685	9 612 ▼
Metric scale	arbitrary	0.0098 m/unit via 42″ door (best-effort)
Render coverage	0.99	0.49 ▼

TSDF weighting merges duplicate shells and removes floaters: 14× fewer voxels that reproject 28% more accurately. Coverage drop is the honest cost of keeping only confident voxels; L1 improves anyway. Absolute scale passes its sanity gate but the monocular cloud is geometrically compressed — treat scale as approximate.

Side-by-side of v1 point-splat reconstruction and v2 TSDF reconstruction of the same hallway view. — **v1 vs v2, same view.** Left: 137k single-view splats stacked into fan shells. Right: 9.6k TSDF voxels averaging ~15 observations each — sparser on screen but geometrically truer.

Side view comparing v1's thick fan of duplicate shells with v2's single flat coherent wall surface. — **Where the 14× reduction came from.** Side view: v1's duplicate single-view shells sweep out a thick fan (left); v2's TSDF weighting collapses them into one coherent wall (right).

New capabilities in v2

Capability	Measured	Note
Free-space carving	4 416 free / 24 283 unknown	Unseen space never filled solid; confidence + provenance per voxel
Watertightness	0.99 backed fraction	99% of surface voxels backed by ≥2 occupied neighbours
Plane semantics	8 RANSAC planes	wall / floor / ceiling by geometry, inlier-count confidence, wall thickness via parallel planes
Metric anchor	42″ door, conf 0.98	Outlier-robust percentile width + scene-size sanity gate (2–40 m)
Voxel confidence	mean weight 7.27	Real accumulated weight everywhere (v1: single unweighted observations)

Honest limits in v2

Limit	Measured	Fix
Absolute scale approximate	~2 m scene, too small for a real hallway	In-frame metric measurement (ARKit depth / ruler)
Render coverage	0.49 (was 0.99)	Back-and-forth / orbit pass spreads the cloud
Scale-consistency CoV	0.576 unchanged	ARKit/ARCore poses remove SfM scale drift
No loop closure	one-way walk, no revisit	Detected + reported, not invented; capture with revisit

Read full v2 narrative →

July 16, 2026 End to end, measured

Real-Video Reconstruction, Phone Walkthrough

Real phone walkthrough of a home (laundry area to hallway), 464×848, 19.4 s, 582 frames. No synthetic data. Full pipeline in 13.7 s (excl. persistent-world build). v2.4 not trained.

Reprojection L10.076

Coverage0.99

SfM tracked24/24

Scale CoV0.576

HAV leaves38.3k

Processing13.7 s

Phone walkthrough vs BRIO webcam (same pipeline)

Metric	BRIO webcam (Jul 13)	Phone video (Jul 16)
Cross-frame scale CoV	≈30	0.576 ▼
Keyframes tracked	14	24 / 24, 0 lost ▲
Mean pose confidence	0.89	0.897 ▲
Alignment edges / samples	56 / 16 626	156 / 19 170 ▲
Surface voxels	6 154	137 685 ▲
Reprojection L1 / coverage	not measurable	0.076 / 0.99

Same pipeline, better capture motion. Forward walking fixed what tiny webcam baselines could not. Three bugs fixed in this run: gauge-normalisation (negative depth), NumPy 2.x .ptp(), and a slow per-voxel render loop (vectorised).

Three keyframes showing original phone frame beside reconstructed voxel render from the same recovered pose. — **The reprojection benchmark, visually.** Each pair: real frame (left) vs reconstruction rendered from the same recovered pose (right). This is the source of the L1 0.076 / coverage 0.99 numbers above.

