DESIGN.md §37.1: hard/soft tiering A/B (harbor-house + maple-court, 3 seeds,
20k evals/run) shows hard-fail mean strictly better under the tiered
comparator on both programmes (harbor 11.67->5.33, maple 19.33->14.00) at
the cost of higher soft/total fails — the intended trade. ACCEPTANCE: PASS.
Filed homemaker-py-p6t as a non-blocking follow-up: race tiered vs flat to
0 hard fails (convergence speed) rather than composition at a fixed budget.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Splits the flat outer-search comparator (-n_fails, fitness) into a tiered
(-n_hard, -n_soft, fitness) so search budget stops being spent polishing
SOFT shape fails (crinkliness/proportion/size/width/edge-too-long/
staircase-volume) while HARD structural fails (missing space, wrong/
required level, level/circulation/vertical connectivity, adjacency,
stairs, covered-outside, storey limits, public access) remain unfixed.
fitness.classify_fail_tier/tier_counts classify every fail string emitted
across fitness.py and graph.py, raising on anything unrecognised so new
fail sites must declare a tier. Validated against all real fail strings in
the checked-in corpus plus every fail-emission call site read from source.
driver.Individual gains n_hard/n_soft (populated from innerloop.Result.
fail_lines); search(use_tiers=...) swaps the comparator when set (default
off, so existing runs are unaffected — inner-loop 0.5^n cliff untouched).
evolve.py exposes --use-tiers / HOMEMAKER_USE_TIERS.
experiments/tier_ab_2g7_3.py runs the acceptance A/B (harbor+maple, 3
seeds, 20k evals) in the background; results pending.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Fixes the stale id()-keyed geometry cache read in parallel staged runs:
substrate_readiness runs in the parent process every tournament/admit
comparison but the parent's score_with_fails (which normally clears
geometry._cache) only runs in pool workers when n_workers>1, so evicted
trees' freed addresses can alias into freshly unpickled ones. Also adds
a defensive clear at collapse_global entry per the bead's recommendation
for the same cache class of hazard.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dq3WAXft8RszMG2CLH7VkU
collapse_global's own commit could relabel a leaf back to the code its
stale share_type names, making share_type == type true again and
resurrecting a multiplicity credit for area never sized for it -- the
commit-door companion to the iio valuation bug. dom.canonicalize_shares()
drops share/share_type whenever share_type != type; called at the top of
collapse_global (covers collapse_global's own commit, 2-opt, and standalone
finish-time use) and _evaluate_full (covers collapse_superposition and
ordinary retype mutations) so the guard is an actual invariant instead of
a per-reader check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dq3WAXft8RszMG2CLH7VkU
Review of fitness.py/solver.py/collapse_cmd.py/innerloop.py/driver.py for
silent score-corrupting bugs (the iio class). Filed with verified repros:
r5a (stale-share resurrection via collapse commit), cvw (parallel staged
stale id()-keyed geometry cache), sd3 (collapse_best keep-better guard
vacuous under baked-in collapse_insearch), pek (shadowed process_storey).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NjyStTdLWFMtdScrpbgQur
Scoped review request for a stronger model to audit fitness.py,
solver.py, collapse_cmd.py, and the collapse_insearch path for
silent numeric bugs analogous to the iio stale-share leak.
Adds a full writeup of the root cause (already fixed in 929be5b) plus a
same-codebase fix-vs-no-fix re-verification: harbor-house qpk-protocol
seeds 1-3 show collapse_insearch=OFF unaffected, but ON diverges by 5-8
fails on 2/3 seeds, non-directionally. Confirms the bug was not merely
theoretical for historical leaf_sharing+collapse_insearch runs, though the
noise is unlikely to have flipped 1ph's aggregate N=20 verdict. Adds a
caveat postscript to §20 (qpk) and files homemaker-py-d86 for the rigorous
historical-commit re-verification this session didn't do.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
_collapse_value and _usage_quality temporarily overwrite leaf.type to probe a
hypothetical candidate code, but graph.leaf_share reads that overwritten type
against leaf.share_type -- so a stale share (left over from a code the leaf
was since retyped away from) spuriously reactivates whenever the probed
candidate happens to equal the old share_type, skewing the Hungarian
assignment's cell value for that (leaf, code) pair. dom.dump/dom.load drops
such stale metadata on reload (dom._emit only serialises share when
share_type==type), so a live search tree carrying it and its dump/reload
round trip fed different values into the same collapse_global call and
landed on different optimal matchings.
