# A08 · Accountable intelligence methods package

**Article:** Accountable Intelligence at the Point of Decision: A Europe PMC Corpus Audit and Randomized Protocol for Evaluating Paper Agents in School-District Evidence Work  
**Author:** Dr. Freedom Cheteni  
**Version:**1.0 · Prepared15 October 2026(Pacific)  
**Publisher:** The Research University · Journal for School Superintendents

This package reports a completed36-record Europe PMC core/JATS feasibility audit and completed artificial-data simulations. The human trial is prospective:160 planned adults, zero recruited participants. No agent accuracy, human performance effect, student benefit, ethics approval, registration, acceptance, DOI or review is invented.

## Files

- `article.md`: original manuscript; source articles remain distinct.
- `study-specification.json`: planning design, actual executed analyses and prerequisites.
- `corpus-audit.json`: exact query,98-hit response boundary,36 audited sources, author/identifier/date/rights/hash records and exclusions.
- `source-audit.json`:16 references with granular full-text/documentation/metadata-only verification; supplemental JATS records.
- `design-analysis.json`: nine artificial-data scenarios,10,000 replicates each, seed/software/MCSE/assumptions.
- `instruments.md`: draft six-dimension0–24 rubric, case architecture, participant brief and blinded reference-standard procedure.
- `analysis-specification.md`: estimand, allocation, Welch inference, missing-data plan, secondary models and reporting boundaries.
- `research_tools.py`: executable audit, simulation and complete-cohort estimator.
- `test_research_tools.py`: synthetic rights, author, version, heteroskedastic-estimator and simulation tests.
- `requirements.txt`: exact NumPy/SciPy versions used.

## Run offline tests and simulations

Use Python 3.12(or a compatible version) and an isolated environment with the pinned requirements. Commands assume this directory:

```sh
python -m pip install -r requirements.txt
python test_research_tools.py
python research_tools.py simulate --output repeated-design.json
```

The simulation uses seed20261015,10,000 replicates per scenario. It does not generate observations that may be represented as human trial data. Same seed/version yields the same numerical outputs.

## Rerun the corpus audit responsibly

```sh
python research_tools.py audit --cache /absolute/nonpublic/A08-sources --output repeated-corpus.json --download
```

Use the documented Europe PMC API, not scraping of article webpages. Keep the cache outside public hosting. `search-core.json` and36 `PMCID.xml`files freeze the source snapshot; existing cache files are read without silently refreshing. The package intentionally distributes no original JATS/PDF/HTML full texts or participant records. A fresh query’s relevance ranking and article bytes can change; compare received hashes and the exact source IDs with the published audit, then date and explain any new snapshot. Do not overwrite the historical result or assume a changed response is an exact reproduction. The hardcoded date window and declared36-record boundary describe this historical analysis; a future review requires an explicitly dated new audit.

Retrieval hashes detect changed bytes, not truth, complete licensing or publisher endorsement. A clear article license does not grant rights to every third-party figure/dataset. NC/ND, ambiguous versioned permission and identifier conflicts are excluded from the source-construction pool. The SPIRIT supplemental reference has conflicting CC BY URL/no-commercial-reuse wording and is citation-only; it is not a benchmark ingestion candidate. An observed F1000core DOIversion2/JATSversion3disagreement is preserved, not labeled misconduct. No observed status signal means no signal in the screened fields at retrieval, not an integrity certificate.

## Complete-cohort estimator

```sh
python research_tools.py estimate-complete --input synthetic-or-authorized-input.json --output estimate.json
```

Input schema:

```json
{
  "dataKind": "synthetic-fixture",
  "rows": [
    {"participant_id":"synthetic-person-001","stratum":"leader-lower","arm":0,"mean_score":12.5,"caseCount":12}
  ]
}
```

The fragment illustrates schema only; a valid test needs four strata with at least two per arm in each. `study-data`requires160 unique adults,20 per arm in each of four strata and12 scored cases each. A provenance label is not ethics certification. The tool rejects duplicate IDs, invalid arms, nonfinite/out-of-range scores, empty cells, zero variance and silently altered actual study sizes. It implements only the complete-cohort estimator; missingness requires the separately specified imputation analysis, not listwise deletion hidden in a success message. It transmits no participant data. Store any actual records under consent, approved restricted access and retention rules.

## Required before a primary human study

Secure funding/partners, independent ethics determination and consent, accessible interface and readable active control, independent case-panel and validated scoring procedures, frozen source/code/model configuration, a preregistered analysis and allocation process. The new rubric is not validated by this methods package. No recruited sample, appointments, permissions or trial registry entry is asserted. Model-assisted source inspection cannot substitute for independent human peer review.

The conceptual companion is A03,`accountable-intelligence-district`. Its earlier review is not inherited by A08. Human authorship, independent handling of an Editor-in-Chief submission, original source attribution and genuine publication dates remain intact.

### Edition 1.1 attribution correction — 16 October 2026

The contributor parser now excludes explicit Academic Editor and Editor roles when a source omits contributor-type attributes, while preserving CRediT writing and editing contributions. The frozen 36-record source audit and affected references were regenerated from the unchanged source bytes. Source eligibility counts, study design and numerical simulation results are unchanged.