ASU Resume Authenticity Audit Skill

Treat a resume as a set of independently verifiable claims. The goal is not to judge whether a person is "good" or "bad", but to determine what is supported by existing evidence, what is directly refuted, and what content remains unverifiable.

Mandatory Deliverables

Unless the user explicitly narrows the scope, generate the following:

```
report-data.json
```
: Standardized claims, sources, conclusions, packaging patterns, timelines, and limitations.
Self-contained
```
report.html
```
: No external styles, scripts, fonts, or images are referenced.
```
report.pdf
```
consistent with the same data model.
A short conversation summary: Prioritize stating the strongest supporting evidence, direct contradictions, and unresolved gaps.

After completing the JSON, use

scripts/render_report.py

to generate HTML/PDF; validate with

scripts/validate_report.py

before delivery.

Evidence and Harm Boundaries

Resume audits involve identifiable individuals, and incorrect conclusions may cause real reputational harm, so always follow these rules:

Only use public information directly relevant to the task. Do not collect addresses, phone numbers, family members, account credentials, private messages, or irrelevant personal information.
Screenshots, accusation articles, anonymous comments, and the candidate's own homepage only prove "someone said this", and cannot automatically become independent evidence.
Use precise statements such as "Public evidence only supports contributor status, not core author status", and avoid personality labels like "liar", "pest", or "fraudster".
Mark as
```
contradicted
```
only when reliable evidence directly conflicts with the claim; otherwise, use
```
unsupported
```
,
```
partially_corroborated
```
, or
```
unverifiable
```
.
Initiating issues, contributing to front-end/documentation, building relationships with maintainers, or obtaining formal community status as a result is not improper behavior. What needs to be verified is whether the resume accurately describes formal roles, technical depth, and personal contributions.
Do not infer integrity based on school rankings or the informal label "double non-first-class". "Double non-first-class" is an informal label and not evidence of fraud. Verify the institution, program, student status, and degree-granting body.
Inferences about gender, appearance, personality, taxes, or motives are usually irrelevant to resume authenticity; exclude them unless authoritative evidence proves a direct connection to a resume claim.
Important evidence favorable to the party must be presented equally with unfavorable evidence.

Read

references/evidence-methodology.md

before assigning conclusion status. Read the corresponding special references before verifying open-source or educational claims.

Workflow

1. Define Scope and Save Source Collection

List all materials provided by the user:

Resume PDF or screenshots;
Articles and social media posts;
Links to code repositories and personal homepages;
Web archives;
Employment, school, award, or paid community background provided by the user.

Read links and documents in full when possible, and check images at their original resolution. If a page is blocked by security policies or login restrictions, state the limitation; use the text provided by the user or other public sources, but do not bypass access controls.

Immediately assign a number to each source (

S-001

S-002

, ...) and mark the source type:

```
official_record
```
: Official records;
```
repository_record
```
: Code repository records;
```
candidate_self_statement
```
: Candidate's self-statement;
```
user_provided_document
```
: User-provided materials;
```
media_or_commentary
```
: Media or commentary articles;
```
anonymous_or_unverified
```
: Anonymous or unverified sources.

2. Break the Resume into Atomic Claims

Do not treat an entire experience as a single claim. Separate roles, scope, results, and causal relationships for verification.

Example:

"As core author, led the project from version 1.0 to 2.0, improving performance by 25%."

Should be split into:

The candidate is formally or de facto a core author.
The candidate led the migration from 1.0 to 2.0.
Performance metrics did improve by 25%.
The improvement was caused by or substantially contributed to by the candidate's work.

Record for each claim:

Original resume text;
Standardized claim;
Claim category;
Timeframe;
Organization or project;
Implied role and responsibility scope;
Evidence required for verification.

See

references/report-schema.md

for data format.

3. First Conduct Deterministic Internal Consistency Checks

Before conducting online research, reconcile the resume with itself:

Recalculate conversion rates, growth rates, and proportions;
Compare dates in resumes, homepages, offers, and posts;
Mark overlapping full-time or internship periods;
Compare role levels used on different platforms;
Identify superlatives like "youngest, first, core" without a clear comparison scope;
Distinguish between overall project metrics and individual impact metrics;
Check for claimed causality without a baseline or control group.

