From 1b3a0adc2ecf6263365ac3ff75915905e0885ddb Mon Sep 17 00:00:00 2001 From: Claude Date: Wed, 26 Aug 2026 07:40:37 +0000 Subject: [PATCH] beads: file homemaker-py-2v1 (connectivity mispricing, P0) Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB --- .beads/issues.jsonl | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/.beads/issues.jsonl b/.beads/issues.jsonl index 57f6306..ce79414 100644 --- a/.beads/issues.jsonl +++ b/.beads/issues.jsonl @@ -1,5 +1,6 @@ -{"_type":"issue","id":"homemaker-py-hxi","title":"Buried circulation and outside space are negative-value: search is rewarded for deleting the circulation spine","description":"Direct consequence of the zero-exposure bug. Measured on a constructed harbor-house seed: deleting a buried O leaf improved the raw score 85x and removed 7 fails; deleting a buried C leaf improved it 62x and removed 6 fails. Programme rooms are held in place ONLY by the missing-space fail cascade, not by contributing value -- deleting a buried k1/da1/m costs +15 fails, so they stay, but nothing positive keeps them. Circulation and outside leaves carry no missing-space requirement, so nothing keeps them at all. Observed live: in a 20k-eval harbor-house run, undivide/core_undivide account for a large share of recorded improvements (16 occurrences in the log) -- the search is literally deleting circulation to score better. This explains three prior negative results as a single mechanism: 18 graded circulation-connectivity (a tie-break signal cannot beat a 60x scalar gradient), 21/22 bridge_circulation (the operator inserts corridor leaves the objective immediately punishes), and the 'level N not connected' hard fails surviving \u003e1M evals in the 3M run.","notes":"Depends on the zero-exposure fix; may need no separate fix if (a)/(c) there restores a value gradient for circulation. Worth re-running the 18 and 21/22 A/Bs afterwards -- both may have been measuring a broken gradient rather than a bad idea.","status":"open","priority":1,"issue_type":"bug","owner":"noreply@anthropic.com","created_at":"2026-08-25T23:14:40Z","created_by":"Claude","updated_at":"2026-08-25T23:14:40Z","dependencies":[{"issue_id":"homemaker-py-hxi","depends_on_id":"homemaker-py-ssz","type":"blocks","created_at":"2026-08-25T23:15:12Z","created_by":"Claude","metadata":"{}"}],"dependency_count":1,"dependent_count":0,"comment_count":0} -{"_type":"issue","id":"homemaker-py-ssz","title":"Zero-exposure leaves score hard quality=0: the objective assigns no value to any interior room","description":"fitness.quality_uncrinkliness returns a hard 0.0 when a leaf has no daylit wall (area_outside==0 =\u003e crink==0 =\u003e 'if not crink: return 0.0'). Because evaluate_leaf MULTIPLIES factors into quality, and process_storey accumulates 'value += quality * rate * area', such a leaf contributes EXACTLY ZERO value while still adding cost. Measured on the full default construction stack (leaf_sharing, depth_balanced, interior_outside, collapse_insearch), 3 seeds each: harbor-house 46% of interior leaves, health-centre 45%, maple-court 56% are zero-exposure. On a converged 20k-eval harbor-house run (seed 1, 57 fails) 14 of 17 crinkliness fails are zero-exposure, and ~470 m2 of the 721 m2 ground floor plate sits at zero value. This is the mathematically consistent limit of the gaussian (1/crink -\u003e inf), so it is a faithful port, not a porting bug -- but it means the objective's gradient does not describe a good building.","design":"Options, none yet chosen: (a) floor the factor at a small epsilon instead of 0 so buried leaves keep a value gradient and remain rankable; (b) make the gaussian one-sided (clip to 1.0 on the compact side) so being LESS exposed than target is not punished as hard as being over-exposed -- architecturally, a compact well-insulated room is not a defect; (c) exempt circulation/store types from the daylight requirement entirely (uncrinkliness_circulation currently uses the same [5/6, 1.1/3] as habitable rooms, so internal corridors -- completely normal architecture -- are guaranteed failures). Any change here invalidates prior fail-count baselines, so it needs its own A/B and a DESIGN.md section.","acceptance_criteria":"A/B on harbor-house + maple-court at fixed budget showing the chosen variant lowers hard-fail count