From Observable Answers to Executable Queries for CM/LM Programs
Brian Theory | Technical report v1.1 | 8 November 2026

START HERE
REPORT.txt is the readable text edition (equations retain TeX notation).
source/main.tex is the complete editable report with inline bibliography.
This package is self-contained: no original workspace files, network, account,
third-party Python dependencies or paid services are needed for its checks.
Python 3.10 or later is required. Run from the extracted package directory:

  python -B code/demo.py
  python -B code/run_checks.py

The demo checks three saved policies, including a correct suboptimal one.
Its fixture prices are explicit: address queries cost 1 each and installation
prices are 2,3,5; the unlock fixture has execution prices 0,2. These exercise
the checker and differ from the report's illustrative unit-price models.
The full command reproduces the supplied expected results in a temporary
directory, checks the matrix identities, and prints PASS. It leaves bundled
code/evidence unchanged. See REPRODUCTION.json for the current public-package run. Finite execution counts are evidence for the supplied
examples, not substitutes for the general proofs.

CONTENTS AND CONNECTION TO THE REPORT
code/static_selection.py: direct grouping, exact conflict covering and lazy
  constraints. Includes the historical four-state no-coarsest example as an
  additional code exercise; that secondary result is omitted from this note.
code/adaptive_acquisition.py: exact worst/unnormalized expected recurrence,
  independent small tree-vector enumeration, addressing and availability cases.
code/policy_checker.py: standalone policy validity/cost verifier, no optimizer.
code/test_certificates.py: five valid fixtures and nineteen deliberate failures.
code/leaf_audit.py: full partial-assignment audit for b=0..3 and sampled adaptive
  transcripts, including repeats and promise-versus-prior counterexamples.
code/run_checks.py: disposable-copy orchestration plus 32 matrix valuations.
evidence/: expected outputs and complete model/policy fixtures.
REVIEW_AND_SCOPE.txt: public review summary and limits.
PUBLICATION_PROVENANCE.json: source and PDF identity; public package scope.
REPRODUCTION.json: the current public-package verification result.
MANIFEST.json: hashes of package files, excluding itself.

QUICK API EXAMPLE
From code/ or after adding it to Python's module path:
  from policy_checker import verify
  result = verify(model, certificate)
Use evidence/certificate_fixtures.json for fully populated examples. The model
contains alpha (string labels), target (integer labels), sensors (binary tables),
execution/installation/prior (exact rational strings), and requires (lists of
[sensor_index, observed_value] prerequisites). Prior weights sum to one.
The certificate contains model_id, roots, nodes and claims. The digest is
computed by model_id(model); it binds canonical JSON, not physical truth.
Roots cover every attained alpha. Nodes are tests with sensor/branches or
leaves with answer. Claims are worst, expected and installed rational strings.
Every stored node/edge must be used by a feasible state. Zero-prior states
remain feasible. Repeats/cycles on feasible paths are rejected. Shared graph
nodes are checked per actual history; global graph acyclicity is not required.
Acceptance certifies this policy's legal answers/costs, never its optimality.
The trusted explicit-table interface is not a hardened hostile-JSON parser.

REPRODUCTION SCOPE
Expected static results retain 9216 exhaustive instances, 400 sampled cases
and 4096 subset-reduction equivalences. Adaptive results retain 300 main
enumerator comparisons. The leaf audit exhausts 177906 partial assignments,
53768 constant-output subcubes, and checks 48 sampled adaptive trees.
The five/19 fixture counts are unchanged; new packaging runs are not new
research cases. Missing historical reviewer harnesses/counts are not claimed
as bundled reproducible evidence. Code is intended as a small exact reference,
not a scalable production optimizer or Tensor Logic runtime.

API SCOPE
adaptive_acquisition.solve operates on one initial observation fiber (or
constant alpha), with a normalized prior. To obtain a conditional expected
optimum for a positive-mass fiber, normalize within that fiber, then multiply
by its original mass before summing across fibers. A zero-mass feasible fiber
still needs a legal exact policy: solve with any normalized auxiliary prior
to check feasibility, and contribute zero to global expectation. Do not pass
all-zero weights to solve. The report gives the
unnormalized mathematical recurrence. The code is not a multi-alpha optimizer
frontend. policy_checker.verify does support multiple alpha roots.

PUBLIC RELEASE STATUS AND TERMS
This public reproduction package accompanies the unchanged technical report v1.1.
It includes the exact reviewed source and PDF, unchanged reference code and
expected evidence, and a public summary of review scope. Private review
correspondence, machine-specific paths and local build artifacts are omitted.
AI-assisted internal review is not human peer review or formal verification.
The paper is an expository synthesis, not a new CM-exclusive algorithm.
No exhaustive novelty or priority clearance has been established.
Source and code are supplied for inspection and reproducibility. No additional
reuse or open-source license is granted by this release. No new license,
affiliation, DOI, external endorsement or venue acceptance is asserted.

The reviewed PDF has seven pages. The source uses standard LaTeX packages and
can be compiled with pdflatex twice. The original paper source and PDF bytes
are unchanged by this public packaging. The manuscript's historical source
access dates are not new verification dates.