Arithmetic contradictions do not rely on external explanations and can usually be given high confidence.

When materials clearly provide numerator, denominator, and displayed percentage, write in the corresponding structured Claim:

json

"metric": {
  "numerator": 200,
  "denominator": 2000,
  "displayed_percent": 10
}

Submit to the deterministic validator for recalculation. Do not guess the denominator from natural language that only states "conversion rate 38%". Inconsistent numerical relationships only indicate inconsistency under the current caliber, and do not automatically equal candidate fraud; if an undisclosed denominator may exist, state this limitation in the analysis.

When materials clearly provide baseline, result, change type, and displayed change value, also record

baseline

result

change_type

, and

displayed_change

in the same

metric

object.

relative_percent

is recalculated as

(result - baseline) / baseline * 100

, e.g., from 100 to 125 is

25%

;

percentage_points

is directly subtracted, e.g., from 60% to 70% is

percentage points, not a

10%

relative change (the relative change is approximately 16.67%). Do not guess the change type or baseline from natural language. If the baseline for relative change is zero, state that calculation is not possible; inconsistent numerical relationships also only indicate the relationship between the provided values, and do not automatically equal candidate fraud.

4. Verify Each Claim with the Strongest Sources

Prioritize primary sources for technical and institutional claims.

Open Source and GitHub Claims

First read

references/open-source-audit.md

, then check:

Author PR and issue search;
Differences between merged, closed-unmerged, open, duplicate, and superseded;
Modified files, code ownership, review records, and release notes;
README acknowledgments, maintainer lists, CODEOWNERS, organization members, and formal appointments;
Contribution types: core engine, features, bug fixes, front-end, website, documentation, testing, generated code, comments, or only issues;
Relationship between contribution time and the claimed version/architecture phase;
Whether star count, ranking, adoption rate, and project success predated the candidate's contributions.

GitHub contribution count does not equal code merged into production; formal committer status for a subproject does not equal author status for the entire core engine.

Employment, Offer, and Internal Project Claims

Prioritize employer records, offer metadata, supervisor confirmation, internal design documents, code ownership, launch records, or metric dashboards. If unavailable, mark as

unverifiable

, do not directly judge as false.

When encountering

owner

0-to-1

core author

architect

, or

lead

, ask:

Who granted or assigned this role?
Which subsystem was actually responsible for?
Which decisions were proposed by the candidate?
Which products passed review and entered production?
Which results can be attributed to the individual rather than the team?

School and Sino-Foreign Cooperative Program Claims

Read

references/education-branding-audit.md

, then verify separately:

Institution where admission and enrollment were granted;
Actual campus or program attended;
Overseas partner institution;
Exchange or visiting status;
Status of the Sino-foreign cooperative program;
Degree-granting institution;
Final degree, diploma, or certificate obtained.

The partner institution's brand can only be used within the scope permitted by official programs and academic credentials. Do not infer fraud because the school is not well-known.

Product, Startup, and Paid User Metrics

Request definitions of registered users, active users, paid users, conversion events, denominators, time windows, revenue, refunds, and test accounts. Recalculate all visible proportions. Third-party traffic estimates cannot prove internal payment data or project attribution.

Awards and "Youngest/First/Top" Claims

Find official award lists and clear comparison sets. If there is no public ranking or exhaustible comparison scope, even if the underlying award or status is true, the superlative should be marked as

unsupported

5. Check Resume Packaging and Inflation Patterns

Read

references/inflation-patterns.md

. Must complete item-by-item evidence verification before judging whether a pattern exists. Common patterns include:

Inflating contributor to maintainer/core author;
Presenting website or documentation contributions as core engine depth;
Transferring project stars, brand, and ranking to the individual;
Using social or visibility work to imply technical authority;
Reciting complete system architecture instead of proving personal outputs;
Attributing team results and adoption rates to the individual;
Denominator switching and pseudo-precise metrics;
Repeating self-assigned titles across multiple own platforms;
Presenting cooperative schools or overseas programs as the main degree-granting body;
Time compression and repeated
```
owner/0-to-1
```
statements.