without inflating soft; DESIGN.md section recording the result; prior baselines re-stated under the new objective.","status":"open","priority":1,"issue_type":"bug","owner":"noreply@anthropic.com","created_at":"2026-08-25T23:14:40Z","created_by":"Claude","updated_at":"2026-08-25T23:14:40Z","dependency_count":0,"dependent_count":2,"comment_count":0} +{"_type":"issue","id":"homemaker-py-2v1","title":"Connectivity is under-priced ~3x against the circulation-\u003ehabitable value gap: the objective is net-positive on severing the spine","description":"Sharper root cause than homemaker-py-ssz, isolated by the 38.6 A/B (which showed none of the three crinkliness modes removes the deletion incentive). Deleting a circulation leaf merges it into its sibling, converting corridor into habitable area. value_circulation=50 vs value_inside=300, so that is a flat 6x value gain. The only counter-pressure is the 'level N not connected' fail, worth 0.5x under value *= 0.5**len(failures). Break-even needs 0.5^k \u003c 50/300, i.e. k \u003e 2.58 -- severing must cost AT LEAST 3 fails to be net-negative, and it costs 1. Net incentive to sever = 6 * 0.5 = 3.0x in favour; measured 4.06x on a well-lit (q_crink=0.736) circulation leaf, so this is NOT the zero-exposure effect and is not fixable inside quality_uncrinkliness. This is the cleanest explanation of why 'level 0 not connected' and 'level 1 not connected' are still present in evolved-3M-nols-3, the best layout found after 1.7M evals: the search is being paid 3-4x to create them.","design":"Options: (a) emit connectivity fails with a multiplicity \u003e= 3 (cheapest, but stacks with the 1i8 cascade-weighting problem and is a magic number); (b) make the connectivity penalty multiplicative and explicit rather than riding the generic 0.5^n (a dedicated building_factor term, sized from the value-rate gap so it tracks value_circulation/value_inside instead of being hard-coded); (c) revalue circulation as infrastructure -- its worth is that it makes other rooms reachable, which the current per-leaf value rate cannot express; the principled version credits circulation with the access it provides rather than its own floor area. (c) is the architecturally correct one and the biggest change. Recommend measuring (a) first purely to confirm the mechanism (does the 3x threshold flip the deletion test?), then designing (b) or (c) properly.","acceptance_criteria":"Deletion test (experiments/diag_exposure_frontage.py value + the ssz A/B harness) shows lit and buried C/O deletions are no longer rewarded; then harbor-house reaches the 15-fail floor in materially fewer than 1.7M evals AND without 'level 0/1 not connected' in the result.","status":"open","priority":0,"issue_type":"bug","owner":"noreply@anthropic.com","created_at":"2026-08-26T07:37:39Z","created_by":"Claude","updated_at":"2026-08-26T07:37:39Z","dependency_count":0,"dependent_count":2,"comment_count":0} +{"_type":"issue","id":"homemaker-py-hxi","title":"Buried circulation and outside space are negative-value: search is rewarded for deleting the circulation spine","description":"Direct consequence of the zero-exposure bug. Measured on a constructed harbor-house seed: deleting a buried O leaf improved the raw score 85x and removed 7 fails; deleting a buried C leaf improved it 62x and removed 6 fails. Programme rooms are held in place ONLY by the missing-space fail cascade, not by contributing value -- deleting a buried k1/da1/m costs +15 fails, so they stay, but nothing positive keeps them. Circulation and outside leaves carry no missing-space requirement, so nothing keeps them at all. Observed live: in a 20k-eval harbor-house run, undivide/core_undivide account for a large share of recorded improvements (16 occurrences in the log) -- the search is literally deleting circulation to score better. This explains three prior negative results as a single mechanism: 18 graded circulation-connectivity (a tie-break signal cannot beat a 60x scalar gradient), 21/22 bridge_circulation (the operator inserts corridor leaves the objective immediately punishes), and the 'level N not connected' hard fails surviving \u003e1M evals in the 3M run.","notes":"Depends on the zero-exposure fix; may need no separate fix if (a)/(c) there restores a value gradient for circulation. Worth