scholar-evaluation – Security skill for AI agents

Scholar Evaluation

Purpose

Provide developmental, evidence-traceable feedback on a scholarly work:

paper, draft, protocol, literature synthesis, or research idea. Use

qualitative judgment first. Optional scores only describe how submitted

evidence maps to a predeclared bounded rubric.

This skill also audits whether a low-stakes assessment process documents its

construct, provenance, rater quality, uncertainty, traceability, sensitivity,

fairness, accessibility, privacy, and human governance.

Hard safety boundary

Never use this skill to automate, recommend, materially influence, or score:

• hiring, promotion, or tenure;
• admissions;
• grants or other funding;
• prizes, honors, or awards;
• discipline, dismissal, or sanctions; or
• any other high-impact personnel decision.

Never rank people. Never reduce a person to a composite score. Never infer

ability, character, integrity, protected traits, future performance, or worth.

A nominal human-in-the-loop does not remove this boundary.

If asked for a prohibited use, stop. Offer developmental comments on a

scholarly work or a process-only audit that does not process applications,

compare people, recommend an outcome, or advise a decision.

Do not issue publication-readiness, accept/reject, or “top-tier” judgments.

Read references/responsible_assessment.md before any organizational use.

ScholarEval status

The referenced ScholarEval project is an **experimental

literature-grounded research-idea evaluation framework**, not validated

psychometrics.

The verified primary record is Moussa et al., *ScholarEval: Research Idea

Evaluation Grounded in Literature*, arXiv:2510.16234v2, revised 2026-02-28.

It reports a retrieval-augmented soundness/contribution framework, a

117-idea four-discipline dataset, coverage experiments, and a user study.

Do not generalize those results to person assessment, consequential decisions,

all disciplines, or this skill's rubric. No peer-reviewed publication status

was verified during the dated review. See references/source_ledger.md.

Metric and prestige policy

Do not score or infer quality from:

• Journal Impact Factor or other journal measures;
• h-index, publication counts, or citation counts;
• altmetrics or attention;
• journal, conference, venue, institution, employer, or geographic prestige;
• author affiliation, reputation, network, or career path.

The rubric validator rejects common proxy-measure criteria.

If a qualified reviewer mentions an indicator descriptively outside the

scoring tools, record its exact purpose, source, coverage, field and time

effects, uncertainty, missingness, biases, gaming risk, and why it does not

directly measure quality. Never hide indicators inside an opaque composite.

Data boundary

Bundled scripts accept only strict local JSON/CSV containing pseudonymous IDs,

bounded ratings, statuses, uncertainty, and local references.

Do not put raw private applications, CVs, letters, reviewer identities,

contact details, protected attributes, or source-document text in inputs,

outputs, logs, examples, or prompts. Keep source content in the authorized

records system and use opaque local references.

Allowed classifications are:

• synthetic
• public_scholarly_work
• deidentified_low_stakes

No script searches the web, loads environment files, reads credentials, calls a

model, executes supplied text, deserializes executable objects, or launches a

process.

Use Bash only to invoke the documented local python3 commands.

Workflow

1. Confirm allowed use and authorization

Record:

• developmental purpose;
• unit of assessment: scholarly_work;
• work type, stage, discipline, language, and audience;
• authorized source location and data classification;
• accountable committee owner;
• conflicts and recusals;
• accessibility and accommodation process;
• appeal or correction route; and
• data purpose, access, retention, and deletion.

Stop on a prohibited decision context or unnecessary private data.

2. Define the construct before criteria

State:

• what quality or support is being examined;
• excluded constructs;
• intended interpretation;
• contexts where the interpretation does not travel;
• evidence requirements; and
• known limitations.

Start with values and disciplinary context, not available metrics.

3. Adapt and validate the rubric

Begin with assets/rubric_template.json, then obtain qualified disciplinary,

assessment-methods, stakeholder, accessibility, privacy, and fairness review.

The template deliberately records content validity as not_established.

Do not change that status without documented evidence for the exact intended

use.

Validate structure:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/validate_rubric.py \
  --rubric assets/rubric_template.json

Read references/evaluation_framework.md for construct, anchor, validity, and

rater guidance.

4. Build traceable evidence records

Reviewers may read an authorized work outside the scripts. Record only stable

local locators and claim references in

assets/evidence_manifest_template.json.

For every criterion, distinguish:

• observed evidence from interpretation;
• supporting from contrary evidence;
• available from unavailable evidence;
• missing from not_applicable; and
• uncertainty from absence.

Failure to find prior work does not prove novelty. Freeze the exact work revision and evidence-access date before independent rating so raters assess the same material. For public papers, check publisher correction/retraction notices and [Crossmark](https://www.crossref.org/services/crossmark/) where available; record unresolved status rather than treating absence of a notice as verification. If the work changes materially, issue a new evaluation linked to the prior revision.

5. Rate independently

Use assets/evaluation_template.json. Each criterion must be:

• rated with an anchor score, bounded uncertainty, evidence IDs, and a local

rationale reference;

• missing with null score/uncertainty and a rationale reference; or
• not_applicable with null score/uncertainty and a rationale reference.

A rated zero requires inspected evidence demonstrating lack of support; unavailable

evidence is missing. Do not encode missing or not-applicable as zero. Raters should train, calibrate,

disclose conflicts, rate independently, and document disagreement.

6. Run local quality checks

Bounded scoring, without labels or recommendation:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/calculate_scores.py \
  --rubric assets/rubric_template.json \
  --evaluation assets/evaluation_template.json

Evidence traceability:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/check_traceability.py \
  --rubric assets/rubric_template.json \
  --evaluation assets/evaluation_template.json \
  --evidence assets/evidence_manifest_template.json

Inter-rater agreement (one evaluation_id identifies one frozen work and round;

raters share that ID within the round):

PYTHONDONTWRITEBYTECODE=1 python3 scripts/summarize_agreement.py \
  --rubric assets/rubric_template.json \
  --ratings assets/ratings_template.csv

Weight sensitivity requires two or more distinct scholarly-work evaluation

files:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/weight_sensitivity.py \
  --rubric assets/rubric_template.json \
  --evaluation /tmp/work-a-evaluation.json \
  --evaluation /tmp/work-b-evaluation.json

Process controls:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/check_process.py \
  --process assets/process_checklist_template.json

The checklist template is intentionally unconfirmed and fails closed.

Instructions and exact schemas are in references/local_tooling.md.

7. Synthesize qualitative findings

Lead with criterion-level evidence, not the composite. For each criterion:

1. cite evidence references;

2. state rated, missing, or not_applicable;

3. explain the anchor interpretation;

4. report score and uncertainty only if rated;

5. note disagreements and context;

6. identify strengths and limitations; and

7. offer non-prescriptive improvement options.

Generate an empty-reference scaffold if useful:

PYTHONDONTWRITEBYTECODE=1 python3 scripts/generate_report_scaffold.py \
  --rubric assets/rubric_template.json \
  --evaluation assets/evaluation_template.json \
  --output /tmp/developmental-report-scaffold.json

The scaffold does not read source documents or draft findings.

8. Human review and release

Before releasing an organizational report, a qualified accountable human

committee must verify:

• construct and rubric provenance;
• content-validity evidence and limits;
• rater training, agreement, inter-rater reliability evidence, and drift;
• evidence traceability and source access;
• missingness, not-applicable rationales, and uncertainty;
• weight sensitivity and order instability;
• disciplinary and subgroup bias review;
• conflicts and r

… (full content is in the download package)