# Dataset coverage and remaining depth

The [generated catalog](../skills/3dviz-pro-max/data/catalog-summary.json) is the count source of truth. A populated direction means it has authored recipes; it does not mean all subjects within it are covered or rendered.

## How the dataset is organized

Recipes describe a complete visualization task. Reusable knowledge describes a decision that can transfer between tasks. Sources support bounded claims. Public evidence records authorship, source checks and actual artifact observations separately.

The physical storage uses manifest layout version 2: each `recipe.<slug>` is one object at `data/recipes/<slug>.json`, and each `knowledge.<slug>` is one object at `data/knowledge/<exact-knowledge-kind>/<slug>.json`. Sources and directions remain arrays. Recipe paths do not encode a primary direction; `direction_ids` is unordered membership used to generate every applicable direction view. The original planning categories remain useful content axes; they do not require a separate loader or skill for each category.

| Content axis | Current representation | Remaining depth to assess |
| --- | --- | --- |
| Topic directions | Direction registry, unordered recipe membership and generated direction views | Breadth within each direction, advanced and mixed-domain cases |
| Styles and themes | 37 style profiles (33 with numeric defaults, five of them named looks) and 18 theme profiles in separate knowledge families | Useful combinations, object-specific treatment and culturally grounded historical claims when needed |
| Representation | 12 geometry profiles, recipe geometry and inspection patterns | Broader subject-specific use and observed geometry checks |
| Object construction | Object archetypes and recipe object craft | More detailed object families, articulation and functional assemblies |
| Materials and lighting | Nineteen material profiles, 22 lighting profiles (eight of them mood rigs with numeric defaults) and recipe guidance | Observed optical conditions, encoding and controlled comparisons |
| Motion and interactions | State contracts, 19 physical behaviors, 12 motion profiles and 11 interaction profiles and inspection patterns | Observed input, accessibility and solver handoffs |
| Composition and labels | Eight presentation profiles, eight composition profiles and recipe guidance | Observed framing, labels and responsive behavior in concrete artifacts |
| Tool and solver integration | Twenty-eight tool adapters, physical contracts and 12 output profiles | Executed import/conversion paths, richer source formats and host integration |
| Reasoning and validation | 99 conditional reasoning rules, 75 domain-validation records, profile records and recipe invariants | Observed application of scoped recommendations and domain checks |
| Reusable geometry | 22 kit blueprints, each a record plus a module under `templates/kits/` plus four captures under `evidence/kits/` | More subjects beyond a village, and a second look family the kits were tuned against |

The blueprint family is the one place where the dataset ships runnable geometry rather than guidance about geometry. Each of the 22 records names an asset under `templates/kits/`, its tunable `params` with the range that was proved, its named `sockets`, a three-band `detail_ladder`, a `footprint_m`, a `poly_budget` and the `fits_looks` ids that `design-context.py --blueprints --look <id>` filters on. Coverage is a village at a diorama scale: six buildings, eight props, four plants and rocks, two walking creatures and two Blender-authored GLB heroes, proved at 1280x720 with ACESFilmic tone mapping at exposure 1.0 and no post-processing. What that does not cover: vehicles, interiors, machinery, any non-European vernacular, and any look outside the hand-painted and stylized families the palettes were tuned in. The captures record what one headless browser drew — they are not evidence that a blueprint looks good, and nothing in the family certifies historical construction, botanical accuracy or structural behaviour. Nine primitives and `layout/village-layout.js` ship as modules with no record, because neither has a capture of its own to point at. Every one of the 22 records declares a `tiers` block, and each tier is either proved by its own captures or says why not. All 22 declare T0, the blockout. The procedural **T2** tier is proved on **18**; the four that are `not-proved` state a measured reason — `quadruped-walker` and `tree-conifer` because a fabric weave on fur and a bark map on needles read worse than T1, `stone-guildhall` and `iron-lantern` because they are glTF records whose geometry never reaches `buildKit`. **T3** is proved on **2** (`timber-cottage`, `watermill`) and declared `missing` on the other 20. **T4** is `missing` on all 22.

