# Independent reproduction notes

The submitted computation ran successfully in the reference harness with an empty results directory, no network, resource limits, and only bundle inputs. All 22 claim-linked result fields matched exactly. Both complete output JSON files matched the submitted files byte for byte.

A separately written standard-library implementation read the three SHA256-verified public CLDF tables and inverted word forms to concept sets, rather than importing any submitted code. It independently matched numerator and denominator counts for all 21 declared concept pairs and every available percentile. Percentile ordering used integer cross-products, avoiding floating-point ranking ambiguity. Its script and outputs are included. This second check also ran offline in a read-only, capability-restricted container.

The node reported no orphan numbers, missing sections or files, uncited references, unlisted citations, duplicate rows, Benford findings or skipped tables. The harness found no hidden-text findings. Those automated flags were reviewed, as were the paper, claims, code, registered plans, provenance, methods, results and external-data descriptions. Duplicate plan/code copies agree with their operative versions.

Hazard verdict: none. The work is a lexical co-occurrence analysis of openly licensed wordlist data, with aggregate counts. It contains no identifiable participant records, credentials or actionable mass-harm capability. Public datasets are referenced by URL and digest with attribution.

These are computational reproduction verdicts. They do not establish the validity of hardness as a tension proxy, independence of language families, adequacy of sparse concept coverage, universality of the linguistic statements, or the absence of phrasal metaphors. The paper acknowledges several of these scope limits; methods and domain review remain separate.