re-running the 18 and 21/22 A/Bs afterwards -- both may have been measuring a broken gradient rather than a bad idea.","status":"open","priority":1,"issue_type":"bug","owner":"noreply@anthropic.com","created_at":"2026-08-25T23:14:40Z","created_by":"Claude","updated_at":"2026-08-25T23:14:40Z","dependencies":[{"issue_id":"homemaker-py-hxi","depends_on_id":"homemaker-py-ssz","type":"blocks","created_at":"2026-08-25T23:15:12Z","created_by":"Claude","metadata":"{}"},{"issue_id":"homemaker-py-hxi","depends_on_id":"homemaker-py-2v1","type":"blocks","created_at":"2026-08-26T07:37:47Z","created_by":"Claude","metadata":"{}"}],"dependency_count":2,"dependent_count":0,"comment_count":0} +{"_type":"issue","id":"homemaker-py-ssz","title":"Zero-exposure leaves score hard quality=0: the objective assigns no value to any interior room","description":"fitness.quality_uncrinkliness returns a hard 0.0 when a leaf has no daylit wall (area_outside==0 =\u003e crink==0 =\u003e 'if not crink: return 0.0'). Because evaluate_leaf MULTIPLIES factors into quality, and process_storey accumulates 'value += quality * rate * area', such a leaf contributes EXACTLY ZERO value while still adding cost. Measured on the full default construction stack (leaf_sharing, depth_balanced, interior_outside, collapse_insearch), 3 seeds each: harbor-house 46% of interior leaves, health-centre 45%, maple-court 56% are zero-exposure. On a converged 20k-eval harbor-house run (seed 1, 57 fails) 14 of 17 crinkliness fails are zero-exposure, and ~470 m2 of the 721 m2 ground floor plate sits at zero value. This is the mathematically consistent limit of the gaussian (1/crink -\u003e inf), so it is a faithful port, not a porting bug -- but it means the objective's gradient does not describe a good building.","design":"Options, none yet chosen: (a) floor the factor at a small epsilon instead of 0 so buried leaves keep a value gradient and remain rankable; (b) make the gaussian one-sided (clip to 1.0 on the compact side) so being LESS exposed than target is not punished as hard as being over-exposed -- architecturally, a compact well-insulated room is not a defect; (c) exempt circulation/store types from the daylight requirement entirely (uncrinkliness_circulation currently uses the same [5/6, 1.1/3] as habitable rooms, so internal corridors -- completely normal architecture -- are guaranteed failures). Any change here invalidates prior fail-count baselines, so it needs its own A/B and a DESIGN.md section.","acceptance_criteria":"A/B on harbor-house + maple-court at fixed budget showing the chosen variant lowers hard-fail count without inflating soft; DESIGN.md section recording the result; prior baselines re-stated under the new objective.","status":"open","priority":1,"issue_type":"bug","owner":"noreply@anthropic.com","created_at":"2026-08-25T23:14:40Z","created_by":"Claude","updated_at":"2026-08-25T23:14:40Z","dependencies":[{"issue_id":"homemaker-py-ssz","depends_on_id":"homemaker-py-2v1","type":"blocks","created_at":"2026-08-26T07:37:47Z","created_by":"Claude","metadata":"{}"}],"dependency_count":1,"dependent_count":2,"comment_count":0} {"_type":"issue","id":"homemaker-py-tdp","title":"Frontage-budget feasibility bound: harbor-house and maple-court are ~3x short of the exposed wall their own crinkliness constraint demands","description":"Closed-form bound, no search needed. Crinkliness fails when 1/crink \u003e 1.6202 (solving gaussian(x,1,5/6,1.1/3)=FAIL_THRESHOLD), and crink = L*h/A, so every interior leaf needs exposed wall length L \u003e= A/(1.6202*h) -- per storey, A_storey/4.86 metres at h=3. Supply vs demand per storey: harbor-house 54m supplied / 148m needed (2.7x short); maple-court 56m / 162m (2.9x short); health-centre 43m / 41m (feasible); programme-house 24m / 12m (2x surplus). The deficit comes from half the plot perimeter being marked 'private' in harbor/maple -- area_outside skips private/fortified edges. This PREDICTS the corpus fail-count plateau: the two programmes that are frontage-short are exactly the two that plateau (harbor 30-40, maple 74-84 fails), and the two with surplus are the two that effectively solve (programme-house reached 1 fail in a 12k-eval run here). Causal check: relabelling harbor's two private edges as open (identical geometry and