Full pipeline run (reconstruction_report.json)

Stage	Metric	Value
Capture	Video frames / extracted / keyframes	582 / 155 / 24
SfM	Frames tracked / mean confidence	24 / 24, 0.897
Alignment	Overlap edges / triangulated samples	156 / 19 170
Alignment	Solved cross-frame scale CoV	0.576
Fusion	Point-cloud points	120 000
Fusion	Surface voxels (confirmed)	137 685 (311)
Components	Proposed (connected-component)	9
Reprojection	Photometric L1 vs original frames	0.076
Reprojection	Coverage	0.99
Persistent world	HAV leaves materialized	38 284
Runtime	Reconstruction processing	13.7 s

Bugs fixed in this run

Bug	Symptom	Fix
Gauge normalisation	Aligned depth went negative	SfM poses define scale; no gauge normalisation
NumPy 2.x `.ptp()`	API break under NumPy 2	Replaced with portable range call
Per-voxel render loop	Slow novel-view / reprojection path	Vectorised render path

Honest weaknesses in this run

Weakness	Measured	Fix
Single-view voxels	137 374 provisional vs 311 confirmed	Orbit / back-and-forth capture
Residual scale drift	CoV 0.576	Metric anchor or AR poses
HAV materialization cliff	fills stall on large SfM span; in-process deadlock	Separate process; reopen from `voxel_grid.npz`
No free-space carving	disabled for speed	Re-enable + TSDF fusion
Weak component labels	extent-ratio heuristic, conf 0.4	Real semantic segmenter

Read full reconstruction narrative →

July 13, 2026 Partial milestone

Reality Capture, Real Webcam Scan

Logitech BRIO, 18 s / 425 frames, real bedroom. End-to-end: keyframes → visual SfM → depth-anything → alignment graph → fusion → real RGB → novel view. v2.4 not trained.

Frames425

Keyframes14

SIFT matches540

Surface voxels6154

Align samples16.6k

Processing10.3 s

Synthetic vs real: domain gap confirmed

Input	SIFT matches	E-matrix inliers	DA depth (qualitative)	Multi-view scale
Synthetic voxel render	0	0	Poor (OOD)	CoV 0.47–0.57
Real BRIO bedroom	540 ▲	516 ▲	Coherent single-view	CoV ≈ 30 (small baseline)
Controlled sensor depth	—	—	N/A (sensor)	Stable (0.918 recall)

Real photos unlock depth-anything and feature geometry. Multi-view metric consistency needs larger baselines, AR poses, or a metric anchor.

Real pipeline run (real_scan_report.json)

Stage	Metric	Value
Capture	Total frames / keyframes	425 / 14
SfM	Tracked frames / mean confidence	14 / 0.89
Alignment	Overlap edges / triangulated samples	56 / 16 626
Alignment	Solved scale CoV	30.4
Fusion	Surface voxels (0.04 m)	6 154
Fusion	Confirmed / provisional	3 200 / 2 954
Runtime	End-to-end processing	10.3 s

Single-view vs multi-view (honest)

Mode	Geometry quality	Scale consistency	Why
Single-view depth	Coherent (closet, dresser, couch)	N/A (one frame)	Sidesteps cross-frame scale coupling
Multi-view fusion	Rough / scattered	CoV ≈ 30	Tiny baselines → noisy SfM + depth scale

Read full real capture narrative →

July 13, 2026 Scan pipeline

Depth Source Comparison

Controlled room, 20-frame trajectory, identical poses and fusion. Only depth source varies. Monocular runs use robust pooled calibration (not frame-0 anchor).

Sensor recall0.918

Sensor backproj0.17

DA3 recall0.564

DA2 recall0.449

DA3 scale CoV0.47

Calib fix2.5×

3-way depth benchmark

Trends vs RGB-D / LiDAR sensor ground truth. Lower backproj/abs err and scale CoV are better; higher recall and IoU are better.

Fusion quality by depth source

Depth source	Backproj err	Surface recall	IoU	Abs depth err	Scale CoV
RGB-D / LiDAR (sensor)	0.17	0.918	0.636	—	—
DA3-SMALL (relative)	1.99 ▲	0.564 ▼	0.229 ▼	1.99	0.47
DA2 Metric Hypersim Small	2.37 ▲	0.449 ▼	0.178 ▼	2.37	0.57 ▲

Source: checkpoints/vwm_scan/DEPTH_SOURCE_COMPARISON.md. DA3 beats DA2 on this synthetic test; neither approaches sensor without reliable depth.

