# Individual-building quality assessment, Hearth 1.3.0

This third phase improves buildings through shared components. The architecture review found no
serious blocker requiring the user's stop condition. The kernel, transport and earlier extension
package are unchanged; [API changes and hashes](api-delta.md) document the boundary. The eight original
lodge images were opened as aesthetic references. No reference program or snapshot was modified,
invoked or used as an implementation template.

## Implemented outcome

Buildings now combine profile roofs with continuous risers, real attached dormers, framed gables,
deeper timber facades, optional accessible balconies, supported masonry chimneys, furniture groups
and planted entrance courts. Roof and facade planning share adjacency decisions so a lower roof
does not collide with an upper window's shutters. Actual route checks include exterior landings,
balcony doors, stair access and furniture interaction positions. Barrel inventories survive reload.

Three lightweight programs select wings, functions, storeys, roof axes/profiles, dormer spacing and
balconies through scoped seeds. They compose the same reusable components and retain locked principal
intent during placement. New roofs can be substituted through a public policy, and an external
facade program can install a dormer through a discovered roof mount. Neither operation needs private
voxel layout or hand-authored provenance.

## Recorded finite results

The [main policy](sampling-policy.json) and [post-fix holdout policy](holdout-policy.json) were recorded
before their respective executions. [Final results](final/results.json),
[frozen environments](final/frozen-inputs.json), [summary](final/summary.json), and
[holdout results](holdout/results.json) retain all inputs and outcomes.

| Batch | Valid inputs / inputs | Incompatible | Defects / budget exhausted | Accepted candidates / candidates | Recorded seconds | Fill cells, successful inputs |
|---|---:|---:|---:|---:|---:|---:|
| Three clients, seeds 0-7 | 24 / 24 | 0 | 0 / 0 | 48 / 48 | 36.371 | 0-121 |
| Frozen environment/building regression pairings | 64 / 64 | 0 | 0 / 0 | 84 / 128 | 63.666 | 0-256 |
| New post-fix pairings | 16 / 16 | 0 | 0 / 0 | 16 / 32 | 17.570 | 49-257 |
| Unresampled stress | 5 / 16 | 11 | 0 / 0 | 6 / 32 | 6.892 | 74-143 |

These are **104/104 supported inputs**, plus a separate **5/16 stress success rate**. Input success
is not candidate acceptance: a complete input can succeed after one support alternative is rejected.
Every input is attempted once; there are no success-only replacement inputs or hidden retries. The
two existing contact alternatives remain bounded by relief five, support height eight and the
existing edit/approach limits. All eleven final stress failures report relief outside that bound.
They are incompatibilities with this declared site policy, not claims of universal impossibility.
Timings are sums of measured per-input wall time on this host, including export/audit for selected
artifacts. Concurrent regression/render work affects them; they are not comparative benchmarks.

No building in these batches excavates terrain. Fill counts include the small chimney bearing plate.
Existing owners, protected regions and support consumers retain the original authority semantics.
Original environment surfaces remain frozen; buildings adapt on forks rather than shaping inputs.
Contact signatures differ in 45 of the 64 regression samples and 12 of the 16 holdout samples. Piers
and stepped contact both remain represented. These observations do not prove every terrain works.

Principal architectural signatures exclude materials, planting, contact geometry and global
placement. They use actual room footprints/adjacency, roof geometry, dormer positions and structural
attachments. The 24 client runs contain **18 observed signatures**: home 4/8, hall 7/8, high house 7/8.
Four locked principal signatures remain unchanged across the 64 environmental pairings, and two
across the new holdout. Optional obstructed dormer mounts are recorded in each result; required rooms,
functions, chimneys and requested balconies are not dropped. This is a finite diversity observation,
not a count or estimate of all possible compositions.

## Failures preserved and generic corrections

The first assessment had **24/24 valid clients but only 31/64 valid adaptation inputs**. Each of the
33 failures contained a real unsupported chimney cell in the pier candidate. Although the bounded
solver reported budget exhaustion alongside the incompatible stepped alternative, these were
implementation defects. `BearingCourse` now spans a small plate from measured anchors. Its 3-by-2
chimney realization adds at most six cells, preserving the original site and edit bounds.

