# Usage: curl -sSL https://seed.show/hiring.resume.screening | bash -s <install-path>
# <install-path> is the directory where the file should land.

set -euo pipefail
[ -z "${1:-}" ] && {
  echo "install requires a path: curl -sSL https://seed.show/hiring.resume.screening | bash -s <install-path>" >&2
  exit 1
}
TARGET="$1"
mkdir -p "$TARGET"
DEST="$TARGET/seed-fold.GE5SpX.folded.md"

cat > "$DEST" <<'PORTDOWN_81B14406'
<!--
This is a `.folded.md` archive — a directory packed into one markdown
file. The markers below are load-bearing; don't edit them directly.

To unpack (requires bash — if you have no shell, tell the user):
  1. curl -sSLf https://fold.dom.vin/skill | bash -s <INSTALL_DIR>
  2. <INSTALL_DIR>/fold/scripts/unfold <this-file>
     (or: unfold <this-file>  if fold/scripts is on your PATH)
-->

---
fold: true
marker: d83e57
at: 2026-05-07T16:16:10Z
root: seed-pack.c3Ljnd
---

<!--fold:d83e57@file path="README.md" mode="644"-->
# Resume Screening Context

## What this is and why the legal stakes are high

Resume screening is the first structured elimination step in a hiring pipeline. Done right: a documented rubric filters candidates against job-necessary criteria, passing qualified candidates into structured assessment. Done wrong: pattern-matching on proxies produces demographic disparity, and the employer is liable — including when the screening is automated.

**AI-assisted screening faces the highest regulatory scrutiny in hiring right now.** EEOC, OFCCP, and multiple state/local jurisdictions have moved specifically to address automated employment decision tools. The legal exposure is not theoretical; it's the subject of active enforcement guidance and the first wave of local law. Agents assisting with screening must understand this context before touching a single resume.

**The hard constraint:** AI agents must not make final screening decisions. They can apply a documented rubric, surface criteria matches, and flag gaps. The pass/fail determination on candidates must have a human in the loop who can document the rationale. This isn't just good practice — it's the posture required to maintain defensible employer liability under current EEOC guidance and, in some jurisdictions (NYC, Illinois), a legal requirement.

---

## Mental model: screening as structured elimination

The purpose of resume screening is a single, precise question: **does this candidate have the demonstrated capability to do this job well?** Not prestige. Not familiarity. Not cultural comfort. Predictive validity for job performance — and only that.

Screening is the first filter in a pipeline. Its job is to pass through candidates who meet minimum qualifications for structured assessment — not to rank candidates, not to select the winner, not to substitute for an interview.

Resumes are a noisy signal. They encode proxies that recruiters and agents have learned to use as shortcuts — school brand, employer brand, GPA, linear career progression — that are consistently weaker predictors of job performance than the actual competencies they allegedly signal. The foundational research: Schmidt & Hunter (1998) meta-analysis reviewed 85 years of selection research; the highest-validity predictors are structured assessments of cognitive ability and specific job knowledge. Brand proxies are not in the top tier. See sources.md.

**The central legal insight:** inconsistency is the primary risk vector. A rubric applied inconsistently across candidates is both ineffective and legally dangerous — inconsistency is direct evidence of disparate treatment, and it amplifies disparate impact. Document criteria before reviewing resumes. Apply them identically.

---

## Structured vs. unstructured screening

**Unstructured screening** is what most people do by default: read a resume, form an impression, make a call. The impression is dominated by heuristics — name, school, employer, career trajectory format — that correlate with demographic characteristics more than with job performance. This is where bias enters and where legal exposure lives.

**Structured screening** replaces impression with criteria. Before any resume is reviewed:

1. **Define minimum qualifications (MQs)** — the non-negotiable requirements where absence makes the job undoable. A minimum qualification must be directly tethered to a task in the job description. "5 years experience" is rarely a valid MQ unless the specific task genuinely requires accumulated experience that cannot be assessed otherwise.

2. **Define preferred qualifications (PQs)** — attributes associated with faster ramp-up or higher ceiling performance. PQs add signal but are not eliminators.

