Answering "are we clean". Generic namespace: yes. Usage namespace: no.
FINISH §39.4. The first sweep missed sites, found by a full re-grep:
graph.py's free-area budget, operators.py host-preference / keep-type /
repair-candidate, fitness.py's ("l","c","k") public-access test, bubble.py's
generic adjacency reference, and -- the important one -- cpsat.py, which was
still matching adjacency by raw startswith. graph.code_matches_requirement is
now the single public answer to "does this leaf count as the thing the
programme asked to be next to", shared by has_adjacency, has_vertical_connection
and cpsat.
RETRACT §39.5. It concluded 2g7.5's CP-SAT seeder win did not survive the
correction. That was wrong. The cause was the missed cpsat matcher above: the
exact solver was optimising a different relation than the scorer checked, so a
failing test reporting an incomplete sweep was misread as a baseline shift.
Re-measured over 6 seeds, cpsat now wins on both programmes (harbor 102/92,
maple 156/154). xfail removed.
REAL BUG UNDERNEATH: CP-SAT was never deterministic despite
num_search_workers=1 and a comment claiming it. neighbors[slot] is a set of
dom.Node, which hashes by id() -- a memory address -- so raw iteration made the
model-build order vary and CP-SAT returned a different equally-optimal
assignment each run (measured 194/180/171/182 over four identical aggregates).
sorted() on the slot indices fixes it. Also paired the wall-clock cap with
max_deterministic_time (solves run ~124ms against a 2s cap, so nothing was
timing out -- latent hazard, not the cause). solve_room_labels is now
reproducible on every captured instance; constructive_topology on the cpsat
path still is not, filed as homemaker-py-fdp (plausible contributor to b8g).
§39.6 THE SECOND NAMESPACE. Usage prefixes b/t/l/k (bedroom/toilet/living/
kitchen) classify programme codes by first letter and stay prefix-based by
design, but they are not inert: has_circulation deletes graph edges from them.
Four corpus rooms are misclassified by spelling -- la1 "Laundry Room" and li1
"Library Corner" as living, br1 "Staff Room" as bedroom, tr1 "Treatment Room"
as toilet. Measured on a health-centre seed: tr1 loses its edge to the adjacent
O, br1 loses its edge to t10 "Staff WC" -- both feed the connectivity fails §38
found persisting. Filed homemaker-py-sel; an explicit usage: key is the fix,
but it changes fitness for correctly-spelled programmes too so it needs its own
A/B.
DOCS. README gains a "Room codes and reserved names" section; CLAUDE.md and
AGENTS.md gain the same summary for agents. audit_programme_config.py now
reports the usage class each code picks up alongside the namespace and
satisfiability checks. DESIGN §37.2's note calling the c/o/s quirk "existing
product behaviour, not a bug" is annotated as superseded.
Corpus audit: zero generic-namespace violations across all ten example
programmes. 346 passed, same 7 pre-existing fixture failures, lint unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB
Supersedes the previous commit's approach. Renaming harbor's four colliding
codes fixed one programme; tightening the matching rule fixes the rule, so a
room may be called anything. cr1/of/st1/st2 are restored and the examples are
byte-identical to their pre-§39 state -- which also means existing .dom
artefacts (evolved-3M*) stay valid, so migrate_ju3_rename.py is deleted.
The rule: Urb has exactly three GENERIC structural types (get_space_types:
qw/C O S/), the leaves the search creates. Measured across the corpus: 154 C,
110 O, 1 S, not one lowercase generic -- while every programme code is
lowercase, including single-character ones (r, t, m, n). Case is the
discriminator, not length. Every generic test was type[0].lower() in (...), a
case-insensitive PREFIX that swept up any programme code starting with those
letters; they now match the generic set exactly. 30 sites across dom, fitness,
graph, operators, programme, shapecurve and bubble.