Fix: neutralise share_type during the probe whenever the candidate differs
from the leaf's real current type, restoring it in the finally block. The
leaf's own current type still legitimately carries a live share.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
Re-ran the §13.1/§13.2-style per-leaf fail-breakdown diagnostic on real
driver.search_staged runs (budget 20000, seeds 0-2, harbor-house and
maple-court) under the current full default stack (leaf-sharing x3,
depth-balanced, interior-O, share-aware edge cap) -- never decomposed by
category since those defaults were flipped on.
Finding: crinkliness (48%) and size (20.6%) now dominate the residual on
both programmes (~69% combined); construction-completeness fails
(missing space, adjacency, level, connectivity) are down to a small
tail (<=6% each). This revises erc.1's old recommendation to deprioritise
compactness-cuts in favour of leaf-sharing -- leaf-sharing is now fully
deployed and crinkliness is proportionally more dominant than ever, so
DESIGN.md §13.11 recommends reopening a compactness/crinkliness-targeted
construction lever as the next concrete step.
Also files two bugs found while validating the methodology: dumping and
reloading a .dom under leaf_sharing+collapse_insearch does not reproduce
the search's own in-process fail count (homemaker-py-iio), and
run_staged_search.py's own sanity rescore omits the collapse_insearch
override (homemaker-py-7ua). experiments/run_and_capture_91f.py sidesteps
this by capturing the true in-process fails list instead of rescoring
from disk; experiments/diag_residual_91f.py tallies fail categories from
those sidecars.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
Filed from a DESIGN.md/memory review: residual diagnostic on the
current full construction stack (91f), 2-opt polish inside
collapse_insearch (fe2), health-centre stack validation (7xb), and
the parallel/BLAS non-determinism noise source flagged in §14 (b8g).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
Build a torch-differentiable local proxy for the ratio-to-fitness path (exact
port of geometry.py's coordinate recursion + the 5 continuous per-leaf quality
factors, with discrete/structural facts frozen from a real fitness.py
snapshot and the 0.5^n cliff relaxed to a sigmoid) and compare Adam ascent
against nm_search on frozen topologies from programme-house and harbor-house.
Result: ~30-35x slower per unit of search progress than nm_search at both
6 DOF and 36 DOF (per-op torch tensor dispatch overhead with no batching
opportunity, plus snapshot/resnapshot cost on par with a full oracle eval),
and no better quality at matched budget. A step-size sensitivity check
confirmed the flagged 0.5^n cliff risk is real, but autodiff doesn't make the
gradient direction any cheaper to obtain here. Not recommended; kept as
reference only, not wired into innerloop.py. Full writeup in DESIGN.md §34.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
The N=3 A/B (previous commits) found the precision-weighted shape
combination improved both example programmes (harbor-house -1.4%,
health-centre -13.9%), but N=3 is a thin sample by this project's own
standard (xyu/9yx use N=15). Two confirmations:
- N=15, plain search, budget=3000 (mirrors xyu/9yx's own protocol exactly):
both programmes trend NEGATIVE (harbor +6.1%, health-centre +6.6%,
Wilcoxon p=0.044)
- N=15, staged search, budget=20000 (true same-conditions replication --
identical to the original A/B except seed count): both programmes AGAIN
trend negative (harbor +6.6% p=0.15, health-centre +4.7% p=0.48)
The same-conditions replication disagrees with the original result's
direction on both programmes. Conclusion: the N=3 positive signal was
sampling noise, not a real effect -- health-centre's -13.9% was driven
substantially by one seed (71->43 fails) that didn't hold up.
multi_use stays default OFF and is not recommended even as a promising
lever -- this is a clean NULL, closing out both halves of §26's original
multi-use-leaves question (path a was NULL/NEGATIVE, path b is NULL after
replication). Mechanism itself is unchanged, complete, and fully tested.