3. **Define a scoring rubric before the first resume is read.** What does "meets MQ" look like? What does "strong on PQ1" look like? The rubric must be applied consistently across all candidates.

4. **Blind the rubric to irrelevant signals.** School name, employer brand, and GPA belong on a "do not score" list unless the role has a documented, validated reason to treat them as predictive. They rarely do.

5. **Apply criteria identically.** The same MQ that disqualifies candidate A must disqualify candidate B. Inconsistent application — even well-intentioned — is both a bias vector and a legal liability.

---

## Minimum qualifications vs. preferred qualifications

This distinction is load-bearing and commonly collapsed.

**Minimum qualifications** are pass/fail. Failing one MQ makes the candidate ineligible. For this reason, MQs must be genuinely job-necessary — not aspirational, not "nice to have if we can get it." Setting MQs stricter than the job requires is the most common route to disparate impact, because credential requirements correlate with race, sex, and socioeconomic status.

EEOC's Uniform Guidelines (29 CFR Part 1607, see sources.md) require that selection procedures — including MQs — that produce adverse impact must be validated as job-related and consistent with business necessity. "We've always required a degree for this role" does not constitute validation.

**Preferred qualifications** are scored, not gated. A candidate missing a PQ remains in the pool; they score lower on that dimension. Preferred qualifications should also be tied to job tasks, but the standard is practical relevance rather than necessity.

---

## What agents get wrong in resume screening

### Using proxies for protected characteristics

Several resume attributes correlate with protected class membership and are not valid predictors of job performance:

- **School name** correlates with race, socioeconomic status, and first-generation college student status.
- **Employer brand** correlates with prior access to networks and credentialing systems that are themselves unequally distributed.
- **Name and graduation year** can surface inferences about race, national origin, and age.
- **Employment gaps** are disproportionately held against women (childcare leave) and people with disabilities or chronic illness.
- **"Culture fit"** is undefined and captures whatever the evaluator's affinity bias points at.

An agent that surfaces "strong profile — Harvard, McKinsey" is doing brand pattern-matching, not performance prediction. That output will amplify historical access inequalities, not job-performance signal.

### Treating GPA and school prestige as predictive

They are not, in general populations. GPA predicts academic performance; academic performance predicts job performance weakly and inconsistently, with effect size varying substantially by job type and industry. For most roles — especially after 3–5 years of work experience — GPA is noise with demographic correlates.

School prestige (selectivity ranking) has similar properties: it measures who was admitted, not what candidates can do. Admissions processes are heavily influenced by socioeconomic background, geography, and legacy status. Treating selectivity as a job-performance proxy inherits those correlations.

The correct test: can you articulate a specific job task that the credential predicts better than a direct assessment of the relevant skill? If not, the criterion is a proxy.

### Inconsistent criteria application

A rubric applied differently across candidates is both ineffective and legally risky. Common patterns:

- Requiring demonstrated experience from candidates in some demographic groups while crediting "potential" for others.
- Applying MQs to the letter for some candidates and waiving them via holistic judgment for others.
- Scoring the same evidence (e.g., a 2-year gap) differently based on unstated context that correlates with protected characteristics.

An agent running screening across a large pool will amplify any inconsistency at scale. Inconsistency in the prompt, in the rubric, or in how ambiguous cases are resolved becomes inconsistency across thousands of candidates.

### Over-relying on keyword matching

Keyword matching measures whether a candidate thought to include the word, not whether they have the skill. It systematically disadvantages candidates who describe the same competency in different terminology — which correlates with educational background, industry of origin, and ESL status. Keywords are a useful pre-filter for clearly-defined technical requirements. They are not a scoring system. See failure-modes.md for the full failure mode with legal analysis.

---

## What AI is changing

### Regulatory attention has shifted to automated tools specifically

EEOC's Artificial Intelligence and Algorithmic Fairness Initiative (active, see sources.md) addresses AI-assisted hiring explicitly: employers using algorithmic screening tools face the same adverse impact analysis requirements as any other selection procedure, and vendor use does not transfer liability. "Our AI vendor handles bias" is not a defense. The employer is responsible for whether the tool produces adverse impact, period.

