§13.8 verdict was positive and monotone-harmless, so default the share-aware
edge-too-long cap to leaf_sharing when share_edge_cap is unset — mirrors the
pll bal+share and §13.6 interior_outside default flips. Explicit
share_edge_cap=False still reproduces the pre-flip control arm.
- fitness.Fitness.__init__: cap defaults to self._leaf_sharing when the conf
key is unset (None); explicit True/False honoured.
- run_staged_search.py: pin conf["share_edge_cap"] = share_edge in both A/B
arms so SHAREEDGE=0 stays a clean control post-flip.
- tests: control arm now pins share_edge_cap=False; new
test_edge_cap_defaults_on_under_leaf_sharing guards the flip.
- DESIGN.md §13.9: rebaseline §13.x floor (maple 80.3→74.0, harbor 34.7→31.0).
Non-sharing runs untouched: programme-house control re-score reproduces
bit-for-bit. 222 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
§13.7 flagged edge-too-long as harbor's top fail class. Dissection showed the
bulk are a leaf-sharing REPRESENTATION ARTIFACT: a share=k leaf aggregates k
same-code rooms, so its walls run ~k× the flat 8 m cap purely for being big —
the same §13.3 leak (size/missing relaxed for shared leaves) on the wall measure,
since edge_cost/outside_edge_cost ignored leaf.share.
Fix: Fitness._edge_cap(*leaves) scales the 8 m cap by the largest type-guarded
leaf_share among adjoining leaves, mirroring quality_size's k×target; non-shared
leaves keep the flat cap so genuine narrow/oversize pathologies stay flagged.
Gated behind a share_edge_cap config knob (SHAREEDGE env), default OFF so the
§13.x controls reproduce.
A/B (full Phase-8 stack, staged, 20k evals, seeds 0/1/2): control reproduces
§13.7 (maple 80.3 exact, harbor 34.7≈34.0); share-aware arm maple 80.3→74.0
(−7.9%), harbor 34.7→31.0 (−10.6%), zero regressions across 6 seeds. Positive
and monotone-harmless (only ever removes a false-positive fail). Verdict:
recommend default-ON; follow-up issue flips the default + rebaselines the floor.
Tests: 6 new unit tests for _edge_cap (221 pass).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JygRv4n2dcyDQqMiDRe7TN
§13.6/ld2 verdict: interior-O light-well seeding is net-positive — harbor
-16.4% (all seeds improve), maple net-neutral (-2.8% mean, no programme
regresses). Mirror the pll bal+share flip: default interior_outside
False->True in driver.search/search_staged and operators.constructive_topology/
lift_base_to_storeys (outside_divisor stays 3). The experiments INTERIORO
A/B override is unchanged. test_interior_outside_… now pins the peripheral
baseline to interior_outside=False explicitly. 215 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Seed O as interior light wells (most-landlocked leaves first, count scaled
by room count via outside_divisor) instead of one peripheral O, attacking the
erc crinkliness residual: seed diagnostic confirms every crinkliness fail is
under-exposed (landlocked), none over-exposed.
A/B (20k evals, seeds 0/1/2, bal+share stack, §13.6): control reproduces §13.5;
interior odiv=3 gives harbor -16.4% (all seeds improve) and maple -2.8%
(net-neutral). Default-optimal divisor 3 found by seed sweep (6 was null).
Lever default OFF; default-ON flip tracked as erc.8.