DESIGN.md §33 rewritten with all three measurements and the honest verdict.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
Follow-up to the previous commit: user flagged that quality_width/
quality_proportion's "stricter of both" (max target, min sigma) combination
for a fused leaf's two codes was an ad hoc hack. Tried two more principled
alternatives and A/B'd all three against the harbor-house/health-centre
example programmes (20k evals x 3 seeds each):
1. stricter-of-both (original) -> health-centre +24.5% worse
2. precision-weighted Gaussian product -> health-centre -13.9% better
3. mixture (max of two Gaussians) -> health-centre +20.4% worse
Landed #2 (fitness._gaussian_product): combining two Gaussian evidence
sources about the same quantity via precision-weighting gives an
intermediate target with a narrower spread, unlike the naive max/min hack.
#3's building block (_clipped_gaussian) is kept, documented, and unit-tested
as a recorded negative alternative -- somewhat counterintuitively, the more
philosophically appealing "let the leaf collapse toward whichever code fits"
mixture model was empirically worse, because max() lets a leaf score 1.0 by
satisfying only the weaker of the two codes' targets.
multi_use stays default OFF -- the precision-weighted result improves both
example programmes on average but isn't the clean sweep needed for a
default flip (harbor-house loses 1/3 seeds). DESIGN.md §33 rewritten with
the full three-way comparison.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
Builds path (b) from §26 -- a leaf permanently serving two DIFFERENT
compatible programme codes at once, extending leaf-sharing's same-code
multiplicity mechanism to different-but-compatible codes. Architect-declared
`co_locate` pairs (validated against interchangeable()'s S1-S4 bounds, no
transitive closure so the b3v chain problem can't recur), threaded through
graph.py's checks via a new leaf_codes() resolver and fitness.py's quality
terms (additive size, stricter-of-both width/proportion). Construction-time
only, gated behind `multi_use` (default OFF, bit-identical when off).
End-to-end A/B (20k evals x 3 seeds x 2 programmes) came back net negative:
harbor-house -4.0% but health-centre +24.5% worse (3/3 seeds), because
fusing different codes' shape targets via stricter-of-both can impose a
tighter joint constraint than either code needed alone, which the tightly-
packed health-centre programme can't absorb. Written up as DESIGN.md §33;
multi_use stays default OFF, no default-flip recommended.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
run_9gp_ab.sh never threaded a worker count through run_staged_search.py, so
every §12.3 arm ran at n_workers=1 (serial) — the one mode §12.4 already
proved byte-for-byte reproducible even before the completion-order
determinism fix (that bug is ProcessPoolExecutor as_completed-only).
Spot-checked empirically: same config run twice gave identical fail counts
at every checkpoint. Closes homemaker-py-h10 as confirmed-null without
re-spending the ~8 core-hours a full sweep re-run would cost.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
N=15 seeds, xyu's own ruin_recreate ON/OFF protocol, against a real 20-room
diverse programme instead of programme-house's duplicated-code sweep: 2.2%
mean-fails delta, p=0.40 two-sided — much weaker than xyu's own inconclusive
6.4%/p=0.059 reading at the same room count, and converging with
harbor-house's null-to-negative result instead. Closes the diversity-axis
gap xyu's larger-N pass could not reach; enable_ruin_recreate stays OFF on a
now-broader evidence base. Issue closed.
A small primary-care health centre: 19 distinct, individually-sized room
codes at n=20 room instances (only one deliberate duplication: two public
WCs), filling the gap between programme-house's duplicated-count sweep
sizes and harbor-house's out-of-range 37 real-diversity instances. Widths
are deliberately tiered (>1.3x gaps at three boundaries) so the auto-derived
interchange relation resolves to three bounded utility/office/clinical
classes instead of one whole-building chain, which a first pass produced.
experiments/run_9yx_sweep.sh repeats xyu's ruin_recreate ON/OFF protocol
(budget 3000, 4 workers, N=15 seeds) against this programme.