OFCCP has signaled parallel focus for federal contractors. Federal contractors must maintain applicant flow data — including data on candidates screened out by automated tools — in a form that supports adverse impact analysis. If the screening tool can't produce that audit trail, the contractor has a compliance problem before any discrimination question is reached.

### State and local law has moved faster than federal

**NYC Local Law 144** (effective July 2023): Employers using automated employment decision tools (AEDTs) in New York City hiring must commission independent bias audits of those tools, publish the results, and notify candidates they are being evaluated by an AEDT. This is the first municipal law requiring bias audits as a condition of use — not just a recommended practice.

**Illinois AI Video Interview Act** (effective January 2020, amended): Requires employers to notify applicants when AI is used to analyze video interviews, explain how the AI works, and get consent. The amendment extended requirements to third-party vendors. Illinois has continued to expand AI hiring law; check for current amendments before advising on Illinois-based hiring.

**Colorado, Maryland, New York (state)**: Active legislative activity on algorithmic employment discrimination as of 2024–2025. Fetch current state of law for specific jurisdiction.

### Proxy discrimination through resume parsing

This is the mechanism most likely to create AI-specific liability. Resume parsing models trained on historical data inherit the demographic structure of that data:

- A model trained on "successful hires" at an organization that hired predominantly from a small set of universities will score those universities higher — not because they predict performance, but because they were common in training data.
- NLP models that parse job title, seniority, and role descriptions can encode gender bias from gendered language patterns in historical job postings and resumes.
- Geographic parsing can function as a proxy for race or national origin depending on the specific location clusters the model learned.

The problem is that none of this is visible in the prompt or the rubric — it's in the model's learned representations. An employer using an AI tool cannot audit this without requesting the vendor's bias audit data.

### What defensible AI-assisted screening looks like

Defensible: AI applies a documented, human-authored rubric to surface criterion matches. Human reviews AI output before any pass/fail decision. Audit trail exists: which criteria were applied, how each candidate scored, who made the final call, when. Adverse impact analysis is run on outcomes before the process is finalized.

Creates liability: AI ranks candidates without a documented rubric. AI makes or recommends pass/fail decisions without human review. No audit trail of criteria application. No adverse impact analysis. Tool selected from vendor without requesting bias audit data.

The line is documentation and human accountability — not whether AI is in the loop at all.
<!--fold:d83e57@file path="failure-modes.md" mode="644"-->
# Resume Screening Failure Modes

This file covers the specific bias failure modes that AI-assisted resume screening amplifies, with concrete examples of how each manifests, and the legal framework that makes each risky.

---

## Legal framework first: two distinct liability theories

Understanding the legal structure matters because different failure modes map to different theories, and the defenses differ.

**Disparate treatment** — intentional discrimination. Treating a candidate differently because of a protected characteristic (race, sex, national origin, age, religion, disability, etc.). Evidence: different criteria applied to different groups, derogatory remarks in notes, facially neutral criteria with demonstrably pretextual application. Defense: very limited. You don't get to show the business was served; intent is the issue.

**Disparate impact** — facially neutral practice with disproportionate exclusionary effect on a protected group. No intent required. Evidence: statistical analysis showing the selection rate for one group is less than 4/5 (80%) of the highest-selecting group — the "four-fifths rule" from EEOC Uniform Guidelines 29 CFR § 1607.4(D). Defense: the employer must show the criterion is job-related and consistent with business necessity. If they can, the plaintiff must then show a less discriminatory alternative existed.

Most AI screening failures produce **disparate impact**, not disparate treatment. The system isn't "intending" to discriminate. It's applying pattern-matching that reflects historical distributions — and those distributions were themselves shaped by prior discrimination. That's what makes AI screening legally dangerous: it can launder structural inequality through an appearance of neutral automation.

---

## Failure mode 1: Prestige bias (school and employer name)

**What it is.** Weighting candidates higher because they attended selective universities or worked at brand-name companies — without evidence that those attributes predict job performance.