Monocular vs monocular (DA3 vs DA2)

Metric	DA3 relative	DA2 metric	Better
Surface recall	0.564	0.449 ▼	DA3
Backproj error	1.99	2.37 ▲	DA3
IoU	0.229	0.178 ▼	DA3
Cross-frame scale CoV	0.47	0.57 ▲	DA3

Calibration methodology (caught bug)

Calibration method	Backproj err	Abs depth err	Recall
Frame-0 anchor (buggy)	6.76	8.10	0.39
Robust pooled (fixed)	2.37 ▼	2.37 ▼	0.449 ▲

Frame 0 was a scale outlier (8.1 vs typical ~3.9). Pooled calibration cut error >2.5×. Lesson: never anchor scale on one frame.

Surface recall, visual comparison

Sens. 0.918

DA3 0.564

DA2 0.449

Read full depth scan narrative →

July 12, 2026 v2.3, not promoted

Appearance v2.3

Data engine + rare geometry training. Headline: 100k stress test found and fixed a fused-renderer memory leak. v2.2 stays production default.

RSS (100k fixed)3056 MB

RSS growth0 MB

Dataset2757

Rare thin_surv.0.965

Std. L1 Δ+45%

Defaultv2.2

100k stress test: memory leak found and fixed

Stage	Result
Pre-fix RSS (45 min)	>9700 MB (monotonic climb)
Root cause	`_repair` unbounded keyframes (~57 KB/frame); `_build` capped at 24
Fix	Shared `_insert_keyframe()` with eviction; both paths use it
8k regression	Keyframes 24/24, RSS +1.4 MB
100k post-fix (52 min)	RSS flat 3056 MB (0 growth)
Compiler invariant	0 model/compiler calls (unaffected)

Source: fused_renderer_leak_fix.json, proof100k_fixed/memory_fix_validation.json. Applies to production v2.2 fused renderer.

Data engine (v2.3 infrastructure)

Component	Result
Dataset scale	1,050 → 2,757 obs (1,707 new rare)
Cycles factory	1.79 obs/s (2.24× speedup)
Active planner	Avoids 90% of Cycles work
Rare generators	25 thin-geometry classes + 17 material families
10k target	Partial (2,757 / 10,000, resumable)

v2.2 vs v2.3 benchmark

Frozen standard held-out (production comparison) and rare-geometry held-out (where v2.3 wins). Trends vs v2.2.

Standard held-out (VWMScore v3)

Metric	v2.2 (prod)	v2.3 continue	v2.3 scratch
Photographic L1	0.018	0.026 ▲	0.041 ▲
Edge IoU	0.691	0.740 ▲	0.872 ▲
StructuralIdentityScore	0.698	0.786 ▲	0.790 ▲
Light-invariance variance	0.011	0.014 ▲	0.012 ▲
Edit drift	0.00107	0.00133 ▲	0.00159 ▲
Gate-fail rate	0.0	0.0 —	0.0 —
VWMScore v3	0.760	0.765 ▲	0.773 ▲

VWMScore edge driven by structure term; standard L1 regresses +45%. Scratch over-sharpens (edge-IoU 0.872). Source: vwmscore_v3_frozen.json.

Rare / compositional held-out (v2.3 wins)

Metric	v2.2	v2.3 continue
Photographic L1	0.055	0.054 ▼
Edge IoU	0.507	0.626 ▲
StructuralIdentityScore	0.743	0.819 ▲
thin_survival	0.853	0.965 ▲
Rare gate failures	0.05	0.0 ▼
Material adherence (err↓)	0.313	0.366 ▲

Rare geometry gains are real; material adherence regressed (training pool 74% rare geometry).

Promotion verdict

Criterion	v2.3 continue
Rare geometry improved	✓ (thin_survival 0.965)
Standard photographic L1	✗ (+45% regression)
Clean win on aggregate	✗ (structure-for-photo trade)
Memory leak fixed (renderer)	✓ (benefits v2.2 prod)
Promoted	✗ — v2.2 stays default

July 12, 2026 v2.2, production default

Appearance Model v2.2

Production default. Structural fidelity without sacrificing photographic appearance. Memory leak fix from v2.3 applies to this renderer.