NOT applied to the SEMANTIC prefixes: l/k/b/t classify programme codes by first
letter (graph.py builds bedroom<->toilet and kitchen<->living relations from
them) and stay prefix-based. Where the namespaces were mixed in one expression
they were split -- has_circulation's ("b","l","k","c") is three semantic
prefixes plus dom.is_circulation; access()'s ("l","c","s") is semantic l plus
the generic circulation set.
New: dom.GENERIC_{CIRCULATION,OUTSIDE,TYPES} + is_generic(); fitness.
_generic_class(), replacing the _t0 dispatch in quality_size/quality_width/
quality_proportion/value_rate -- the four terms that mattered most and that a
first sweep missed, since they dispatch through a t0 variable rather than an
inline test. graph._adjacency_target resolves a generic adjacency requirement
(programmes write "adjacency: [c, o]") to the generic set while every other
requirement keeps Perl's prefix semantics.
Two subtleties: S is in both generic sets but takes the OUTSIDE parameter
families -- a first translation tested circulation first and silently gave S
the circulation params, caught by test_get_space_params_sahn_proportion. And
validate_codes survives, narrowed to a code spelled exactly C/O/S, which is a
genuine ambiguity; merely starting with c/o/s is now fine.
Invariant asserted as a test: test_scoring_is_invariant_under_programme_code_
spelling relabels one tree and its config together and re-scores. Bit-identical
across 12 comparisons (6 seeds x collapse on/off).
Re-baseline (seed 1, 20k, original names): 58 fails (15 hard / 43 soft) against
the real 37-instance programme, with cr1 at 79.1 m2 vs declared 80 (was 32.9
and 17.1), of/st1/st2 all present and in band, and one fail naming any of them.
57 -> 58 on a 5-instance-harder programme is within noise: "did not regress".
Fallout (§39.5): 2g7.5's CP-SAT seeder win does not survive. Over 6 seeds --
harbor real 102/114 (cpsat loses), harbor old-effective 98/99 (tie, so the win
was already marginal), maple-court 156/144 (cpsat wins). maple is the control:
the solver did not regress, harbor's programme changed. Test xfail'd with that
reason plus a maple companion; both assign_solver flags stay default off.
Filed homemaker-py-w6x to re-check other narrow-margin harbor A/Bs.
345 passed, 1 xfailed, same 7 pre-existing fixture failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB
Closes homemaker-py-ju3. DESIGN.md §39.3.
The class: key from the bead's design was deliberately NOT built. Auditing
every use of the prefix rule first showed it runs deeper than c/o/s -- l/k/b/t
carry real adjacency semantics (graph.py builds bedroom<->toilet and
kitchen<->living relations from first characters) -- so re-plumbing the type
system would invalidate the whole corpus and every baseline, for a problem
whose damage is the silence, not the convention. Two findings made the smaller
fix sufficient: no corpus programme has ever declared a bare c/o/s code, so
check_space_counts' skip only ever discarded declared rooms; and nothing
references harbor's four codes in any adjacency or co_locate list.
- programme.validate_codes raises on a reserved-prefix code, with the full
explanation. Called from BOTH parse paths (programme._parse_spaces and
fitness.Fitness._load_programme parse conf["spaces"] independently, so
validating one would leave the other door open). l/k/b/t stay unreserved.
- harbor-house and harbor-house-l0 renamed: cr1->fr1, of->ao, st1->gs1,
st2->gs2. New prefixes are unused in harbor and semantically neutral, and
the two storage codes still share a prefix, preserving the structure
evaluate_building's per-code plot-ratio term depends on. name: unchanged.
- experiments/migrate_ju3_rename.py migrates .dom files written before the
rename (--check dry-runs). Pre-rename artefacts, notably evolved-3M*.dom,
must be migrated or their leaves read as unmatched generics.
- test_collapse_global's c/o/s exclusion test now uses a generic C leaf, which
is what the exclusion is actually for; it previously relied on a programme
code colliding, which is no longer possible.
