# Independent reproduction notes

I read the complete analysis, registered code, deviations, claims, paper, and result tables before accepting the harness verdicts. The registered and executed `colex.py` files are byte-identical. The isolated, network-disabled run reproduced all 15 claimed fields exactly. Public CLICS4 v1.0 wordlist archives were fetched by the harness and checked by declared byte count and SHA-256.

A separate standard-library CSV/ZIP reader independently rebuilt the four HEAVY-target family coverage and colexification counts and all 15 physical-to-affect family coverage and colexification counts. Every count agreed. This independent program reads only the required concepts and does not import or call the submitted code. Its results are in `independent-check.json`.

As a sensitivity calculation, I enumerated all 455 distinct three-of-fifteen labelings with exact rational arithmetic on the analysis's four-decimal rates. The one-sided tail is 103/455, approximately 0.226374, versus the declared seeded Monte Carlo value 0.2232. Both are nonsignificant at 0.05. This does not replace the declared Monte Carlo result and is not a discrepancy in its reproduction. The plan's phrase about all labelings and 10,000 permutations is ambiguous; the submitted registered implementation samples labelings and the paper reports that implementation.

The attestation establishes the declared computations on this frozen database, rather than a universal absence of lexical or phrasal expressions. A zero count in the selected gloss coverage does not establish impossibility in all languages, and the labeling test requires a concept-exchangeability assumption for inferential interpretation. Lexical identity does not by itself test a mechanism of depression. Those limits do not alter the numerical reproduction verdict.

The code and public aggregate linguistic data contain no operational biological, chemical, cyberattack, or other harmful procedure. Hazard screen: none.