Eighteen-policy authored adaptive/fixed/conditional-SoftBoN replay
7 October 2026 Pacific (America/Los_Angeles)

This standalone standard-library source draft reproduces 216 AUTHORED outputs:
two tasks x three bank seeds x two abstract caps x eighteen policies. It is not
real model data, a fitted real-data study, serving/runtime qualification, an
H1 grant or adoption, an efficacy finding, or a complete published SoftBoN
method reproduction. All reviewed production code and original eight test
modules are exact copies. No private operational paths, keys, transcripts,
private task/reference/prompt data, model weights or historical user data are
included. Source commitments are public-safe hash references, not signatures
or self-issued authority.

The method/configuration/fixture rules are declared INLINE in those source
files: original_comparison.py defines the seeded banks, source labels and
source/evaluation families; comparison.py freezes conditional SoftBoN n16,
positive inverse_temperature1, judgeA and separate sampling seeds; allocation.py
and integrated_controls.py freeze all four allocation rules, source-only
marginals and seed formula. No need to obtain an original private directory.
The loader CLI declares separate illustrative late labels; nothing is tuned
against them. Source fitting retains its original eight-candidate source banks
and limited prefix registry, despite sixteen-candidate evaluation capability.

Unpack to a NEW owner-private directory. On a POSIX host, for a fresh path:
  mkdir -m 700 replay
  unzip reproduction.zip -d replay
  chmod 600 replay/*
  cd replay
Do not apply these permissions to existing directories or unrelated files.
The captured loader REQUIRES the input directory0700 and source/descriptor
files0600 owned by the executing account. Python files need no execute bit.
Verify manifest hashes against an independently trusted reference BEFORE
executing code. The manifest does not authenticate itself. Loader own bytes,
initial interpreter, stdlib and bootstrap remain externally trusted; its
checks do not establish immutable executable-FD or native runtime provenance.

Then:
  python3 -I -B load_comparison.py --directory "$PWD" --closure-sha256 ddf99695690a387a6d4f2d53069042e25a6b5ee4097d06c248088cd4cbdec0c1 --output-directory "$PWD/results"
  python3 -B -m unittest test_full test_controls test_integration test_parent_comparison test_softbon test_numeric_regression test_settlement_regression test_dependency test_public_reproduction
  python3 -O -B -m unittest test_full test_controls test_integration test_parent_comparison test_softbon test_numeric_regression test_settlement_regression test_dependency test_public_reproduction
The result directory must be fresh. Each generated JSON artifact is bounded
32MiB, stored owner-private and fsynced. expected-artifacts.json records exact
retained reviewed output hashes and sizes; the additional standalone test
actually launches isolated Python (also optimized when parent is optimized)
and compares all six generated artifacts, including all216 outputs and views.
Exact output bytes were checked on the recorded interpreter/platform; floating
arithmetic/serialization across platforms may change hashes. Such differences
need review, not assertions of empirical disagreement or model qualification.
No universal hard whole-lifecycle/filesystem-finalization bound is claimed.

All17 operative source files are captured and checked BEFORE local imports;
controlled imports use those captured bytes, no ambient local sys.path or
sys.modules. An explicit alias binds integrated_controls.comparison to the
unchanged original thirteen-arm module. Ordinary comparison is the reviewed
fourteen-arm SoftBoN branch. All original thirteen policy outputs, four
allocation outputs and SoftBoN outputs match the reviewed parents. Distinct
parent model-fit and successor comparison-plan commitments remain separate.

The complete18-policy execution registry is retained. Analysis has two frozen
OVERLAPPING views, sizes14 and7; the analysis consumer's sixteen-policy cap is
UNCHANGED. Shared adaptive records are identical, not additional reruns.
All seventeen unique joint contrasts appear once per cap: fifteen exploratory
fixed/allocation/SoftBoN controls and two original matched-adaptive contrasts
(joint-additive primary, joint-placebo diagnostic). These are reused paired
task/seed draws, not independent studies, multiplicity-controlled inference,
calibrated correctness, simultaneous confidence guarantees or powered H1.
Unknown outcomes and expenses remain unknown; true known expense is retained
before overquote refusal; no partial-pool SoftBoN fallback or cost imputation.
The fixed pool's scalar reward is not assumed a correctness probability.
The conditional published Algorithm2 kernel is exp(lambda*r)/sum(exp(lambda*r));
its exact-Fraction stabilization does not reproduce iid generation, tuning,
logit access, repeated judging or a full published-method pipeline.

A known-law/synthetic mechanical replay cannot establish actual adaptive value,
learnability, source/target transfer, fresh-bank generalization or production
performance. Real source rights, counter/template/serving parity, authenticated
cost and grading/native/clock/collector authority remain external and absent.
No original H1 protocol or scientific freeze is changed. No model/native/VM
execution, site publication or external spend is part of this draft.