Quality has a stated ceiling. The procedural T2 tier is stylized-game quality: canvas-drawn textures and vertex-baked occlusion, built in the browser with no shipped files. T3 adds what a runtime cannot compute — bevelled arrises, a UV atlas and light baked into textures — and the two heroes above ship at that tier, baked in Blender from the modules' own geometry. Film or photoreal work needs authored T4 assets with hand-built topology and edge flow, and **none ship here**: T4 is a schema slot every record declares missing. A tier a record does not prove is a tier it does not claim.

Inline invariants are not automatically counted as standalone validation records. Source count is not a quality score. A recipe with a numerical formula still needs an implemented solver or explicit analytic evaluation before it becomes a simulation.

## Completion accounting

The canonical catalog currently contains **223 recipes, 440 knowledge records in 18 populated families, 451 sources and 3,599 public evidence events**. All 24 directions have at least six recipes. The four original depth batches remain a historical checkpoint: 72 additional recipes, 24 geometry/lighting/output profiles, 96 conditional reasoning rules and 72 domain-validation records. Their curation records distinguish source reads and editorial checks from runtime checks, which were not run for those batches.

Generated pages cover all 24 directions and 18 populated knowledge families, supporting progressive reading. Retrieval indexes `applies_when` as well as subject terms. Current `reasoning-rule` and `domain-validation` counts are 99 and 75 respectively; the completed 96-rule retained budget remains a historical target, while validation profiles and inline recipe checks remain separate guidance.

The original authoring budgets are 144 purposeful recipes, 96 routing/composition recommendations and 72 domain-validation records. All three budgets are complete as authored and source-reviewed data. The final batch of the original recipe budget added 72 records, 71 sources and 380 evidence events, including its final independent review. The 24 shared profiles are separate from those budgets. These are content budgets, not reasons to pad the dataset; completion of an authoring budget does not certify implemented solvers, rendered output or host compatibility.

A subsequent reviewed depth batch added 32 records: eight each for object archetypes, material profiles, physical behaviors and tool adapters. Each of these four families contained 11 records at that checkpoint. The batch adds 26 sources and 144 evidence events; three water revisions separately add six events and portable historical snapshots. Independent review corrected model assumptions and implementation boundaries before promotion. Runtime status remains `not-run` for these data records.

The handbook expansion adds 12 recipes and six knowledge records, backed by 18 additional sources and 108 evidence events. Recipes cover finite Fourier synthesis, reciprocal mapping on the Riemann sphere, Mandelbrot escape orbits, one linear phase-portrait family, first-reflection room acoustics, the single-qubit Bloch ball, and six bounded logic/discrete tasks. Inspection guidance covers synchronized slices, linked plot selection, metric measurement, uncertainty intervals, detail insets and reversible guided steps. Independent review corrected certificate ancestry, input exhaustion and self-request handling, and checked numerical/coordinate boundaries before integration. All 18 records remain source-reviewed authoring guidance without runtime certification.

The object and science expansion adds 18 knowledge records, 17 sources and 116 evidence events: zipper, bearing and valve construction; corrugated board, FDM and composite surfaces; slider-crank closure, belt traction, Hertz contact, cantilever bending, shaft torsion and slack-cable engagement; and six scientific data bridges/checks. The data guidance covers DICOM frame/value interpretation, segment overlap/reference grids, fixed-topology trajectory identity, per-frame periodic images, VTK field association and CF packed-field decoding. At that checkpoint, the families contained 14 object archetypes, 14 material profiles, 16 physical behaviors, 14 tool adapters, four motion profiles and 75 domain-validation records. These are bounded authoring contracts and proposed checks, not implemented importers, executed solvers or certified medical pipelines. Two additional evidence events and a historical snapshot record a framed-shelter vocabulary revision after the larger catalog exposed an existing retrieval case failure; its model content and source claims are unchanged.

The mixed-model research expansion adds 12 independently reviewed recipes, 17 net-new source records and 92 public evidence events. Four cover predator-prey phase time, root-zone irrigation, closed-lane car following and construction sequencing; four cover RAG retrieval provenance, one-head attention aggregation, a small-network forward pass and a two-parameter loss slice; four cover dental motion, supplied brain atlases, deposited binding sites and orbital isosurfaces. These are bounded, source-reviewed authoring contracts with explicit provenance, coordinate and dimension limits. No environment solver, importer, patient model, protein or orbital parser, retriever, neural model or renderer was executed. At that checkpoint, all 100 authored known-intent retrieval cases passed their lexical regression checks; this does not validate scene behavior or subject correctness.