Extends y51's n=18 synthetic sweep (strongest of four sizes at N=10) to
N=15 seeds, matching f1d's own confirmation sample size. Effect shrank
(9.3%->6.4%, two-sided Wilcoxon p 0.098->0.059) but didn't evaporate or
reverse — an ambiguous middle case, not a clean confirm or null. Refiled
option (b) (non-synthetic third example programme) as homemaker-py-9yx
since extending N alone doesn't address the interchangeable-room-code
confound §24 already flagged. enable_ruin_recreate stays default OFF.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
N=15 driver.search sweep of construction_beam_width 1 vs 4 (protocol
identical to c94's original 5-seed run, DESIGN.md §29): 6W/4L/5T, mean
fails 57.0->56.6, Wilcoxon p=0.84. Excluding seed 2's outlier the mean
flips slightly negative (56.3->56.9), confirming the §29 5-seed
"improvement" was that one outlier. construction_beam_width stays
default 1 on confirmed rather than precautionary grounds. DESIGN.md §30.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
The raw-seed diagnostic (single constructed tree, byte-identical across
beam widths) was wrongly taken as proof a full driver.search run would
also be byte-identical. Actually running it (5 seeds/programme, budget
1500, n_workers=1) shows harbor-house diverges: 2 wins/1 loss/2 ties vs
greedy — a full bootstrap population hits beam-vs-greedy tie-breaks a
lone raw seed sample missed. programme-house stayed tied 5/5. Verdict
corrected from a confident null to inconclusive/mixed on harbor-house;
practical disposition unchanged (construction_beam_width stays default 1).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
operators._assign_adjacency_aware gains beam_width (default 1 = exact
prior greedy behaviour), threaded through constructive_topology/
lift_base_to_storeys/driver.search/search_staged as
construction_beam_width. Verified functioning on an adversarial
synthetic case, but byte-identical raw-seed output to greedy at every
width tested (1/4/8/20) on both example programmes -- no headroom for
the beam to find on this repo's programmes. DESIGN.md section 29.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8agJBT2ZpmF3ErW7wi2wY
homemaker-py-9wi's 2-opt adjacency polish (collapse_global's local_search
kwarg) was validated on harbor-house alone; this closes homemaker-py-cdl
by extending the A/B sweep to programme-house (34 more files, 46 total):
0 regressions, 2 improvements, rest identical.
collapse_global's own default stays False since it also runs every
fitness eval via collapse_insearch (qpk) on the unmerged tree, where the
2-opt pass would add cost to a hot path the sweep never measured.
Instead default it on at the two one-shot finish-time call sites:
homemaker-collapse --local-search, and a new homemaker-evolve
--collapse-local-search wired through driver.collapse_best's
**collapse_kw.
The Jacobi adjacency relaxation in collapse_global (94g) re-solves a linear
assignment each round holding neighbours' labels fixed from the previous
round, which can 2-cycle between labellings that satisfy zero adjacency
requirements even when a fully-satisfying permutation exists (proved by
test_two_opt_polish_escapes_jacobi_plateau on a minimal 4-cell chain).
Fitness._two_opt_adjacency_polish runs after the Jacobi fixpoint and tries
swapping the labels of every same-level pair of supply leaves, keeping a
swap only on strict improvement -- monotone by construction. Gated behind
collapse_global(local_search=...) / homemaker-collapse --local-search,
default off pending a broader sweep (homemaker-py-cdl). Swept the 11
harbor-house evolved-*/3m/materialised .dom files: 0 regressions, 1 real
improvement (evolved-anneal-3M.dom 21->19 fails).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Locates the threshold f1d's programme-house/harbor-house split implied,
using four synthetic sizes (10/14/18/22 rooms) derived from programme-house
by scaling its bedroom+ensuite module count, since no natural third example
programme sits between the two. Results are noisy and non-monotonic (n=10
mild win, n=14 clean null, n=18 strongest trend at p=0.098, n=22 near-null)
rather than a clean decay with room count -- documented in DESIGN.md #24.
enable_ruin_recreate stays default OFF; filed homemaker-py-xyu as a
low-priority follow-up (larger-N at n=18, or a non-synthetic third example).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds operators.mutate_ruin_recreate: un-divides one wing of a storey and
rebuilds it with the same adjacency-aware constructor the seeders use
(_assign_adjacency_aware, generalised with a new `scope` param), seeded
from the surviving circulation bordering the wing. Gated behind
enable_ruin_recreate (default off) / --ruin-recreate, same pattern as
reassociate/bridge_circulation.
