Independent reproduction report

All four assigned computational claims were checked, with eleven result comparisons.
Verdicts: {"C1": "reproduced", "C2": "reproduced", "C3": "reproduced", "C4": "reproduced"}. Both freshly executed implementations match all declared result values within their specified tolerances. The complete result objects are also equal.

The supplied code was read before execution and run in a Python 3.12.15 container with an initially empty results directory. The independent verifier was written from the described sampling, seed formula, algorithms, and summary statistics, without importing the supplied implementation or reading declared results. It uses an in-place adjacent-pair reduction rather than the supplied slice-based implementation.

The independent numerical reference sums the integer numerators of the generated dyadic numbers exactly, rounds the integer total once to binary64 with ties to even, then rescales by an exact power of two. All generated inputs have denominator 2**53, and these totals are far from overflow and underflow. Thus scaling after rounding introduces no additional rounding. The same construction is used for the sum of magnitudes. math.fsum agrees with both exact references in all 400 cases. The independently reconstructed algorithm outputs, all summary fields, and growth exponents match the supplied-code rerun and declared results.

Both runs used no network, a read-only root filesystem, no Linux capabilities, no privilege escalation, an unprivileged numeric user, and memory, process, CPU, and time limits. Only computation code and a dedicated empty results directory were mounted. The supplied run took 31.415 seconds and the independent run 46.566 seconds, together within the declared two CPU minutes of wall-clock execution. These are observed durations on this verifier's machine, not benchmarks.

Integrity findings: the assigned integrity object reports no orphan numbers, missing sections, missing files, duplicate-row or Benford issues, or skipped tables. I inspected all nine supplied files and found no hidden Unicode control, formatting, surrogate, or private-use characters, no raw HTML, no personal information, and no instructions aimed at a verifier. The inputs are generated from fixed seeds; no field or human observations are involved.

Statistical limitation: for the 50 condition numbers, the supplied code reports sorted(values)[25], the upper central order statistic. For an even sample, the conventional median averages the two central observations. This convention reproduces exactly, but readers should not interpret the reported condition-number statistic as that conventional median. exact-reference-audit.json separately records conventional medians. In addition, this finite seeded experiment does not establish correctness of compensated summation for arbitrary inputs, nor universal portability across every floating-point platform. These cautions do not change the computational reproduction verdicts.

Hazard screen: none. This is a numerical summation experiment with harmless standard-library code and synthetic inputs; it provides no meaningful capability for mass harm or unauthorized attacks.