- operators: interior_outside + outside_divisor through constructive_topology,
lift_base_to_storeys, _assign_adjacency_aware (fix n_circ budget for >1 O)
- driver.search/search_staged threading; run_staged_search.py INTERIORO/ODIV env
- test_interior_outside_seeds_landlocked_wells_and_scales_count
- experiments/run_interioro_ab.sh; DESIGN.md §13.6
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
erc.7/§13.5 verdict: depth_balanced + leaf_sharing (factor 3) is the
winning Phase-8 stack. Flip the three knobs to default-on so
homemaker-evolve inherits them; env-var A/B overrides (DEPTHBAL/
LEAFSHARE/LEAFSHAREFAC) unchanged. 214 tests pass, no snapshot churn.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_grow_leaves grew a random caterpillar, so equal-target rooms landed at
wildly different binary-tree depths — the depth-driven size maldistribution
Diagnostic B (§13.2) localized (same code at 0.05x and 14.7x target). The
depth_balanced flag always splits a shallowest leaf instead, growing a
near-complete tree so the proportion-aware sizing pass hits each target with
cut fractions near their proportional value.
Floor probe (diag_depth_balance.py): depth spread collapses 7->1, the giant
ratio falls (maxR 12->8 harbor / 16->6 maple), %undersize 54->25 / 42->22,
and the achievable floor drops -12% harbor / -11% maple at EQUAL leaf count.
Additive with leaf-sharing (bal+sh3 beats §13.3 share3-alone). Default OFF,
214 tests pass; threaded through driver.search/search_staged and exposed via
DEPTHBAL in run_staged_search.py. End-to-end 20k A/B running.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the area-derived share recovery with explicit, type-guarded per-leaf
multiplicity: construction stamps leaf.share=k and leaf.share_type=code; the
fitness (graph.leaf_share) honours k only while leaf.type==share_type, so any
retype/undivide auto-invalidates a stale share — no operator resets, and a
small leaf cannot retype its way into covering rooms it does not provide. Two
Node fields survive the whole search via deepcopy (genome.decode is unused in
the hot path); .dom emits `share` only on a live shared leaf.
This closes the §13.3 missing-fail leak: floor probe missing 17–44 → 0, and the
achievable floor drops −39% harbor (120.3→73.3) / −32% maple (194.7→133.0) with
no re-emergence as size fails.
Flag threaded through driver.search/search_staged → constructive_topology /
lift_base_to_storeys, exposed via LEAFSHARE/LEAFSHAREFAC in run_staged_search.py
(injects the objective into inner-loop + final-score fitness so both A/B arms
share one programme dir). run_leafshare_ab.sh runs the staged 20k A/B.
Smoke-tested end-to-end (harbor, factor 3, re-score OK). 214 tests pass;
default-OFF reproduces baseline.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Same-code rooms collapse into fewer, larger SHARED leaves so the ~1.8/leaf
shape tax (§13.1) is paid once per group. Multiplicity k is recovered from
area (k=clamp(round(area/target),1,max_share)) — no genome change — and used
in two default-OFF sites: graph.check_space_counts counts coverage (Σk vs
req.count) so one leaf covers several rooms without a missing fail, and
fitness.quality_size centres on k×target (σ scaled by k). Construction:
operators._share_rooms groups instances; _size_divisions_from_targets sizes
shared leaves to k×target via leaf_mult.
Floor probe (experiments/diag_leaf_sharing.py, harbor+maple, seeds 0/1/2,
+innerloop): total fails −27% harbor / −16% maple at share3, shape factors
fall ~linearly with leaf count (confirms §13.1). Cap: 17–44 missing fails
leak because depth maldistribution (§13.2) keeps shared leaves below k×target
so round() undercounts; inner loop can't close it. Net still positive.
Default-OFF reproduces baseline exactly (214 tests pass). Driver plumbing +
staged 20k A/B remain; §13.3 records the next design fork.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The constructive seeder was never nondeterministic: _assign_adjacency_aware
ends every max/min with a unique leaf-idx tiebreak and uses set unions only
for membership, so iteration order never leaks. constructive_topology(seed=0)
is byte-identical across processes for every example programme. The cited
"sig 4480 vs 16064" was a measurement artifact — Python's builtin hash() of a
str is salted per process (PYTHONHASHSEED), so an identical signature hashes to
different ints run-to-run.