Identifying a pattern is only an investigation signal and cannot be used alone as evidence of fraud.

6. Assign Evidence Status

Use only one status per claim:

```
corroborated
```
: Strong independent evidence supports the key content of the claim.
```
partially_corroborated
```
: The fact is true, but the role, scope, causality, or magnitude exceeds the evidence.
```
contradicted
```
: Reliable evidence directly conflicts with the claim.
```
unsupported
```
: Insufficient evidence remains after reasonable search.
```
unverifiable
```
: Only currently unavailable internal or private records can verify the claim.

Also record confidence level (

high

medium

low

) and explain what evidence could change the conclusion.

7. Build Report Data

Create

report-data.json

according to

references/report-schema.md

. Keep references brief, and every important conclusion must point to a source. Accusations must be attributed, for example:

The article author alleges...; original screenshots are missing from current materials, and independent verification has not been completed.

Do not write directly:

The party forged...

Unless direct, authoritative evidence has proven the conclusion.

8. Generate HTML and PDF

Run in the skill directory:

bash

python3 scripts/render_report.py \
  --input /absolute/path/report-data.json \
  --html /absolute/path/report.html \
  --pdf /absolute/path/report.pdf

The HTML must remain self-contained and use the included paper-style evidence report design, including at least:

Scope and limitations;
Evidence status cards;
Investigation flow chart;
Highest-risk claims;
Complete claim-evidence matrix;
Packaging patterns and counter-evidence;
Timeline;
Sources and next verification steps.

9. Validate and Visually Inspect

Run:

bash

python3 scripts/validate_report.py \
  --data /absolute/path/report-data.json \
  --html /absolute/path/report.html \
  --pdf /absolute/path/report.pdf

pdftoppm

is installed, render the PDF as PNG and check each page to ensure no text cropping, missing Chinese characters, broken tables, or overlapping elements. Modify before delivery if issues are found.

10. Deliver with Calibrated Language

Conclusion order:

Strongest confirmed facts;
Strongest direct contradictions;
Largest unverifiable gaps.

Do not repeat incendiary language from articles. Clearly distinguish between

literal fiction

role/scope inflation

unsupported marketing statements

, and

merely lack of evidence

Attached References

```
references/evidence-methodology.md
```
: Source hierarchy and status judgment rules.
```
references/open-source-audit.md
```
: GitHub/Apache contribution and role audit.
```
references/education-branding-audit.md
```
: Institution, Sino-foreign cooperation, exchange, and degree representation.
```
references/inflation-patterns.md
```
: Defensive resume packaging classification.
```
references/report-schema.md
```
: JSON format read by the report generator.
```
references/asu-case-study.md
```
: Limited case study based on public records and user materials; do not treat it as eternal or comprehensive fact.

Attached Tools

```
scripts/render_report.py
```
: Deterministic HTML/PDF generator.
```
scripts/validate_report.py
```
: Data structure, evidence reference, HTML, and PDF smoke check.
```
assets/report_template.html
```
: Self-contained Chinese HTML report template.
```
evals/evals.json
```
: Chinese test tasks and expected behavior.

asu-resume-audit-skill

NPX Install

Tags

SKILL.md Content (Chinese)

ASU Resume Authenticity Audit Skill

Mandatory Deliverables

Evidence and Harm Boundaries

Workflow

1. Define Scope and Save Source Collection

2. Break the Resume into Atomic Claims

3. First Conduct Deterministic Internal Consistency Checks

4. Verify Each Claim with the Strongest Sources

Open Source and GitHub Claims

Employment, Offer, and Internal Project Claims

School and Sino-Foreign Cooperative Program Claims

Product, Startup, and Paid User Metrics

Awards and "Youngest/First/Top" Claims

5. Check Resume Packaging and Inflation Patterns

6. Assign Evidence Status

7. Build Report Data

8. Generate HTML and PDF

9. Validate and Visually Inspect

10. Deliver with Calibrated Language

Attached References

Attached Tools