A/B (qpk protocol, DESIGN.md §23): initial uniform-weight run was
underpowered (fired ~1/32 children), null. A weight=3.0 follow-up
(_MUTATION_WEIGHTS["ruin_recreate"]) showed a statistically significant
win on programme-house across 15 seeds (8W/1L/6T, mean fails 7.07->6.00,
Wilcoxon p=0.041) but no consistent effect on harbor-house across 8 seeds
(3W/2L/3T). Kept default off pending a size-threshold follow-up.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Combined follow-up to 8sh (DESIGN.md §22): raised bridge_circulation's
_MUTATION_WEIGHTS entry to 2.0 (lj3) and re-ran the qi6/qpk-protocol A/B
at 4x the sample size (qjg) in one sweep, since the two variables were
confounded if tested separately. Result is null in the opposite direction
from 8sh's small-N signal -- no total-fail benefit (p=0.71 programme-house
N=20, p=0.69 harbor-house N=12) and a higher rate of trajectory-divergence
-induced new not-connected fails than at the original uniform weight.
Reverted the weight bump; enable_bridge_circulation stays default off.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDZjAATDWW1xFfc7xnJqSt
Adds operators.mutate_bridge_circulation: retypes the cheapest path between
two disconnected circulation components to circulation, directly clearing a
'level N not connected' fail instead of relying on the qi6 graded comparator
key (measured negative, DESIGN.md §18). Gated off by default via
driver.search's enable_bridge_circulation flag and evolve.py
--bridge-circulation, mirroring enable_reassociate's clean-toggle pattern.
qi6/qpk-protocol A/B (DESIGN.md §21) is directionally positive but mixed at
N=3/N=5 (never worse on total fails; clears 2/5 baseline not-connected fails
vs qi6's 0/4; one seed's RNG-trajectory divergence adds 2 new not-connected
fails) — kept default off pending a larger-N confirmation sweep
(homemaker-py-qjg) and a mutation-weight bump experiment (homemaker-py-lj3).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDZjAATDWW1xFfc7xnJqSt
20-seed programme-house sweep (vs the original 5) resolves the qpk A/B's
mixed 3/5 result as small-sample noise around a true small positive: mean
fails 7.95->7.10 (~10.7%), 11W/6L/3T, paired t-test p~0.028. Flips
collapse_insearch's default from OFF to ON in evolve.py and driver.py
(_overrides_for/_fitness_for/_evaluate/search/polish_finish); opt out with
--no-collapse-insearch. fitness.Fitness itself is unchanged.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDZjAATDWW1xFfc7xnJqSt
conn_grade ON vs OFF (qpk protocol, experiments/run_qi6_ab.sh): harbor-house
(budget 2500, seeds 1-3) byte-identical output in every seed — the secondary
comparator key never fired. programme-house (budget 3000, seeds 1-5) 3/5 seeds
tie exactly; seeds 1/2 diverge to a different topology but the fail delta is
adjacency/crinkliness/width/access/size, never connectivity. Zero of 4 cases
where a not-connected fail was present got cleared by the grade.
Mechanism (b)/(c) (graded proximity as tertiary comparator key) is falsified,
not just unconfirmed. Kept default OFF (already was). Closed qi6; filed
homemaker-py-8sh for the remaining candidate (mechanism (a): an explicit
insert/relocate-circulation operator that doesn't depend on the search
stumbling onto a fail-count tie).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GDZjAATDWW1xFfc7xnJqSt