programme, perimeter labels only) cut zero-exposure leaves 52%-\u003e19% and seeder crinkliness fails 16.5-\u003e12.0 over 6 seeds.","design":"Ship the bound as a pre-flight check (a 'homemaker-programme-check' style report, or a warning from evolve.py at startup): given a plot and programme, report required vs available frontage and the courtyard area needed to close the gap. This is the cheap, no-ground-truth-needed version of what 2g7.2 objective calibration wanted -- it answers 'is this programme even satisfiable' without needing traced human plans.","acceptance_criteria":"A tool/check that reports the frontage budget per programme dir; DESIGN.md section with the derivation and the corpus table; the plateau programmes flagged as frontage-infeasible-as-specified.","status":"open","priority":1,"issue_type":"task","owner":"noreply@anthropic.com","created_at":"2026-08-25T23:14:40Z","created_by":"Claude","updated_at":"2026-08-25T23:14:40Z","dependency_count":0,"dependent_count":0,"comment_count":0} {"_type":"issue","id":"homemaker-py-2g7.4","title":"Exact shape-curve inner loop (Otten/Stockmeyer DP) replacing Nelder-Mead","description":"The classic slicing-floorplan result applied to our exact representation: each leaf's size/width/proportion constraints define a feasible-shape region; these compose bottom-up through the slicing tree as piecewise shape curves, yielding in ONE linear pass (no iteration): (a) whether ANY ratio assignment satisfies all per-leaf shape constraints, and (b) the ratios that realize a chosen point on the root curve. Today the same question costs an 80-eval NM run per child (~all of the 3M-eval budget) and answers it only approximately. Plan: (1) prototype on harbor-house-l0 with a rectangular plot approximation; (2) validate against innerloop.optimise — DP-feasible topologies must score \u003e= NM result when polished, DP-infeasible must never reach 0 shape fails under NM; (3) wire as a PRE-FILTER: prune shape-infeasible children before any native eval, and warm-start NM from DP ratios (or replace NM entirely where the plot is near-rectangular; keep NM as final polish for skew). CAVEATS to model honestly: crinkliness/access/adjacency are NOT in the DP (graph terms, not per-leaf shape) — the DP handles the size/width/proportion family only, which is fine for pruning; equal-offset skew-quad geometry means DP areas are approximate — measure the approximation error on real plots first (harbor plot is a near-rect quad). Expected payoff: 100-1000x cheaper feasibility, turning topology search into enumerate-and-prune and unlocking the racing/MAP-Elites/CP issues. Cf. §34: autodiff failed on wall-clock; this is a different attack — exactness via structure, not gradients.","acceptance_criteria":"on harbor-house-l0: DP verdict agrees with NM-polished shape-fail outcome on \u003e=95% of 200 random topologies; measured speedup \u003e=50x per feasibility decision; approximation error on the skew plot quantified","status":"closed","priority":1,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-08-02T09:15:04Z","created_by":"Bruno Postle","updated_at":"2026-08-02T22:42:15Z","started_at":"2026-08-02T18:39:05Z","closed_at":"2026-08-02T22:42:15Z","close_reason":"Prototype PASS: 99.0% agreement (\u003e=95%), 93.6x speedup (\u003e=50x), approximation error quantified (7.5% bbox overestimate). See DESIGN.md §37.2. Not wired into product this session -- follow-up homemaker-py-6xh filed.","dependencies":[{"issue_id":"homemaker-py-2g7.4","depends_on_id":"homemaker-py-2g7","type":"parent-child","created_at":"2026-08-02T10:15:04Z","created_by":"Bruno Postle","metadata":"{}"}],"comments":[{"id":"019fc645-557c-7351-9349-d693e5eefa2b","issue_id":"homemaker-py-2g7.4","author":"Bruno Postle","text":"Post-close correction (user review): the prototype's rectangular\napproximation used an axis-aligned global bbox for each quad's (w,h) --\ncorrect only by coincidence on harbor-house-l0's near-axis-aligned plot\n(~7.5% area error). A real building's orthogonal walls need not align to\nthe plot's survey/CRS axes; confirmed by rotating the plot 45deg, where\nbbox error jumped to 102%.