EAIPAM methodology

Transparent evidence

The language model may assist with structured evidence extraction, but programmed rules determine the final category scores and risk bands.

EAIPAM scoring profile

Reference Integrity30%
Citation-to-Claim Alignment30%
Scientific Accuracy25%
Evidence Depth15%

Reproducibility protocol

1

Fingerprint

Calculate a SHA-256 hash for the exact uploaded file.

2

Version

Lock parser, rubric, model, prompt, dataset and reference snapshot.

3

Score

Apply fixed numerical rules to stored structured findings.

4

Preserve

Store the evidence and approved report as an immutable record.

Calibration and training governance

Administrators can upload verified human-drafted, AI-drafted, mixed or indeterminate manuscripts. Every record must include provenance, permission status, subject area and label confidence.

No silent learning.

A newly uploaded example cannot change live assessments. It enters an experimental dataset, undergoes validation and becomes active only through a newly published EAIPAM training profile.

Responsible interpretation

EAIPAM reports evidence and editorial risk. It does not claim that stylistic patterns alone prove authorship by a particular model or tool. Reference integrity, scientific accuracy and citation-to-claim alignment remain the dominant assessment components.