# The proposed identification procedure is validated on synthetic graphs of varying size and on seven real-world datasets spanning biology, agriculture, and social science (Section 5).

**Verdict: FALSIFIED AS LITERALLY REGISTERED.**

This compound empirical-validation claim fails on its directly executable synthetic half. The package runs the printed SEM, selection/reproduction loop, sample size, graph sizes, PC alpha, GES penalty, and 50 seeds per cell. It does not substitute oracle CPDAG precision for finite PC/GES.

| Native synthetic result | PC | GES |
| --- | ---: | ---: |
| grid cells where oriented-only mean precision beats all-adjacency precision | 3/15 | 14/15 |
| same count at `d=20` | 0/5 | 4/5 |
| total algorithm/data runs | 750 | 750 |

The paper says its interpretation is consistently more precise for both algorithms at `d=20`. The native rerun yields the opposite PC ordering at all five reported generations. This remains true across the separately executed selection-strength and graph-density sensitivity sweep in `outputs/variants.json`.

The real-data audit is transparent about its different evidence levels:

| Real-data check | Result |
| --- | ---: |
| paper datasets named and arithmetically audited | 7 |
| printed percentage arithmetic mismatches | 0 |
| six pseudo-ground-truth datasets with oriented > unoriented precision | 5/6 |
| pooled oriented precision over those six | 50/138 = 0.3623 |
| pooled unoriented precision over those six | 15/37 = 0.4054 |
| raw datasets independently rerun | PanTHERIA only |
| PanTHERIA shape | 626 x 8, matching the paper |
| PanTHERIA log/alpha-0.05 oriented precision | 0.40 |
| PanTHERIA log/alpha-0.05 unoriented precision | 1.00 |

The printed seven-dataset arithmetic is internally consistent, but it cannot rescue a compound validation claim when the native synthetic headline fails and the only raw real-data rerun also reverses the precision ordering.

## Destructive control

Changing one printed DGRP numerator makes the arithmetic consistency gate fail. Flipping one byte of the actual PanTHERIA input changes its pinned SHA-256 and is detected before analysis.

## Evidence

Native synthetic pipeline: `outputs/synthetic_native_pipeline.json`. Raw PanTHERIA and seven-table audit: `outputs/realdata_native_and_source_audit.json`. Executed controls: `outputs/destructive_controls.json`.

## Limitations

Only PanTHERIA is independently rerun from raw data. The other six datasets are assessed from the paper's printed integers, and six of the seven evaluations use pseudo ground truth that the paper itself calls an unreliable surrogate. The falsification therefore targets the compound empirical-validation claim; it does not assert that the seven analyses were never run.