VWMScore v30.773

Edge IoU0.691

Structural ID0.698

Thin survival0.841

Gate failures0

10k frames0 runs

Architecture fix summary

Change	Effect
Decoupled intrinsic decoder	Stops light-invariance from flattening final-head features
Structure-guided final decoder	G-buffer boundary map (never Cycles RGB) guides edges
Anti-blur + bleed + structure-aware locality	Edges preserved; boundaries never flattened by locality loss
Root cause	intrinsic~lightinv cosine +0.89 on shared trunk (amplification imbalance)

Invariant proof (10,000 camera frames)

Test	v2.2 result
Neural model invocations (camera-only)	0
Compiler invocations (camera-only)	0
Median FPS	48.6
Modules import clean	85

4-way benchmark (VWMScore v3)

Frozen benchmark on same held-out split. v2.2 promoted (Pareto-dominates v2 on all six axes). Trends vs v2 (was production).

Quality, structure, and aggregate score

Metric	v2 (was prod)	v2.1	v2.2
Photographic L1 → Cycles	0.019	0.014 ▼	0.018 ▼
Edge IoU	0.580	0.531 ▼	0.691 ▲
StructuralIdentityScore	0.460	0.341 ▼	0.698 ▲
Light-invariance variance	0.025	0.013 ▼	0.006 ▼
Unchanged-region drift	0.00124	0.00084 ▼	0.00115 ▼
Geometry-gate failure rate	0.0	0.01 ▲	0.0 ▼
VWMScore v3	0.670	0.650 ▼	0.773 ▲

Source: benchmark_final/benchmark_full.json. Trends in v2.1/v2.2 columns vs v2 (previous production default).

Photographic L1, version comparison

v2 0.019

v2.1 0.014

v2.2 0.018

Structural torture house

Thin geometry: chair legs, railing balusters, cables, handles, keyboard keys, vent slats. Trends vs v2.

Hard-structure survival

Metric	v2	v2.1	v2.2
StructuralIdentityScore	0.568	0.511 ▼	0.768 ▲
thin_survival	0.543	0.513 ▼	0.841 ▲
Edge IoU	0.425	0.368 ▼	0.525 ▲

No edge hallucination: boundary precision held at 0.95 on standard set (aggressive-loss candidates painted fake edges).

July 12, 2026 v2.1, not promoted

Appearance Model v2.1

16 held-out scenes · 96 views · edit-locality + relight supervision · VWMScore v2 benchmark. Best photographic L1; v2 wins aggregate score and stays default.

Photographic L10.014

VWMScore v20.833

Observations1050

Edit drift0

Light variance8×10⁻⁶

Gate failures1%

New capabilities measured in v2.1

Capability	v2.1 result	Notes
Edit-locality loss	wired	EMA 0.0005; 270 edit pairs
Unchanged-component drift	0	8 sequential edits torture test
Intrinsic light invariance	8×10⁻⁶	Matched Cycles relight groups
Dataset audit	1050/1050	0 rejected
In-Luanti capture	not re-run	path wired; headless proof only

Training & dataset (v2.1)

Metric	v2.1	v2
Total observations	1050	600
Standard pairs	600	600
Edit pairs	270	-
Relight groups	180	-
Training steps	5000	3000
Training time	~180 s	105 s
Peak VRAM	215 MB	215 MB
Batch mix	40/30/30	standard

Historical: v1 / v2 / v2.1 (VWMScore v2)

Earlier unified evaluation. Superseded by VWMScore v3 benchmark above.

Quality, geometry, and aggregate score (archived)

Metric	v1	v2	v2.1
Photographic L1 → Cycles	0.033	0.019 ▼	0.014 ▼
Edge IoU	0.635	0.580 ▼	0.531 ▼
Geometry-gate failure rate	0.0	0.0 —	0.01 ▲
Intrinsic light variance	-	0.0	8×10⁻⁶ ▼
Unchanged-component drift	-	0	0 —
VWMScore v2	0.717	0.871 ▲	0.833 ▼

Trends vs prior column. v2 won production default at this stage; later superseded by v2.2.