Re-baseline (seed 1, 20k evals, same settings as §38's run): 57 fails against
the 32-instance effective programme -> 55 against the real 37-instance one,
with all five previously-lost room instances now placed inside their declared
sigma bands (fr1 87.2 vs declared 80, was 32.9/17.1; ao/gs1/gs2 were absent
entirely) and no failure naming any of the four codes. At one seed each,
57 vs 55 is within noise -- the robust result is the room placement, not the
count. Historical harbor numbers are not comparable to post-ju3 ones; filed
homemaker-py-t3s to restate 2v1's acceptance figure once evolved-3M is
migrated.
346 passed (+10 new), same 7 pre-existing fixture failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB
Answers "are any config requirements actively fighting the engine". One is.
§39.1 NEGATIVE (recorded): no room spec in any corpus programme is internally
contradictory. Using shapecurve.leaf_constraints' validated FAIL_THRESHOLD
inversions, every code admits an (area, aspect) satisfying size, width,
proportion and crinkliness at once, and none needs more than one exposed side.
The "estimated targets are mutually unsatisfiable" hypothesis is falsified.
§39.2 SEVERE: Urb's type system is prefix-based (c = circulation, o/s =
outside) and programme codes share that namespace. A code starting with those
letters is silently reinterpreted, with three unannounced consequences:
check_space_counts SKIPS it outright (never required, no missing or too-many
fail); get_space_params returns generic *_circulation/*_outside params before
consulting self.spaces; and is_circulation/is_outside flip, changing value
rate, crinkliness exemption, and whether it supplies daylight to neighbours.
harbor-house is affected (maple-court, health-centre, programme-house are
clean): cr1 "Common Room with Fireplace" has all three declared targets
overridden (size 80.0 -> 0.0/14.0) and is valued at 50/m2 not 300; of x2 and
st1/st2 lose width/proportion and are treated as outside space. 5 of 37 room
instances (14%) are silently optional.
Measured: the two cr1 leaves converged to 32.9 and 17.1 m2 against a declared
80, with no too-many-spaces fail despite count:1; of/st1/st2 are absent from
the result with zero fails. Compounds with §38.2 -- the largest room in the
programme sits on the wrong side of the x6 circulation value gap, so the
objective is paid to shrink it.
Benchmark validity: every harbor-house fail count in this document was
measured against a 32-instance effective programme, not the 37 its config
declares.
Adds experiments/audit_programme_config.py (namespace + satisfiability
reports). Filed homemaker-py-ju3 (P0). No src changes; 336 passed, same 7
pre-existing fixture failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MJ84Feep79Hhm3E4zZJmnB
Created homemaker-py-8oq (review the 2g7.7 LLM-repair plan with a more
capable model) and homemaker-py-57z (live acceptance-benchmark follow-up,
blocked on ANTHROPIC_API_KEY availability), linked as blockers/dependents
of homemaker-py-2g7.7 per the planning session on 2026-08-05.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uCyZJCh5mBuA2yEFxgQpo
homemaker-py-7ua is fixed but only covers run_staged_search.py's
LEAFSHARE/MULTIUSE-gated rescore path. Baseline (LEAFSHARE=0/MULTIUSE=0)
runs still MISMATCH for the same underlying reason -- driver.search_staged
has no collapse_insearch param and always runs with it on -- tracked as
the new homemaker-py-4ok.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uCyZJCh5mBuA2yEFxgQpo
driver.search_staged has no param to disable collapse_insearch, so its
inner evaluator always runs with search()'s collapse_insearch=True
default. The final _native_score() sanity rescore loaded patterns.config
fresh with no override, so it silently diverged from search-time conf
whenever leaf_sharing was on, producing a false MISMATCH. Verified with
a smoke run: pre-fix MISMATCH, post-fix OK with identical scores.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uCyZJCh5mBuA2yEFxgQpo
driver.collapse_best built its evaluator with _fitness_for's default
collapse_insearch=True, so collapse_finish's base_fails/cand_fails were
both measured through score_with_fails' own auto-collapse pass -- base
silently equalled collapsed on 5/5 probed files, making the "keep only