**How it manifests in AI screening.** Training data or prompt instructions that encode prestige as signal. A model asked to "identify the strongest candidates" will draw on patterns in its training that associate "strong" with recognizable institution names — because those associations are pervasive in professional text corpora. The output looks objective ("this candidate has an excellent background") but reflects inherited prestige hierarchies.

**Concrete example.** A screening prompt flags candidates from FAANG companies and Ivy League schools as "top-tier" and candidates from state schools or unknown employers as "needs further review." The effect: candidates who were admitted to selective schools (a process heavily correlated with parental income, legacy status, geography, and race) are systematically ranked higher. Candidates from historically Black colleges and universities (HBCUs), community colleges, or international universities not in the model's training distribution are disadvantaged with no performance-relevant rationale.

**Why it's legally risky.** School selectivity and employer brand both correlate with race and socioeconomic status. Weighting them produces disparate impact. When challenged, the employer cannot demonstrate job-relatedness because the research literature does not support prestige as a predictor (Schmidt & Hunter 1998, Sackett et al. 2022 — see sources.md). Defending this practice requires validation evidence you almost certainly don't have.

**What structured screening does instead.** If a specific credential is genuinely required (e.g., licensure), make it an explicit MQ tied to a job task. Otherwise, screen for the demonstrated skill the credential supposedly signals — not the credential itself.

---

## Failure mode 2: Affinity bias (similar background)

**What it is.** Rating candidates more favorably when their background resembles the reviewer's own — or, in AI systems, resembles the profiles in whatever corpus was used to define "good."

**How it manifests in AI screening.** Models trained on or evaluated against examples of "successful" hires in a homogeneous organization will reproduce the demographic profile of that organization. If the training set of "successful engineers" is predominantly male and from a handful of universities, the model will score resumes toward that profile. The model is not reasoning about performance; it's doing pattern completion.

**Concrete example.** An AI screening tool built on a company's existing "high performer" pool — which is 80% white men from four universities — systematically scores resumes from that demographic higher, even when the underlying competencies are identical. Women who have taken non-linear career paths (common with childcare leave), candidates from different cultural naming conventions, or candidates who describe the same skills in different terminology are rated lower without a performance-relevant basis.

**Why it's legally risky.** If the model's training data encodes historical discrimination, the model reproduces it at scale. The employer is liable even if the model was built by a vendor — EEOC AI guidance is explicit that vendor use does not transfer liability. Disparate impact analysis will expose the statistical signature; "we used an AI" is not a defense.

**What structured screening does instead.** Define competencies before reviewing resumes. Score against the rubric, not the gestalt impression. Blind names, schools, and employers from the scoring pass where possible — blind review consistently reduces demographic bias in outcomes.

---

## Failure mode 3: Recency bias

**What it is.** Over-weighting recent experience or the most recent employer, under-weighting older but highly relevant experience.

**How it manifests in AI screening.** In a resume, the most recent role tends to get disproportionate attention. A candidate with 15 years of directly relevant experience whose most recent role was adjacent scores lower than a candidate with 3 years of experience in the exact current category.

**Concrete example.** A candidate returns to data engineering after three years in management. The AI scorer focuses on the management role (most recent), identifies it as "not a match" for a senior data engineering position, and downgrades the candidate — despite 12 years of prior data engineering experience at the relevant seniority level. The management transition is penalized rather than understood as a development arc.

**Why it's legally risky.** Recency bias intersects with age discrimination when it systematically disadvantages candidates whose long career records are front-loaded with experience from years ago. Career arcs that include breaks, pivots, or management transitions correlate with sex (women more likely to have caregiving-related gaps) and disability. A facially neutral "focus on recent experience" instruction can produce demographic disparity at scale.

**What structured screening does instead.** The rubric specifies how many years of relevant experience are required and where on the resume they can appear. "10 years of experience in X at any point in career" is a different criterion from "current role in X."

---

## Failure mode 4: Halo effect

**What it is.** One strong signal inflating the overall assessment. In resume screening: a prestigious credential or employer causes the reviewer (human or AI) to rate all subsequent evidence more favorably than warranted.

