# Finite assessment, root seed 0

This report summarizes the retained [final results](final-corpus/results.json). The
[sampling policy](sampling-policy.json) was fixed before evaluating the corpus. Every input is kept,
including incompatible stress inputs. There is no rejection resampling. Timing is observed on this
CPU environment, with some rendering/testing work concurrent; it is not a portable benchmark.

| Batch | Valid inputs / attempted inputs | Feasible candidates / attempted candidates | Incompatible inputs | Budget exhausted / defects | Measured seconds |
|---|---:|---:|---:|---:|---:|
| Independent environment/building pairings | 64 / 64 | 84 / 128 | 0 | 0 / 0 | 32.777 |
| Three building programs, seeds 0–7 | 24 / 24 | 48 / 48 | 0 | 0 / 0 | 22.288 |
| Settlement seeds 0–7 and counts 3, 6 | 10 / 10 | 80 / 80 | 0 | 0 / 0 | 65.221 |
| Bounded stress sampling | 9 / 32 | 11 / 64 | 23 | 0 / 0 | 13.515 |

Candidates are complete alternative building placements: stepped perimeter work and piers under
the same locked design. The settlement denominator counts alternatives across its forty building
placements. The solver compares both permitted strategies; this is not a retry after silently
moving a locked building. All batches report zero retries. Stress incompatibilities comprise
**21 relief-limit** and **2 support-height-limit** inputs. Both declared alternatives failed the
same measured hard condition for each incompatible input. This establishes incompatibility with
the delivered policy, not impossibility under every conceivable building technique. A separate
unit test exercises search-budget exhaustion with an unvisited valid candidate.

The 64 pairings use sixteen independently frozen complete environmental scenes and four independent
building seeds. The first eight environment expressions use additive fields. The last eight use the
held-out blend/transform composition. Expressions, seeds, water levels, known domains and frozen
geometry identities are saved in [adaptation-inputs.json](final-corpus/adaptation-inputs.json).
Tests also cover degenerate coefficients, local perturbations, a compatible independent environment,
protected observations, unchanged cells outside authorized work, and hollow support beds.

Observed principal building signatures are **18 / 24**: dwelling **5 / 8**, atelier **8 / 8**, commons
**5 / 8**. The settlement corpus has **10 / 10** observed signatures. Signatures use realized room
extents, purpose/connection topology, levels/stairs, roof geometry and path organization; they omit
names, material color and global transforms. The adaptation-only corpus has three principal designs
across its four building seeds. Repeated principal structures are retained and counted honestly.

Contact geometry is measured separately. There are **44 / 64** observed contact signatures in the
adaptation corpus, **24 / 24** in the building corpus, **10 / 10** in settlements and **9 / 9** among
successful stress inputs. Chosen strategies are 46 piers / 18 stepped in adaptation, 6 / 18 in
buildings, 17 / 23 across settlement buildings, and 9 / 0 in successful stress inputs.

| Batch, valid inputs only | Foundation fill range | Excavated solid cells | Existing cells replaced by sitework | New sitework cells |
|---|---:|---:|---:|---:|
| Adaptation | 0–159 | 0 | 3–168 | 3–87 |
| Buildings | 0–70 | 0 | 1–3 | 0–73 |
| Settlements | 16–216 | 0 | 369–797 | 19–336 |
| Stress successes | 40–174 | 0 | 3–141 | 1–199 |

Fill is the foundation's measured support volume. Replacement and new-cell counts cover sitework
components, including approaches and paths. Water-to-post changes are replacements; zero excavated
solid cells does not mean zero site disturbance. Authorized pool excavation and rejected excavation
under a supported tree are exercised separately in contract and containment tests.

The final test command `python -m pytest -q test` passed **134 tests in 97.27 seconds**. Tests damage
actual geometry and NBT to verify support, stairs/headroom, exits, paired doors, lighting, furnishings,
enclosure, inventories, pool containment, permissions and transport identity. They also verify
substitution, independent extension, phases, transforms, atomic failure, provenance and regeneration.
See [pytest-final.txt](pytest-final.txt).

[replay.json](replay.json) records independent seed-0 runs with `PYTHONHASHSEED=3` and `8675309`,
including output paths containing spaces. Both recover the same content, bounds and semantic digest:
`e81b08063278b231b60eaa16e4120973ed688bb3fd7e6a8da439daa4e7f5caf9`.
The default is five buildings, 125,475 occupied cells and 782 component instances. Seed 1 produces
three buildings and was also exported/reloaded successfully. The output contains one region and
Minecraft data version 3955; every explicit blockstate property and relevant inventory NBT is checked.

Representative samples are listed in [../samples/replay-manifest.json](../samples/replay-manifest.json).
[Visual evidence](../previews/README.md) includes both sides of the settlement, distinct building
massing, a wet-site adaptation and room/stair cutaways. Cutaways are geometry-only previews; the
full originals, not the slices, are used for validation and provenance.

These finite results do not prove universal feasibility or beauty. The documented movement,
lighting, support and roof models bound the claims. Historical failures and their generic fixes
remain in [development-findings.md](development-findings.md).