The experience, mathematics and physics expansion adds six independently reviewed recipes, two creative theme profiles, eight source records and 48 public evidence events. The recipes cover a bounded quadratic Julia filled-set escape inspector, ideal two-qubit Bell-state marginals and correlations, one rigid rectangular-room pressure eigenmode, one homogeneous isotropic P/S plane wave, accessible museum route/object inspection, and a reversible branching spatial storybook. Cyberpunk and steampunk records are authored fictional construction/material/light/motion proposals rather than genre definitions or historical claims. The recipes keep finite iteration, ideal quantum expectations, lossless cavity assumptions, unbounded-medium wave kinematics, venue-specific accessibility guidance and narrative/scene checkpoint ownership explicit. No fractal renderer, quantum experiment, room or elastic solver, narrative engine, museum artifact, accessibility audit or renderer was executed.

The first Blender, glTF and Three.js handoff wave added 17 independently reviewed knowledge records, 26 net-new sources and 186 public evidence events. Six cover evaluated geometry, hard-surface normals, UV/bake correspondence, Geometry Nodes instances, transform/pivot conversion and deformation-animation ownership. Seven cover linear color, metal/rough ORM, tangent normals, procedural-shader bake/rebuild decisions, IBL/shadow ownership, transparency and thick transmissive volume. Four cover kinematic-to-dynamic velocity transfer, compound collider mass properties, physics timeline lifecycle and baked deformation versus runtime simulation.

The second wave adds 12 independently reviewed records, 19 net-new sources and 109 evidence events. Four object archetypes cover functional assemblies, modular sockets, construction-aware deformation cages and camera-resolved hero detail. Material profiles cover layered decals and geometry/bump/displacement/parallax boundaries. Tool adapters now cover Draco/Meshopt/KTX2 decoder lifecycle and `KHR_texture_transform` semantics/fallback. Lighting profiles cover lighting-only lightmaps, emissive practicals, local probe scope and renderer/per-light shadow update ownership. At the second-wave checkpoint, affected family counts are object 18, material 19, tool 22 and lighting 13; other checkpoint counts include geometry 10, motion 6, physical 18 and reasoning 97. The [handoff reference](../skills/3dviz-pro-max/references/blender-threejs-handoff.md) connects these decisions. Both waves are source-reviewed contracts; no Blender scene/export, glTF round trip, decoder, Rapier run, Three.js render or visual comparison was executed.

The third wave adds 21 independently reviewed profiles, 24 net-new sources and 172 public evidence events. Seven presentation/composition profiles cover labels, disclosure, provenance, sectional depth, scale transitions, linked comparisons and navigable density. Six motion profiles cover root-motion ownership, track masks, event markers, morph composition, loop seams and interruption. Six interaction profiles cover manipulation, snapping, hierarchy selection, transactional history, annotation and modality mapping. Two geometry profiles cover projected LOD discrepancy and morph correspondence. Review corrected root-sample double counting, whole-vector morph ownership, ping-pong event direction and post-commit capture-loss cleanup. No animation, input or geometry runtime was executed.

The fourth wave adds 14 records, 19 net-new sources and 138 public evidence events: five diagnostic inspection patterns, three validation profiles, four output profiles, one animation-pointer adapter and one breakable-joint contract. Eight overlapping domain-validation candidates are explicit subcases in the three profiles instead of duplicate records. Inspection views do not prove acceptance; unavailable required capabilities remain unmet acceptance. Physics readout, Fetch cancellation, canvas capture and plugin support have explicit source/version boundaries. No runtime was exercised.

Further coverage is open-ended: deeper subjects, mixed-domain cases and broader paraphrase evaluation are separate from completing the retained authoring budgets. New demo work, visual upgrades and showcase production are deferred. Existing artifact observations remain valid only for their recorded revisions and scopes. Host installation, marketplace discovery and hosted CI/release checks remain separate from dataset checks.