**How it manifests in AI screening.** A model that processes a resume top-to-bottom updates its "this is a strong candidate" prior when it hits a recognizable brand early, then interprets subsequent evidence charitably. The same experience described in the same words gets rated differently depending on whether it appears after a prestige marker or after an unrecognized employer.

**Concrete example.** Candidate A: Google → 3 years as an analyst → 2 years at an early-stage startup in an ambiguous role. AI scorer: "strong candidate, solid progression." Candidate B: state university → 3 years as an analyst at a regional firm → same 2 years at the same startup in the same ambiguous role. AI scorer: "limited brand recognition, unclear trajectory." The actual competency evidence in years 4–5 is identical. The halo from Google inflated Candidate A's entire subsequent record.

**Why it's legally risky.** Halo from prestige markers amplifies prestige bias (above) — same demographic correlates, same disparate impact mechanism. Additionally, inconsistent scoring of identical evidence is direct evidence of disparate treatment if it correlates with protected characteristics.

**What structured screening does instead.** Score each criterion independently before forming any overall assessment. The rubric asks "does this candidate meet the requirement for [specific competency]?" separately for each competency — not "is this a good candidate overall?" Holistic impressions are the mechanism of halo; criteria separation disrupts it.

---

## Failure mode 5: Employment gap penalization

**What it is.** Treating gaps in employment history as negative signals without a job-performance rationale.

**How it manifests in AI screening.** A model asked to identify "consistent, progressive career trajectories" will flag gaps as inconsistencies. Most models will not ask why the gap exists; they will register it as a negative pattern.

**Concrete example.** A candidate who took 18 months off to care for a parent with a terminal illness has a gap that the AI flags as a concern. The same 18 months another candidate spent at an employer produces a positive signal. The caregiving work — often involving project coordination, crisis management, financial administration — is invisible to the screen.

**Why it's legally risky.** Employment gaps correlate with sex (women disproportionately provide unpaid caregiving), disability (health conditions and treatment), and national origin (immigration transitions, family obligations in other countries). The ADA further constrains how disability-related gaps can be treated. Penalizing gaps without a job-performance rationale produces disparate impact on multiple protected classes.

**What structured screening does instead.** Remove gap analysis from the screening rubric unless a documented, validated business necessity requires continuous employment. Screen for the competencies; if continuous recent practice matters (e.g., a licensed professional maintaining CE hours), make that the explicit, job-anchored criterion — not "no gaps."

---

## Failure mode 6: Keyword-based exclusion at scale

**What it is.** Excluding candidates who have the skill but not the specific term. At scale, keyword matching becomes a systematic exclusion mechanism rather than a useful pre-filter.

**Concrete example.** A job description requires "Salesforce experience." A candidate with 5 years of Salesforce administration has described it as "CRM management" and "Salesforce.com" throughout their resume. The keyword screen rejects them. A candidate with 6 months of Salesforce exposure who used the exact term passes.

**Why it's legally risky.** Terminology gaps correlate with ESL status, educational background, industry of origin, and career age. Candidates who learned skills in non-US contexts, in different industry vocabularies, or before a product was rebranded use different terminology. Systematic keyword exclusion can produce disparate impact on national origin and age grounds.

**What structured screening does instead.** Use keywords as a preliminary flag, not an eliminator. Confirm that keyword matches are backed by substantive evidence, and that keyword misses aren't masking substantive evidence.

---

## Failure mode 7: Proxy discrimination through resume parsing (AI-specific)

**What it is.** An AI model encodes demographic correlates into its scoring through the statistical structure of its training data — without any explicit instruction to consider protected characteristics. The discrimination is in the model's learned representations, not in the prompt.

**How it manifests.** Resume parsing models learn from historical corpora. Those corpora reflect historical hiring patterns, which were themselves shaped by prior discrimination. The model doesn't "know" it's encoding race or sex; it's learning that certain clusters of features co-occur with "strong candidate" labels in training data. Those feature clusters may be highly correlated with protected class membership.

