2 commits

Author	SHA1	Message	Date
Claude	a76ed3b9b8	Quality aggregation: divide out how many questions a leaf was asked ... 39.17 left the search's storey choice unexplained and blamed value_rate. It is not the rate, or not only. Measured over the twelve baseline runs, value/cost by leaf kind: outside ground 7.40, roof terrace 2.69, room 0.34, circulation 0.02. A terrace returns 2.7x its cost where a room returns a third of it, so filling upper storeys with terrace is not the search leaving value on the table -- it is by a wide margin the most profitable thing the objective offers. 7% of the corpus area produces 32% of its value. Most of that gap is mean quality: 0.986 for a terrace against 0.223 for a room. Quality is a PRODUCT of factors and the kinds are not asked the same number of questions -- an outside leaf is exempt from size, crinkliness and access, so 3 of 7 factors can ever bite it against a room's 6. Each exemption is individually right (no programme size target; uncovered outside is lit by definition; ground-level outside needs no access). The consequence is not: a leaf exempt from the two harshest factors out-scores one judged on them and doing well, purely by not being asked, and quality multiplies the value rate. Stated generally, and this is not about outside space: under a product, adding any new quality criterion mechanically devalues every leaf it applies to, including leaves that score 1.0 on it. The objective's scale should not depend on how many things it measures. quality_aggregate="geometric_mean" (default OFF, "product" is stock) divides that out. Computed in log space so six small factors cannot underflow the product before the root is taken; a zero factor still gives zero, so a fully buried leaf is worth nothing either way. Telling "exempt" from "asked and scored 1.0" needs factor_is_asked, which restates conditions that live inside the quality_* methods. That duplication can drift, so tests/test_fitness_aggregate.py pins it against every leaf in the corpus: wherever the predicate says exempt, the factor really is 1.0. Fail set byte-identical everywhere, and for a stronger reason than 39.13/39.14 had: evaluate_leaf emits each fail from the factor itself before anything is combined, so no aggregation can move one. Score effect +37% to +169%, reaching all four programmes where the crinkliness changes reached two; room value/cost 0.34 -> 0.66, circulation 0.02 -> 0.07. Deliberately not fixed: a terrace still out-earns a room 4:1, which is the rates (value_supported = value_inside = 300 against costs of 110 and 200), not the aggregation. That is a design judgement for the programme author, and 39.16 is a standing reminder that "this inherited constant looks wrong" has been wrong twice already in this section. Left open on ecx with the numbers. A/B running; verdict to follow. Refs homemaker-py-ecx. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB	2026-09-05 18:57:00 +00:00
Claude	366a047a60	Make the crinkliness factor one-sided: stop billing the daylit wall twice ... The tail rescale shipped in cd392e7 is a measured NULL as a search intervention -- 12 of 12 pairs byte-identical on harbor and maple, 8000 evals from a plateau, not merely underpowered. Of course it is: the whole failing tail is 0.034% of corpus value. Looking at the rest of the factor, prompted by the owner, found something much larger above the threshold. crink = area_outside/area = (L*h)/A, so 1/crink = A/(L*h) is the room's mean depth from its daylit wall in storey-heights. That is the right variable for a daylight rule, and the fail boundary it implies (1/crink = 1.62, i.e. 4.86 m at h=3) is a sensible one that agrees with 38.3's frontage bound derived independently. What is wrong is hanging a TWO-sided gaussian on it: * The near side penalises a room for having MORE daylit wall than target -- while leaf_cost's siblings edge_cost and outside_edge_cost already charge that same wall at exterior_wall=100 and boundary_wall=133.3 per m2. The wall is billed once in cost and again as lost value. * It never earns its keep as a failure either: the over-exposed branch only reaches FAIL_THRESHOLD above crinkliness 21.5, and the corpus maximum is 3.95. It has never produced a single fail; it only removes value. * 133 of the 318 passing graded leaves in the 500k baseline (42%) sit on that side, mean quality 0.810. crinkliness_shape="daylight" (default OFF, "gaussian" is stock) clips it: a room shallower than the gaussian's peak scores 1.0, because daylight is a sufficiency requirement and surplus is the cost model's business, not this factor's. Clipping at the PEAK rather than at FAIL_THRESHOLD is deliberate -- it keeps the factor continuous and preserves the graded approach to the daylight limit, where clipping at the threshold would put a 10x cliff on the exact boundary the 0.5**n fail multiplier already steps on. Fail set byte-identical on all 21 corpus artefacts for all four shape/tail combinations, so stock stays a valid yardstick for every arm. Area-weighted crinkliness quality 0.480 -> 0.513, leaf quality product 0.2722 -> 0.2831; per-artefact score +0.2%..+19.6%, and unlike the ramp it reaches health-centre and programme-house, where the tail change was 0.000%. Note "daylight" clips the OPPOSITE side from 38.1's superseded compact_ok, which forgives being buried; composing either with those modes is refused. ab_9gj_ramp.py becomes ab_9gj_crinkliness.py and takes named arms, since it now covers both changes; its first arm is the baseline and the yardstick. Refs homemaker-py-9gj. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB	2026-09-05 07:27:02 +00:00

Renamed from experiments/ab_9gj_ramp.py (Browse further)