## Numeric coverage and provenance

Records that used to describe a look in adjectives now carry numbers a scene can be built from. 33 style profiles, eight mood lighting profiles and three camera / time-of-day records (46 knowledge records in total) hold an authored `defaults` block stating palette, sky, fog, light rig, tone mapping and exposure, camera, materials, motion and post in one comparable shape; five of the style profiles are named looks written as complete art directions. 21 recipes hold numeric `defaults` with a worked expected value: 18 recipes split evenly across anatomy, physiology and molecules, plus three corrected mechanics, astronomy and robotics recipes, alongside three state tables (cardiac cycle, lever classes, slip systems). Where a record declares no tolerance of its own, `knowledge.validation-default-tolerance-rule` supplies the default. Four factual tooling records cover glTF punctual light units, the Blender view-transform to tone-mapping table, headless glTF export and the glTF material extension matrix.

The `defaults` block states its own provenance, and the distinction matters more than the numbers. **`source-backed`** defaults cite claim IDs read from a registered source — 13 of the 21 recipe blocks are source-backed. **`authored`** defaults are editorial starting points tuned against a stated exposure and mood: all 46 knowledge blocks and the remaining eight recipe blocks are authored, because their sources print no numbers. No renderer, solver or measuring instrument produced any authored value. Every block also carries `known_limits`, and any of them may be changed when a brief asks for something else.

## Handbook comparison

The [handbook coverage audit](handbook-coverage.md) maps the maintainer's broader 3D handbook to current records. All 28 listed style terms are represented, while many specific scientific models, interaction patterns and experience formats still need dedicated recipes. A generic representation or populated direction does not establish subject-specific completeness.

## Evaluate the skill without generating a demo

Use the [brief behavior probes](../evals/skill-behavior/brief-cases.json) to review whether the agent preserves small-edit scope, combines creative choices, researches factual meaning, assigns physical state correctly and synthesizes a design instead of dumping search output. Their expected outcomes are review criteria, not recorded passes.

Three initial [non-rendering response observations](../evals/skill-behavior/observations.md) cover a scientific heart plan, a mixed-material fictional setting and a title-only edit with no supplied artifact. They demonstrate bounded responses in this repository session, not installation or rendered quality.

The [Latent Village application walkthrough](../evals/skill-behavior/village-application/README.md) applies the completed catalog through 14 actual searches, 19 resolved records and source-code inspection. It records irrelevant search candidates, applicability decisions and a proposed design tied to concrete implementation gaps. It is an analysis result, not evidence of an improved render.

All 249 authored known-intent lexical cases pass, including Blender/Three.js object, surface, lighting and physics handoff prompts. All unit tests and the schema/evidence, generated-index and package checks also pass; 209 protected baseline files and the original 414-source prefix remain exact. Broader paraphrase robustness remains a separate evaluation. Structural validation checks data contracts and evidence linkage. Neither substitutes for independent domain review or future rendered evaluation.

The reviewed mathematics, logic and puzzle subject batch adds twelve recipes, fourteen sources and 92 events. Finite numerical/discrete models retain exact input/state boundaries and explicit Blender-to-Three.js construction; inverse history is distinct from imported-state solving. These are source-reviewed authoring contracts, not executed solvers. The subsequent scientific and applied collections are also complete, as recorded below.

Nineteen further scientific recipes cover supplied biomedical records, mechanical/XPBD/contact models, chemistry trajectories/orbitals and environmental/observer/scenario models. They add 23 net-new sources and 146 public evidence events. Each separates source facts, authored numerical assumptions, object construction and runtime ownership; no clinical, calibration, forecast or device validation is implied.

The completed Blender/Three.js depth program adds **74 knowledge records and 49 subject recipes (123 total)**, with 146 net-new sources and 1,016 public evidence events. The final applied collection contributes cities/operations 5, AI/quantum 5, art/story 5 and XR 3. Overlapping candidates reuse existing records; this is coverage of the approved collection, not exhaustive coverage of every 3D subject. A clinical intervention recipe still requires a named procedure/version, primary guidance and qualified review before collection. Actual artifact, specialist-data and device validation remain separate from documentary acceptance.