The original 64-input sweep is therefore regression evidence after that fix. The subsequent holdout
uses eight new independent field seeds and previously untried blended/transformed expressions,
crossed with two different structure seeds. All environments are frozen before placement. Its
post-fix first execution and the final railing-correction replay are retained separately.

`first-assessment/`, `second-assessment/` and `before-railing-*` contain prior results and matching
artifact locations. [Development notes](development-notes.md) retain roof-cap, facade/porch,
approach, planting-support, bearing and railing findings. Failed unit runs are retained as logs;
the final suite passes without weakening or deleting any earlier test.

## Regression, identity and public commands

- **200 passed in 164.42 seconds**, [full regression](full-regression.txt). All 179 earlier tests
  remain, plus 21 building-quality cases. New cases cover domain boundaries, all four rotations,
  real glass identity/attachments, independent roof substitution, branch regeneration, rollback,
  permission overflow and deliberately damaged cover, support, guards, doors, lights and access.
- **15 selected full artifact pairs** audited with **2,265 nodes**, **2,017 nonempty-component
  structural queries**, and **314 inventory-bearing cells**. Every node, relation, dependency,
  operation, port use, displacement record, reverse index and exact current-cell set is compared;
  geometry and all NBT are checked independently of palette/property ordering. Manifests live under
  `samples/building-quality/final/` and `holdout/`.
- `python generate.py` succeeds with **five buildings, 128,021 blocks, 908 components**. Its single
  region has world bounds `(-49,-4,-23)` through `(49,38,104)`. The final semantic digest is
  `e1dca5ad9a482e9407abf304740fa82869839837389b3271539c04d687b0bc38`.
- [Default full-structure audit](default-structure-audit.json) additionally checks all 908 nodes,
  868 nonempty-component queries and 88 inventory-bearing cells against regenerated content.
- Seed 0 matches across `PYTHONHASHSEED=3` and `8675309` in [replay.json](replay.json). Seed 1 and an
  output path containing spaces also succeed: three buildings, 91,399 blocks and 611 components,
  digest `1ce03599aa862a97ffbafb2ca343764ba3565051e68ee05072b6bfa099736afa`.
- [Inspection output](inspection.json) prints an actual roof-window glass chain and verifies the
  identical query and current cells after export/reload. The [native import guard](portability.json)
  blocks numpy, MCIO, scipy, shapely and ctypes during new generation. **No new Pyodide execution
  was performed.** Earlier externally reported browser results remain historical, not new evidence.

## Visual review and remaining quality gap

Twenty final proc images cover home 0, hall 6, high house 1 and wet adaptation 00, each from three
exterior angles with a floor cut and a vertical circulation cut. Labeled camera/cut records and
[review notes](../../previews/building-quality/README.md) accompany them. Baseline and development
images remain separate. All eight task-root lodge references were inspected as outcomes only.

The stronger stepped roof silhouettes, roof windows, timber depth, sheltered entrances and varied
heights improve the original blocklike buildings. Cuts expose continuous stairs, room boundaries,
functional stations and clear routes. The wet sample shows actual piers beneath the preserved
building, not a giant leveled platform. Some balcony fence segments are poorly represented in full
proc renders despite correct decoded states and passing damage tests; the retained parser probe and
isolated fence render distinguish this unresolved visual issue from absent schematic guards.

The result remains **below the lodge references' aesthetic quality**. Regular bay rhythms, large
plain rear panels, repeated dormer proportions and broad open lawns remain apparent. Furnishing is
functional but less layered and intimate than a hand-composed interior. Attics are not usable rooms;
chimneys do not simulate heating. The support, light, movement and roof-cover models retain their
documented discrete approximations. No claim is made of universal beauty, full Minecraft physics,
arbitrary roof flashing, automatic city layout or validation beyond the recorded finite domains.

Rerun commands and the complete supplemental R01-R30 ledger are in [GENERATOR.md](../../GENERATOR.md).
Use new output directories when assessing a later revision so these records remain historical.