if fails don't increase" safety guard vacuous and understating 94g's
real effect in logs. fitness.collapse_finish now forces canonical
(collapse_insearch=False) scoring for its own measurement regardless of
self's config; collapse_best now builds its evaluator canonically too
(matching what homemaker-fitness reports for the written .dom) and
threads max_share/conn_grade through. Same-family fix in
search_annealed's no-polish-budget rescore branch, which silently
defaulted to collapse_insearch=True via _evaluate's default.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uCyZJCh5mBuA2yEFxgQpo
Research spike (DESIGN.md §37.8): classical rectangular dualization assumes
one-vertex-one-rectangle, which breaks on harbor's circulation hub (an
emergent-shape multi-leaf region, not a fixed single module) and is overkill
on the room-only adjacency graph (a trivial 3-edge matching already fully
satisfied by §11.7's seeding). No literature precedent for the multi-storey
stacking constraint either. Not prototyping; no code changes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014uCyZJCh5mBuA2yEFxgQpo
CP-SAT assignment does not clearly beat greedy end-to-end: worse on
harbor-house hard fails, roughly a wash on maple-court, and the reassign
operator fired in only 1 of 18 runs. Keeps assign_solver="greedy" and
enable_reassign=False as defaults; closes the bead's remaining acceptance
criterion. Raw results in experiments/results/.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Adds src/homemaker_layout/cpsat.py (OR-Tools CP-SAT) as an exact alternative
to operators._assign_adjacency_aware's greedy/beam room-code placement,
wired in as assign_solver="greedy"|"cpsat" (EXPERIMENTAL, default "greedy",
byte-identical to before) through constructive_topology/lift_base_to_storeys/
driver.search, plus a new operators.mutate_reassign in-search repair
operator (driver.search's enable_reassign=False default, mirrors
enable_ruin_recreate). Both found and fixed a resize-fragility bug (a
second CP-SAT pass against settled geometry, operators._cpsat_relabel_settled)
and a CP-SAT symmetry-blowup stall (explicit interchangeable-code grouping).
Seeder-level A/B on harbor-house is a solid, low-noise positive (~13% fewer
real fitness-scored secondary-adjacency fails, 10 seeds). Full driver.search
A/B is only pilot-scale (budget=3000 vs the bead's own 20k target) and
inconclusive -- both flags stay default-off pending a larger-N confirmation.
Full writeup: DESIGN.md §37.7. Bead left in_progress (own acceptance
criteria not fully met); homemaker-py-5bv tracks the deferred post-collapse
repair item.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Generalise shapecurve.py's DP to process dom.levels(root) bottom-up per
storey instead of assuming a single free tree. A divided node's split is
free only per solver.free_branches' own criterion (below is None or
undivided there) -- geometry.coordinate always mirrors a below-linked
node's corners from the storey below regardless of whether that storey's
counterpart is divided, so every free region at any storey reduces to the
exact same single-region problem the pre-existing _check/realise already
solved. New _region_roots finds below-fixed leaves (checked directly,
gridless) and below-fixed-box/free-split fringe nodes per storey;
_solve_all_levels realises each storey before checking the one above and
snapshots+restores on any infeasibility, preserving solve()'s all-or-nothing
and is_feasible()'s never-writes contracts across the whole tree.
eligible() now allows any storey count.
Validated on the real (non-de-risked) examples/harbor-house: 200 random
2-storey topologies, DP-vs-NM agreement 99.5%, 0 false negatives, 117.7x
speedup (DESIGN.md §37.6). Full suite 397 passed.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Adds shapecurve.is_feasible() (a non-mutating refactor of solve()'s check
phase) and a shapecurve_prune flag composing the DP's exact feasible/
infeasible verdict with operators.predicted_shape_fails' existing heuristic
prune: DP-feasible vetoes a heuristic prune outright; DP-infeasible only
hard-prunes when the incumbent already has zero total fails (exact, since
infeasible proves the shape-fail floor is >=1); otherwise defers unchanged
to today's heuristic threshold. Conservative by design since a wrong prune
is unrecoverable.