July 12, 2026 v2, superseded

Appearance Model v2

Former production default. Promoted after v1; superseded by v2.2.

VWMScore v20.871

Photographic L10.019

Surface variance0.0

Gate failures0

Camera FPS209

Quality benchmark, deferred vs v1 vs v2

Metric	Deferred	v1	v2
Photographic L1 → Cycles	0.316	0.033	0.019 ▼
Edge IoU (geometry)	-	0.635	0.580 ▼
Multi-view surface variance	-	-	0.0
Geometry-gate failure rate	-	0.0	0.0
VWMScore v2	-	0.717	0.871 ▲

Lower L1 is better. v2 wins aggregate VWMScore v2 on unified benchmark. See 3-way compare for v2.1.

Photographic L1, visual comparison

Def. 0.316

v1 0.033

v2 0.019

Training & dataset

Metric	v2	v1 (ref)
Training pairs	600	50
Scenes	100	10
Scene classes	10	1
Training steps	3,000	600
Training time	105 s	13 s
Peak VRAM	215 MB	395 MB
Parameters	1.45M	compact U-Net
Silhouette IoU (parity)	0.97	0.97
Pairs rejected (parity)	0	-

Runtime invariant, model runs only at compile time

Test	v2 result
Model runs during house compile	17 (once per view)
Model runs, 1,000 camera-only frames	0
Median FPS (camera-only)	209
p95 frame time	5.5 ms
Test suites green	all
Modules import clean	83

Failed iteration (honest), intrinsic-only v2 before dual-head fix

Metric	v1	Intrinsic-only v2
VWMScore	0.792	0.584
Photographic L1	0.033	0.135
Outcome	-	not promoted

Dual-head architecture was added and re-benchmarked; promoted v2 is the row in the quality table above.

July 12, 2026 v1, Superseded

Appearance Model v1

First real checkpoint, 2 held-out scenes (1008, 1009) · narrow procedural distribution · proof that the full pipeline works.

L1 improvement21×

Held-out L10.016

Silhouette IoU0.97

Edge IoU0.72

Peak VRAM395 MB

Camera FPS242

Quality benchmark, narrow held-out (scenes 1008, 1009)

Metric	Deferred	v1 model	Change
Photographic L1 → Cycles	0.343	0.016 ▼	21× better
Views improved	-	100%	-
Silhouette IoU (camera parity)	-	0.971	-
Edge IoU (geometry)	-	0.72	-
Geometry-gate failure rate	-	0.0	-

v1 channel ablation later showed only deferred RGB contributed, v1 learned tone/GI, not structure. Source: vwm_appearance_v1.report.json

Photographic L1, visual comparison (v1 held-out)

Def. 0.343

v1 0.016

Training & dataset

Metric	v1
Training pairs	50
Scenes × views	10 × 5
Held-out scenes	1008, 1009
Training steps	600
Training time	13 s
Peak VRAM	395 MB
Resolution	160²

Runtime invariant

Test	v1 result
Model runs during compile (6 views)	6
Model runs, 500 camera-only frames	0
Camera-only FPS	242
1,000-frame proof median FPS	89
Test suites green	46
Modules import clean	82

Cross-milestone comparison

Headline metrics across checkpoints. Current default: v2.2. Latest evaluated: v2.3 (not promoted).

Headline metrics at a glance

Metric	v2.2	v2.3	Best
Photographic L1 (standard)	0.018	0.026 ▲	v2.2
thin_survival (rare)	0.853	0.965 ▲	v2.3
Edge IoU (rare)	0.507	0.626 ▲	v2.3
VWMScore v3 (standard)	0.760	0.765 ▲	v2.3 (marginal)
100k RSS leak	pre-fix >9.7 GB	fixed 3056 MB flat	v2.3 fix
Training observations	1050	2757 ▲	v2.3
Production default	✓	not promoted	v2.2
Compiler invariant	✓	✓	both

v2.3 improves rare geometry and fixes long-session memory; v2.2 kept for standard photographic quality. See v2.2 vs v2.3 for full tables.

Read full milestone narratives →