The real run-to-run noise was parallel-only: driver._run_batch admitted futures
via as_completed (completion order), and admit() is order-sensitive (accrues
n_evals per result; keeps the first individual of an equal-key tie as best). A
long parallel run diverged 167 vs 161 fails (maple seed 0). Fix: admit futures
in submission order (block on each result in turn; all still run concurrently),
reproducing the serial admission sequence. Two workers=4 runs are now
byte-identical. Serial (workers=1) was already byte-for-byte reproducible.
Per-seed numbers are reproducible only at a fixed worker count; serial != parallel
is expected (children/iteration 1 vs n_workers changes batch granularity).
- driver: iterate futs in submission order, not as_completed
- test: test_search_parallel_is_reproducible (fails on pre-fix, passes on fix)
- DESIGN.md §12.4: corrected the reproducibility note
Closes homemaker-py-xcy
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Threads circ_divisor (default 3 = unchanged) through
operators.constructive_topology/lift_base_to_storeys and
driver.search/search_staged; env CIRCDIV in run_staged_search.py. Adds
experiments/run_c3g_ab.sh.
Motivation (DESIGN.md §12.3 diagnostic): the maple shape residual is
over-granular construction (73 small leaves -> crinkliness+size). Cheap raw-seed
probe: a coarser spine lowers the SHAPE floor (maple 135->110, harbor 83->66)
but raises access/adjacency, leaving the raw TOTAL floor flat-to-worse. Because
§12.3 showed shape is the HARD residual and access/adjacency are cheap to
repair, only an end-to-end A/B settles whether trading them pays — this is the
plumbing for that run. Tests green (default path byte-identical).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Land the two evidence-supported parts of the re-scoped 9gp capstone as
operators on the existing decoded Node tree (no Polish-expression rewrite),
each default-OFF and measured against the §12.2 leu.2 baseline.
9gp.1 shape-feasibility pre-filter: operators.predicted_shape_fails lays a
topology out at its proportion-aware target geometry and counts shape fails
(size/width/proportion/crinkliness); driver._evaluate prunes clearly-infeasible
topologies before the inner loop (1 eval vs ~80), guarded so nothing that could
beat the incumbent is discarded. search/search_staged feasibility_filter,
feasibility_max_shape_fails (env FEAS/MAXSHAPE), default OFF.
9gp.2 M3 Wong-Liu reassociate: operators.mutate_reassociate adds associativity
(a|b)|c <-> a|(b|c) on same-orientation live cuts — the canonical-slicing move
missing from swap(M1)/rotate(M2), attacking the §11.4/§11.5 reachability
bottleneck. enable_reassociate (env REASSOC), default OFF (weight 0 -> baseline
byte-identical).
Unit tests (operators + driver) green, full suite 211 passed; maple-court smoke
run clean under native fitness. A/B sweep handed off per the plan; DESIGN.md
§12.3 documents the design and the pending measurement.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Size each constructive-seed cut from leaf TARGET areas (division=[f,f] gives
left area-fraction f) and pick each cut's rotation for child squareness — both
derived from target dims, topology/type assignment untouched. Area-only
regressed (slivers); rotation choice is what makes it pay.
End-to-end (20000 evals, 3 seeds, staged): harbor 85.3->74.0 (-13%, best 69),
maple-court 151.7->136.0 (-10%, best 126). PROP=0 reproduces the §11.7/§12.1
baselines exactly. programme-house regresses at fixed budget (deeper local
optimum walls off the undivide restructuring path) but a budget sweep shows
it's convergence speed, not a worse asymptote (PROP=1 reaches 1 fail at 150k).
Default-on (seed_proportion_aware=True, env PROP=1).
cq1: n_storeys now honours storey_minimum, not just level: keys — programme-house
(storey_minimum:2, all rooms level:0) was seeded one storey short and fell
through to plain search. New programme.storey_minimum()/n_storeys_for();
driver.search passes min_storeys to the seeder; search_staged routes on the max.
No-op for harbor/maple; programme-house single-stage 8.0->5.0.
