Replication Package: The Semantic Share φ — A Registered Experiment with Large Language Model Producers and Reviewers
Abstract
Replication package for the OSF-preregistered experiment (https://osf.io/uj9rv, DOI: 10.17605/OSF.IO/UJ9RV, registered 2026-10-06 before any data collection) accompanying the working paper "Observable but Incomprehensible: Delegation under Verification Capacity Constraints" (Wu, 2026). Contains the frozen pre-analysis plan (v1.4) and the frozen G0 analysis-code deposit (both byte-identical to the OSF registration attachments), all collection code with deterministic plan builders (md5-anchored), the executed plans, the frozen artifact bank (1,260 items), all four confirmatory phases' raw per-unit records — 22,784 units: 13,860 Layer-S + 3,200 H4 + 900 H5-generation + 4,824 H5-review, collected 2026-10-08 through 2026-10-13 across five model families (deepseek-v4-pro, kimi-k3, qwen3.8-max-0902, glm-5.3, doubao-seed-2-1-pro-260915) — plus the twelve pilot run directories (excluded from confirmatory analysis by design), supervisor and budget-watchdog logs, the full deviation log (D1–D35), the cross-phase cost ledger, the data-integrity report, and the missing-unit disclosure table (83 terminal parse failures; 96 H5 channel-gate exclusions, all DeepSeek empty-output generations). The H5 producers' private-channel planted-defect reports (the scoring ground truth for the H5 detection analyses, 804 kept items) are included at data/h5_private_reports.jsonl. Audit truth is each phase's raw.jsonl. See README.md for the data availability statement, license (data CC-BY 4.0; code Modified BSD), plan-regeneration md5 anchors, and reproduction instructions. Version 1 (2026-10-13): data frozen at collection completion, before confirmatory analysis; executed analysis outputs will be deposited in a subsequent version.
