FORMAL VERIFICATION SCOPE — arXiv preprint
Learning Orthant Comparability: Proper–Improper Separation and
Compression without Side Information

The numbering below refers to the main manuscript, not the supplement.
The Lean sources are an additional check of specified mathematical results;
they are not a formalization of the whole paper or a novelty certificate.

1. Covered mathematical claims

- Sign/face representation, endpoint symmetry, cardinality formulas, parity
  constraints, and the affine positive version-space representation.
- Theorem 4.1: existence of a positive sub-sample of size at most d-1 with
  identical version space; tightness of this preservation requirement on a
  path construction. This is not a lower bound for all compression schemes.
- Theorem 4.2(a): closure is one-sided sound and fits every realizable mixed
  sample. A negative closure prediction is not a certificate of true
  incomparability on an unseen pair.
- Theorem 4.2(c): at most d-1 online mistakes, all false negatives, and the
  matching worst-case deterministic lower bound against full-history learners.
- Theorem 4.3: a fixed map from positive samples to positive sub-samples of
  size at most d-1, with closure reconstruction fitting every realizable mixed
  sample, using no auxiliary information. The map is noncomputable and uses
  Classical.choose. Existence and correctness, not an executable efficient
  selection procedure or its implementation correspondence, are checked.

2. Additional but restricted PAC coverage

Theorem 4.4's binomial-tail comparison and exponential tail bound are checked
for finitely supported distributions, represented by a finite probability
vector and an arbitrary map into the pair domain. The confidence corollary
is checked where its displayed threshold is at most one. No general-measure
reduction, unrestricted distribution-free PAC theorem, running-time bound,
or efficient PAC-learnability claim is certified by this Lean development.

3. Modelling correspondence

Lean uses ordered pairs (also allowing equal endpoints) and finite sets of
positive constraints. The manuscript uses unordered pairs of distinct points.
Endpoint symmetry permits the mathematical restriction using a fixed choice
of representatives, but a quotient-type interface is not formalised. Those
representative choices must not encode hidden auxiliary bits. The lower-bound
constructions assume two ordered levels in each coordinate. Upper bounds
hold under the weaker assumptions stated in the individual Lean theorems.

4. Explicit exclusions

Theorem 4.2(b), implementation complexity, an efficient implementation of the
compression map, the component-count dimension formula, proper-learning
hardness and proper/improper complexity separation, later PAC-Bayes and
random-threshold results, and all LLM/agent applications are not formalised.
The finite Python checks are separate certificate checks, not general proofs
or empirical validation of an AI application.

5. Build and axiom check

Lean: leanprover/lean4:v4.35.0.
Mathlib commit: 56a005855ea17e0b0958f73ec801ae140b0a1679 (v4.35.0).
All nine dependency revisions are pinned in lake-manifest.json.

A source-only copy was rebuilt on 24 October 2026, using existing dependency
artifacts matching the pinned revisions, but no old LOC build artifacts.
All ten LOC modules and the LOC entry point compiled. The original repaired
checkout also compiled and passed the same audit. Audit.lean transitively
checks 287 declarations in the LOC namespace and permits only:
  propext, Classical.choice, Quot.sound.
There are no sorry/admit placeholders or user-added axioms in the LOC proof
sources. Non-fatal linter and deprecated-syntax warnings remain. This check
trusts Lean's kernel and the pinned dependencies; it is not an independent
reconstruction of the compiler or a clean-network-download test.

The proof sources and Audit.lean are unchanged from the audited development.
The finite-record pullback in Definition 2.1 is an expository extension:
Lean formalizes the order on feature vectors, not a separate type of
labelled records with a possibly noninjective feature map.

Reproduce after extracting the source archive and installing elan:
  cd LOC
  lake exe cache get
  lake build
  lake env lean Audit.lean

The first build needs network access to obtain the pinned toolchain and
dependencies. No AT project, author-specific absolute paths, symbolic links,
or bundled .lake directory are required by the distributed source package.