New maple-court best (126) saved as generated.dom. 204 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_assign_adjacency_aware gains fixed_circ (seed the connected-dominating-set from
given circulation leaves) and secondary-adjacency-aware room placement: codes
with the most non-c adjacency requirements are placed first, each onto the open
slot satisfying the most of its requirements against already-typed neighbours
(clustering k1<->da1, da1<->o). lift_base_to_storeys(reqs, adjacency_aware=True)
grows the upper-floor circulation spine off the inherited vertical core and
assigns rooms around it; threaded through driver.search_staged
(seed_adjacency_aware) and run_staged_search.py (ADJ env).
End-to-end staged harbor, 20000 evals, mean total fails over 3 seeds:
ADJ=0 99.0 (reproduces the §11.4 staged lex baseline exactly), ADJ=1 85.3
(-13.7, -14%; best 78). New best harbor configuration overall: staged baseline
99.0 -> single-stage adjacency-aware (§11.6) 90.7 -> staged + adjacency-aware
lift 85.3. Staging and adjacency-aware seeding compose.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
operators._assign_adjacency_aware spends ~one extra leaf per three rooms on a
greedy connected-dominating-set of circulation leaves (read from the geometric
leaf_graph, type-independent), so every room borders a connected circulation
spine and adjacency-to-c + access are satisfied by construction. Default-on via
constructive_topology(adjacency_aware=True), threaded through
driver.search(seed_adjacency_aware) and run_search_scaled.py (ADJ env).
End-to-end single-stage, 20000 evals, mean total fails over 3 seeds:
harbor 110.0 -> 90.7 (-17.5%; ADJ=0 reproduces the §11.2 105 baseline exactly),
programme-house 12.3 -> 9.3 (-24%). Adjacency-aware single-stage harbor (mean
90.7, best 85) beats the §11.3 staged best of 95 — the first Phase-6 fail-count
reduction from seeding. Follow-ups (lift_base_to_storeys, secondary adjacencies)
filed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
genome.signature: ratio-invariant structural topology hash (per-storey tree
shape + cut orientation + leaf types), the cheap stand-in for the 9gp canonical
encoding. driver gains niche_by_signature (one individual per topology, replaces
the fitness-scalar dedup) and restart_patience (soft restart: keep elites,
refill with fresh seeds); SearchResult gains n_distinct_signatures /
diversity_history / n_restarts.
Diversity criterion MET (final-pop distinct ~5/16 -> 16/16). Gate NOT met:
blank-slate programme-house mean fails 12.3(legacy)/12.7(niche)/13.0(restart)
over 3 seeds at 20000 evals; harbor staged 95/94/108. Niching is a tie within
seed noise, restarts strictly worse — falsifies the premise that the
fitness-scalar dedup causes premature convergence. Both flags default-off,
kept for reuse. Epic c4c complete.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Implement a graded proximity comparator key (-n_fails, grade, fitness) behind
a default-off use_grade flag: fitness._leaf_grade / score_with_grade sum
f/FAIL_THRESHOLD over failing per-leaf quality factors; scalar fitness and fail
count stay untouched so the inner-loop 0.5^n cliff (§5.4) is unaffected (0/9
regression check: PASS). Read once per child in driver._evaluate off the
already-optimised tree; threaded through search_staged (Stage 2 only).
Harbor staged A/B (20000 evals, seeds 0/1/2): lex 95/96/106 (mean 99.0) vs
lex+grade 99/98/102 (mean 99.7) — grade wins 1/3, no plateau escape. Premise
falsified: within a fixed fail-tier 0.5^n is constant so fitness still spans
~6 orders of magnitude; grade above fitness displaces that working signal.
Verdict: reject; lexicographic (-n_fails, fitness) stands. Flag kept default-off
for reproducibility / possible reuse as a §11.5 diversity signal.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Search the genome in causal dependency order. Stage 1 evolves a single-storey
base over the level-0 room set (programme auto-derived to a tempdir), ranked
with a substrate-readiness bonus (reserved core × divisible capacity) so the
base is selected as a good substrate, not just a good ground floor (anti-§4.2).
