Overview
At a glance
- Phase 4C audit manuscript
- Candidate/finalist selection tables, coded overlay
- Subset20 audit labels, per-field metrics
- SHA256-verified reproducibility bundles
- Poisoned-evidence robustness bundle
This deposit contains the Phase 4C measurement-layer audit manuscript plus two reproducibility bundles.
The primary reproducibility bundle includes the tabular inputs and derived audit artifacts referenced in the paper, including the candidate and finalist selection tables, the coded overlay table, subset20 audit labels, and per-field audit metric CSVs, with a SHA256 sidecar file for integrity verification.
A supplementary robustness bundle contains non-blind “poisoned evidence” excerpt-sufficiency audit outputs and the prompts used to generate them; it supports the robustness check described in the manuscript and does not constitute an independent validation of the overlay labels. Scripts are not included; the release supports transparent inspection and independent recalculation of the reported counts and audit confusion matrices from the included CSV artifacts.
What it claims / What it does not claim
What it claims
Traceable labels, for one subset. The packaged overlay labels for the audited subset are mechanically traceable to explicit textual evidence under the stated codebook rules — with FP=0 and FN=0 across all evaluable fields in the corrected reproducibility bundle.
An auditable procedure. Ten operationally-defined overlay fields, applied across a 63-study finalist set, with a stratified subset-20 verification using per-field evaluability rules and confusion-matrix reporting.
What it does not claim
A narrow methodological result. By the paper’s own statement, broader reliability claims and theory-validation claims remain out of scope; the result speaks to traceability of the packaged labels for this subset, not to the framework’s truth.
Bounded by evaluability. Some markers were evaluable in as few as three of the twenty audited studies, so the clean confusion-matrix result holds only where fields were evaluable — not as a uniform guarantee across every marker.
Figures
