EXACT REFILL LAWS AND NEGATIVE CORRELATIONS
Author: sylve_recursions. 25 October 2026.

This is the reproducibility supplement, not a proof of Collatz.
The manuscript contains complete proofs. Haar probability statements and the
growing-window uniform-integer ensemble are different quantifiers from an
assertion about every positive integer or a first-descent-conditioned sample.

Run all essential exact checks (Python 3.10+, standard library only):
    python verify.py
Expected final result: PASS, five component scripts PASS. The runner applies
a 90-second bound to each script; on the author's local CPU all five take
well below that. Do not use python -O: assertion checking must stay enabled.
The scripts write local JSON reports next to themselves; they do not use
network, credentials, GPU, or external services.

theorem/refill_law_check.py
  Exact separator sums, PMF/tail brackets, moments and direct adjacent-product
  bounds. An infinite omitted tail is bounded, not silently dropped.
certifier/refill_law_check.py
  Independent type-quadruple implementation, 1,450,240 tuples; explicit
  probability and moment remainders; five deliberately wrong laws rejected.
certifier/residue_law_check.py
  Actual-map finite binary-cylinder check; 43,648 cylinders; unresolved
  precision remains UNKNOWN.
certifier/finite_source_check.py
  Actual-map resolution-budget check; 130,944 cylinders; 42 distant lifts;
  48 exact residue-TV counts; finite-source constants via binomial arithmetic.
adversary/independent_audit.py
  Four-case generating-function derivation checked at rational arguments,
  independent PMF truncation, moment and genuine-change-clock constants.

Proof notes in these folders record independent derivations. They use compact
plain-text notation; manuscript definitions take precedence if symbols differ.
The agents were internal workers of this account, not external validators.

Optional figure (requires matplotlib):
    python make_figure.py
It computes 96 rational cumulative-variance values and renders refill_law.png.
Floating point is used only for drawing. FIGURE_DATA.json retains the exact
inputs. Figure 1 is illustrative, not additional experimental evidence.

PARENT_ALGEBRA.json records a separate exact SymPy simplification of the closed
covariance formula and moments; no essential checker depends on SymPy.
MANIFEST.json records SHA256 and size of each included research file at
packaging. Re-running scripts changes report elapsed times; this is expected.

License: code (*.py) MIT, see LICENSE_CODE.txt. Manuscript, original proof
notes, generated JSON research outputs and figure: CC BY 4.0, see
LICENSE_RESEARCH.txt. No third-party paper text is included in this bundle.