Stage 2 lifts the best base into a full multi-storey design — preserving the
inherited core, instantiating each upper storey's required set by construction —
and searches the deltas with the base mutable at low probability (base_p=0.15).
New: programme.{n_storeys_required,partition_rooms_by_storey,write_stage1_programme},
graph.substrate_readiness, operators.{lift_base_to_storeys,_pick_weighted_by_storey},
base_p threading, driver.search rank_bonus_fn/seed_factory/base_p hooks +
search_staged orchestrator, experiments/run_staged_search.py, tests/test_staging.py.
Result (harbor, 20000 evals, seed 0): staged 95 fails vs single-stage 105
(-10, -9.5%), gain in crinkliness 27->18 + edge 12->8. Anti-bungalow confirmed
(Stage-2 core moves all noop — core inherited, not carved). Programme-house
regression PASS (warmstart-2f4 still reaches whole-pop 1-fail).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make the required programme room set a constructive invariant instead of
something the topology search must stumble onto by random divide+retype.
- operators.constructive_topology: bootstrap seeder that sizes each storey to
its required rooms (partitioned by level; level-free rooms distributed),
+1 core C and +1 O per storey, then assigns types. Stochastic for population
diversity. Wired into driver bootstrap when the programme has required spaces.
- operators.mutate_place_missing: repair op that inserts a missing required
space by dividing a host leaf into [room | remainder]. Lex-safe host ranking
(generic O first, never displace a required room); honours required level.
Weight 2.0 in the mutation mix; noops cheaply once the set is complete.
A/B on harbor-house (20k evals, seed 0, identical config):
old random-bootstrap 133 fails (103 missing, 77%)
new constructive 105 fails ( 12 missing, 11%) -21% total, missing-stack
collapsed; seed head-start 163->139.
§4.10 regression PASS: warmstart-2f4 still reaches a 1-fail population at 50k.
Verdict (DESIGN.md §11.2): construction is necessary and reframes the
bottleneck to quality-fail packing of a complete dense design (crinkliness/
size/access/edge) -> unblocks §11.3 staging, motivates §11.4 graded objective.
Follow-up filed (homemaker-py-s44): adjacency-aware seeding.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Python fitness always pins quality_daylight to 1.0 (URB_NO_OCCLUSION semantics),
but oracle.py was invoking urb-fitness.pl without the flag, causing outside leaves
to receive real sun-model daylight scores and producing a ~9% gap.
Changes:
- oracle.py: add URB_NO_OCCLUSION=1 to score_batch env
- oracle.py: Score.fail_lines now parses structured YAML failures from the
llm-agent-mcp branch and converts them to plain-text equivalents, so
parity tests can compare oracle vs native failure sets regardless of format
- Regenerated all 36 corpus .score/.fails files with URB_NO_OCCLUSION=1
(no .score files are tracked in git; the script generates them locally)
- All 183 tests pass; closes homemaker-py-gpx
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a compound operator (e.g. level_compound_fix) creates a new
internal node and explicitly sets its division ratio, that ratio was
silently overridden: warm_x0 received parent.ratios which had no entry
for the new node, so Nelder-Mead started at the default 0.5 instead
of the operator's intended 0.25 (for rrl/rrr). Result: NM evaluated
the compound topology at the wrong geometry and scored 3 fails instead
of 1 — so lex always rejected the compound child, making
level_compound_fix invisible to the outer search.
Fix: for nodes that are genuinely newly divided (not divided in the
parent tree at the same path), inherit the child's operator-set ratio
rather than defaulting to 0.5. Structural mutations (e.g. swap) can
reveal hidden level-N nodes that retain stale pre-writeback ratios —
those are correctly excluded by checking parent_node.divided.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When level_fix alone displaces a required room (e.g. t3) it triggers 5
oracle fails for the missing room — far worse than the starting 2 fails,
so lex always rejects it. level_compound_fix atomically: moves the
constrained room to its required floor AND re-inserts the displaced room
by splitting the sibling of the largest C leaf on that floor. The C
sibling is guaranteed adjacent to C (shared parent split), so the
displaced room keeps its required C-adjacency.