Validated 0/400 false negatives across two structurally distinct plots
(harbor-house-l0 + a newly-added programme-house sweep, the first genuinely
non-rectangular plot this DP has been checked against). The real
driver.search A/B on harbor-house-l0 measured NULL (byte-identical off/on)
for a root-caused, pre-existing reason: predicted_shape_fails rarely
triggers organically at this scale, so neither new branch had an opening to
fire -- not a defect in this change. Full writeup: DESIGN.md §37.5.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Promotes the validated shape-curve DP (experiments/shapecurve_spike.py,
2g7.4, DESIGN.md §37.2) from a reference-only spike into
src/homemaker_layout/shapecurve.py, and wires it into driver._evaluate as a
warm-start for innerloop.optimise: when eligible (single storey, no
leaf_sharing/superpose/max_share/multi_use) and no caller-supplied x0, the
DP's exact shape-feasible ratio point is written onto the tree before NM
runs, off by default (shapecurve_warmstart=/--shapecurve-warmstart).
Caught and fixed a latent bug promoting the spike: realise() could leave
numpy.float64 in `division`, which yaml.safe_dump can't serialise — the
original spike never round-tripped through dom.dumps so this was never hit.
A/B on harbor-house-l0 (experiments/ab_shapecurve_warmstart.py, budget=2000,
5 seeds): mean total fails 16.6 (on) vs 19.6 (off), ~3.5x mean fitness
improvement; mean hard-fail count alone was a noise-level wash at this
sample size. Full writeup in DESIGN.md §37.4.
Deliberately deferred to new tracked beads (children of 2g7): DP-exact hard
pre-filter (wkh), multi-storey below-link support (koo), leaf_sharing/
co_type modelling (tym), true skew-quad polygon algebra (ekc) — 6xh stays
in_progress pending those.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Composer half of the ground-truth bead: examples/harbor-house/drawings/
harbor-house 1.svg turned out to be a Bonsai/Blender render of 3m.dom's own
IFC (32 IfcSpace paths == 3m.dom's upper-storey leaf count), not a human
trace, so no usable reference exists yet -- this builds and tests the
pipeline that will consume one once traced. compose.py parses storey-N
Inkscape layers of cut-lines + labels against a boundary-stub .dom (plot/
height/elevation only, no room shapes to keep aligned across storeys) and
recursively detects guillotine cuts, mirroring geometry.py's own
division-line algebra; non-slicible regions and label mismatches are
reported by location rather than guessed at. homemaker-compose CLI added.
Renamed dom._link to public dom.link since compose.py needs to re-link from
outside dom.py. Full design writeup in DESIGN.md sec 37.3; actual human
tracing of harbor-house/programme-house is tracked as follow-up under
2g7.1, still open.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
Prototype + validation for an exact size/width/proportion feasibility DP
over a frozen slicing topology, replacing the ~80-200 eval Nelder-Mead
inner loop's approximate answer to the same question with one bottom-up
pass (experiments/shapecurve_spike.py). Leaf feasible regions are exact
FAIL_THRESHOLD-inversions of fitness.py's quality_size/width/proportion;
internal-node composition runs on a shared discretised grid.
Validated on harbor-house-l0 (experiments/validate_shapecurve.py, 200
random topologies vs NM minimising shape-fail-count directly): 99.0%
agreement (0 false negatives), 93.6x speedup at grid_n=150, plot-level
bbox approximation error quantified at +7.5% (root-causing both observed
false positives). All three acceptance criteria cleared -- see DESIGN.md
§37.2 for full results and the caveats/scope not covered (multi-storey,
leaf_sharing/co_type, true skew-quad regions). Kept as a reference spike,
same status as experiments/autodiff_spike.py (§34); production wiring
into driver.py filed as homemaker-py-6xh.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LSwQwpEaHFBkeVSDDWd75S
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
