cd392e77c5
quality_uncrinkliness evaluates a gaussian at x = 1/crink, so its exponent grows like 1/crink^2 and underflows a double to exactly zero below crink ~ 1/15. Measured over the twelve 500k cold-start runs (39.12): 430 leaves carry a minimum-exposure requirement, 112 fail it, and those 112 span quality 1e-300..1e-1 while contributing 0.034% of total value on 23% of the floor area. Every value in that range is numerically zero beside a passing leaf's ~1, so the search cannot rank two layouts that differ only in how exposed their under-lit rooms are. This is wider than the bead's diagnosis (a flat 0.0 for zero-exposure leaves) and it explains why 38.1's `floor` mode measured as a no-op: max(q, 0.01) maps 110 of the 112 onto one constant, replacing a flat zero with a flat 0.01. crinkliness_tail="ramp" (default OFF, "gaussian" is stock) replaces the tail -- only the tail, only below FAIL_THRESHOLD, only on the compact side -- with a straight line in crinkliness meeting the gaussian exactly at the crossing. _crink_at_fail_threshold inverts the gaussian there using the same truncated _E the factor is evaluated with. Deliberately conservative: nothing at or above FAIL_THRESHOLD moves, so no calibration changes and no leaf crosses the threshold. The fail set is byte-identical on all 21 committed corpus artefacts, the four init.dom seeds included -- asserted in tests/test_fitness_crinkliness_tail.py, not assumed. That invariance is also what makes it legal to score both arms of the A/B under stock (the 38.9 trap's one exemption). A fully buried leaf still scores exactly 0; this restores an ordering within the failing region, it does not forgive it. Composing with 38.1's superseded modes is refused, since both rewrite the same tail. Score effect on the baseline artefacts: +0.3%..+2.8% on harbor and maple, exactly +0.000% on health-centre, programme-house, and every init.dom -- a programme with no partially-exposed failing rooms has nothing to grade, and neither does any starting layout. The ramp is a mid-search signal by construction, so experiments/ab_9gj_ramp.py defaults to seeding each run from a 500k plateau artefact rather than cold. The module-level math import replaces a now-redundant local one. DESIGN.md 39.13 and the A/B verdict follow in a separate commit. Refs homemaker-py-9gj. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB |
||
|---|---|---|
| .beads | ||
| .claude | ||
| examples | ||
| experiments | ||
| src/homemaker_layout | ||
| tests | ||
| .gitignore | ||
| AGENTS.md | ||
| CLAUDE.md | ||
| DESIGN.md | ||
| pyproject.toml | ||
| README.md | ||
homemaker-layout
Programme-driven building-layout search over slicing trees. A clean-room Python successor to the Perl Urb project, intended to eventually be 100% Python.
Why a rewrite
Urb represents a building as a binary slicing tree where room sizes are derived top-down from division ratios. That makes room area an emergent property of every cut above it, which:
- gives the genome low locality (a cut near the root rescales every descendant),
- makes target room sizes nearly impossible to hit, so the gaussian size penalty dominates fitness, and
- defeats crossover (transplanted subtrees lose their proportions).
homemaker inverts this: leaves carry target dimensions from the programme and division ratios are solved bottom-up for a fixed topology. The evolutionary search then only explores topology + types + adjacency.
Phase plan
Solver experiment: port Urb's geometry, re-solve ratios from programme targets, score the result against the original via the Perl oracle.✓Native Python fitness (retire the Perl oracle).✓Memetic search: canonical slicing genome + high-locality operators + Nelder-Mead inner loop.✓Penalty reshaping: lexicographic✓(-n_fails, fitness)outer-search comparison.Representation upgrade: canonical slicing encoding + bottom-up shape feasibility, scaled to larger programmes.✓- Search-quality experiments (current): a long running series of
opt-in levers tried against the
harbor-house,health-centre, andprogramme-houseexample corpora — leaf-sharing, finish-time cell→room collapse, ruin-and-recreate LNS, 2-opt polish, multi-use/co-located leaves, adjacency-graph and bubble-diagram fitness signals, and more. Most of these are negative/null results kept as opt-in flags or reference code rather than defaults. SeeDESIGN.md§11 onward for the full, numbered experiment log with methodology and results for each.
Layout
src/homemaker_layout/dom.py— read/write Urb.domYAML into aNodetree.src/homemaker_layout/geometry.py— faithful port of Urb's top-down geometry.src/homemaker_layout/programme.py— parsepatterns.configspace requirements.src/homemaker_layout/solver.py— bottom-up ratio solve (scipy).src/homemaker_layout/fitness.py— native Python fitness evaluator.src/homemaker_layout/fitness_cmd.py—homemaker-fitnessCLI (drop-in forurb-fitness.pl).src/homemaker_layout/collapse_cmd.py—homemaker-collapseCLI: finish-time global cell→room relabel of a.dom.src/homemaker_layout/graph.py— leaf-adjacency graph for programme-driven checks.src/homemaker_layout/genome.py— topology genome: base-floor tree + per-storey deltas.src/homemaker_layout/operators.py— high-locality mutation and subtree crossover.src/homemaker_layout/innerloop.py— ratio optimisation inner loop (Nelder-Mead / CMA-ES).src/homemaker_layout/driver.py— memetic search outer loop.src/homemaker_layout/evolve.py—homemaker-evolveCLI entry point.src/homemaker_layout/oracle.py— legacy Perl shim, kept for cross-validation only.src/homemaker_layout/bubble.py— 3D bubble-diagram adjacency fitness-signal prototype (DESIGN.md §27); validated null, not wired intofitness.py— reference only.
Room codes and reserved names
Leaf types live in three namespaces that share a first character. Only the first is enforced; the other two are conventions the fitness function reads, so a room's spelling can change how it is scored.
1. Generic structural types — C, O, S (reserved). The leaves the
search itself creates: C circulation, O outside, S sahn (an outside court
that also serves as circulation). Always uppercase. A programme code spelled
exactly C, O or S is rejected at load.
2. Programme room codes — anything else, lowercase. k1, b1, cr1,
of, and single-character codes like r or t. These may start with any
letter: since DESIGN.md §39.4 the generic tests match C/O/S exactly, so
naming a room cr1 no longer makes it circulation. (Before that fix it did —
and silently dropped it from the required-space check entirely.)
3. Access requirements — the usage: attribute. Every space declares one
of living, kitchen, bedroom, toilet, utility, none. Mandatory, no
fallback, and a missing or unknown value is a load error. It replaced a
first-character convention (b/t/l/k) under which a room silently
inherited another room's connectivity rules from its spelling — la1 "Laundry
Room" was trimmed as a living room (DESIGN.md §39.7).
spaces:
la1:
usage: utility # controlled, drives engine behaviour
name: Laundry Room # free text, building-specific
A usage value exists only where the engine treats it differently, so the vocabulary is closed: a new access class means new code, not new config. Check a programme with:
python experiments/audit_programme_config.py
which reports reserved-name collisions, the usage class each code picks up, and whether each room's size/width/proportion/crinkliness targets are mutually satisfiable at all.