\n\nFixed: (w,h) now measured from edge lengths (edge0+edge2)/2,\n(edge1+edge3)/2 -- rotation-invariant by construction -- and the\nparent/child composition rule (which dimension sums vs. is shared) is now\nderived EXACTLY from child.rotation parity (verified float-exact\nidentity) instead of the empirical geometric heuristic the closed version\nused (which, tried alone without the parity fix, regressed accuracy\n99.0% -\u003e 95.5%).\n\nRe-validated: 99.0% on harbor-house-l0 unrotated (matches original, 0\nfalse negatives), 100% at 97x speedup on the SAME plot rotated 45deg.\nSee DESIGN.md §37.2 (Correction 1 / Correction 2) for full detail.\nhomemaker-py-6xh (production wiring follow-up) still applies.","created_at":"2026-08-03T06:17:40Z"}],"dependency_count":0,"dependent_count":1,"comment_count":1} {"_type":"issue","id":"homemaker-py-2g7.3","title":"Hard/soft fail tiering: 'solved' = zero hard fails","description":"Lex-by-total-count treats a crinkly wall the same as a missing room, so search polishes shape taxes instead of fixing structure — the 3M-run best still carries 'level 0/1 not connected' and wrong-level fails after 1.7M evals. Split fails into HARD (missing space, wrong/required level, level connectivity, circulation connectivity, stairs, covered-outside) and SOFT (crinkliness, proportion, size, width, edge-too-long) tiers. Outer comparator becomes (-hard, -soft, fitness); 'solved' is defined as zero hard fails. GUARDS: (1) the inner-loop 0.5^n cliff must keep protecting against trading into new fails (§4.5/§4.9 — rerun the 0/9 inner-loop-protection check); (2) rerun the §4.9 outer A/B: the scheme must not reintroduce the scalar pathology; (3) §11.4 warns comparator reshaping alone does not escape topology basins — the claim here is narrower: budget stops being spent on soft fails while hard fails remain, and reporting becomes meaningful. The tier map lives in fitness.py next to the fail emission sites so new fail strings must declare a tier. Can start before the calibration issue lands but final tier assignments should be reviewed against its findings.","acceptance_criteria":"tiered comparator behind a flag with A/B on harbor+maple (3 seeds, 20k evals): hard-fail count at budget strictly better or equal on mean, no §4.9 regression; report shows hard/soft split","notes":"ACCEPTANCE A/B COMPLETE — PASS (2026-08-02, experiments/tier_ab_2g7_3.py,\nharbor-house + maple-court, 3 seeds, budget 20000, leaf_sharing=True,\nn_workers=4, wall ~2h53m):\n\n harbor-house hard mean: flat 11.67 -\u003e tiered 5.33 (soft 29.00 -\u003e 42.33)\n maple-court hard mean: flat 19.33 -\u003e tiered 14.00 (soft 71.33 -\u003e 87.67)\n\nHard-fail mean strictly better on BOTH programmes at fixed budget — the\nrequired acceptance bar. Soft/total rise as expected (budget redirected from\npolishing shape fails to structural ones). Full per-seed log at\nscratch/tier_ab_2g7_3/log.txt (not committed — scratch output, regenerate via\nthe script if needed).\n\nGuards: (1) inner-loop 0.5^n cliff untouched by construction (no diff to\ninnerloop.py or the existing 0.5**len(failures) line) — not re-measured\nempirically, doesn't need to be. (2) tiered key is still lexicographic, not a\nblended scalar, so structurally immune to the §4.8 scalar pathology;\nencoded as tests/test_driver.py::test_use_tiers_prefers_fewer_hard_over_fewer_total_fails.\n\nDESIGN.md §37.1 written up with full table and rationale. Feature lands\ndefault-off (--use-tiers / HOMEMAKER_USE_TIERS / driver.search(use_tiers=)),\nso no existing reproduction changes.\n\nFollow-on (not blocking, filed separately): convergence-SPEED comparison\n(evals to 0 hard fails, tiered vs flat, same budget) — this A/B measured\nfail composition at a fixed budget snapshot, not time-to-solved.","status":"closed","priority":1,"issue_type":"feature","assignee":"Bruno Postle","owner":"bruno@postle.net","created_at":"2026-08-02T09:14:14Z","created_by":"Bruno Postle","updated_at":"2026-08-02T17:51:18Z","started_at":"2026-08-02T09:58:53Z","closed_at":"2026-08-02T17:51:18Z","close_reason":"Acceptance A/B passed on both harbor-house and maple-court (hard-fail mean strictly better under tiering); guards verified; DESIGN.md §37.1 written up.","dependencies":[{"issue_id":"homemaker-py-2g7.3","depends_on_id":"homemaker-py-2g7","type":"parent-child","created_at":"2026-08-02T10:14:14Z","created_by":"Bruno Postle","metadata":"{}"}],"dependency_count":0,"dependent_count":0,"comment_count":0}