On programme-house this jumps the warmstart from 2 fails (l1 wrong
level + t3 size) to 1 fail (staircase volume), which lex accepts as an
improvement and provides a new base for further mutation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Scores .dom files using fitness.Fitness.score_with_fails(), writes .score
and .fails side-cars in the same format as urb-fitness.pl, and respects the
same skip-if-up-to-date / FORCE_UPDATE caching semantics.
Closes homemaker-py-g0b.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
core_divide: divides a C leaf simultaneously on ALL storeys that share that
path, maintaining staircase consistency as an atomic invariant rather than
requiring multi-step recovery.
core_undivide: reverses core_divide consistently across all floors, merging
a C sub-core back into a single C leaf everywhere.
level_fix: atomically moves a level-constrained room to its required floor
by retyping the largest leaf there and vacating the wrong-floor leaf to C.
Requires `reqs` (SpaceReq dict); disabled (zero probability) without it.
mutate() gains `reqs=None` parameter; driver.search() passes its already-
loaded reqs so level_fix fires during the main memetic loop.
Together these let the optimiser escape the deceptive valley around the
2-fail warmstart: level_fix moves l1 to level 0 (reducing fails 2→1),
then core_divide can split the C core to accommodate the displaced t3.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Cross-storey equivalent of mutate_retype. Directly addresses
level-constraint failures ("l1 on wrong level") by moving a room type
from one floor to another without changing topology or geometry.
Registered in MUTATIONS at default weight (1.0); no drastic geometry
perturbation so it does not need the reduced level_add/delete weight.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously level_add copied the top storey exactly, duplicating all
named programme rooms and immediately triggering space-count failures
for every room on the new floor. The lex outer-search comparison
(-n_fails, score) then always rejected the multi-storey child because
its fail count was far higher than the single-storey parent.
Fix: retype all named-room leaves on the new storey to generic C or O
before admitting the child. The outer search then retypes them
incrementally via the normal retype operator. This allows level_add to
produce designs with the same fail count as the parent (storey_minimum
fail removed, no duplication fails added), making the multi-storey
transition visible to the lex selector.
Result on programme-house cold start (init.dom, 100k evals, 4 workers):
before: 6 fails, single-storey, stuck after 40k evals
after: 4 fails, two-storey, still improving at 100k
Also adds examples/harbor-house/ from urb/examples for future runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Outer search now ranks individuals by (-n_fails, fitness) instead of raw
fitness scalar. This prevents high-score 3-fail designs from displacing
2-fail designs in tournament selection and population replacement — the
root cause of the §4.8 pathology where flag count dominates geometry.
Inner loop is unchanged: it still optimises against the raw 0.5^n fitness
scalar, so the cliff that prevents trading into new failures remains intact
(0/9 regressions in experiments/penalty_reshape.py).
Also removes stale _CHILD_INNER_KW = {"sigmas": (0.05,)}: this was left
over from the CMA-ES era; the NM inner loop default (homemaker-py-d6d)
does not accept a sigmas parameter.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Bakeoff with native fitness shows NM wins at all DOF sizes: +9% at
child_budget=80 for programme-house (6-7 DOF), and decisively at
harbor-house scale (35-40 DOF) where CMA-ES exhausts its convergence
detector after ~3 generations (46 evals) and adds failures on 12/15
runs. NM uses the full budget, is parameter-free, and has zero new
failures across all test cases.
- Add nm_search() to innerloop.py; change optimise() default to "nm"
- Add nm_search to parametrised test cases
- Add bakeoff_native.py and bakeoff_harbor.py experiments with results
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