**Concrete examples:**
- A model trained on job titles and seniority descriptors from historical postings absorbs gendered language patterns. "Administrative" and "support" roles are learned as lower-status; "technical" and "leadership" roles are learned as higher-status. Candidates who occupied roles described with feminized language score lower even when the underlying work was equivalent.
- A model trained on name → outcome associations from historical hiring data (without explicitly using name as a feature) can encode proxy discrimination through co-occurring features like university name, neighborhood, or organizational name that are geographically and racially patterned.
- A model that learns to parse "leadership experience" from specific role title formats disadvantages candidates from industries or cultures where leadership is described differently.

**Why it's legally risky.** This failure mode is nearly impossible to detect from the prompt or rubric — it requires auditing the model's outputs across demographic groups. Under EEOC AI guidance, the employer is responsible for whether the tool produces adverse impact regardless of whether the mechanism is understood. Under NYC Local Law 144, using an AEDT without a bias audit from an independent auditor is itself a violation, irrespective of whether discrimination is demonstrated.

**What defensible practice looks like.** Request the vendor's bias audit before deploying any AI screening tool. If no audit exists, do not deploy. After deployment, run adverse impact analysis on actual screening outcomes (not just the tool's theoretical properties) before using results to make decisions.

---

## Failure mode 8: Absent audit trail (AI-specific)

**What it is.** AI screening runs without a documented record of what criteria were applied, how each candidate was scored, and who made the final determination. When a decision is later challenged, the employer cannot reconstruct what happened.

**How it manifests.** An agent screens 500 resumes, produces a ranked list, and a human selects from the top 50. No record exists of the criteria applied, the scores assigned to each candidate, or why any candidate was excluded. The process appears efficient; the audit trail is empty.

**Why it's legally risky.** Without documentation, the employer cannot demonstrate the process was job-related and consistently applied. When a rejected candidate files an EEOC charge or a discrimination lawsuit, the employer's only defense is "our process was fair" — which they cannot prove. For OFCCP-covered federal contractors, the record-keeping requirement is explicit: applicant flow data must be maintained in a form that supports adverse impact analysis. "The AI did it and we didn't log it" does not satisfy this requirement.

**What defensible practice looks like.** Before any AI screening step, document: the criteria being applied, the rubric for each criterion, and who is responsible for final pass/fail decisions. After the screening step, retain: the AI's output for each candidate (scores or assessments by criterion), the final pass/fail determination for each candidate, the identity of the human who made or reviewed that determination, and the date. This documentation is not optional — it's the evidence that a challenged process was fair.

---

## Failure mode 9: Disparate impact through proxy variables (AI-specific)

**What it is.** A screening criterion that appears facially neutral produces disproportionate exclusion of a protected group because it correlates with a protected characteristic. The criterion is a proxy, not a predictor.

**How it manifests in AI screening.** AI systems are particularly prone to discovering and using proxy variables because they can detect correlations that humans would not recognize as problematic. A model optimizing for "candidates similar to current high performers" may learn that ZIP code, university name, or even resume formatting style predicts group membership — and use those signals without any human recognizing that it's doing so.

**Concrete examples:**
- ZIP code or city neighborhood as a signal: encodes race and socioeconomic status in most US metropolitan areas.
- University graduation year without age as an explicit feature: a model can infer approximate age from graduation year and penalize older candidates.
- Resume length and formatting conventions: vary by educational background and cultural norms; a model trained on US-standard resume formats will disadvantage candidates from cultures with different conventions (e.g., CVs that include personal information or that list credentials before experience).

**The four-fifths test as minimum standard.** For any AI-assisted screening step, run the four-fifths calculation on actual outcomes before finalizing the process: if the selection rate for any protected group is below 80% of the highest-selecting group's rate, adverse impact exists and the burden shifts to the employer to demonstrate job-relatedness. This calculation needs to be run on real screening output, not on theoretical properties of the rubric.

**What structured screening does instead.** Enumerate the features the AI is permitted to use. Prohibit features that correlate with protected characteristics unless a documented, validated job-relatedness justification exists. Treat any feature not on the approved list as off-limits — the model should not be discovering its own criteria.

---

## Summary: the audit test

Before deploying any AI-assisted screening step, the employer (or their agent) should be able to answer:

1. **What criteria are being applied?** List them explicitly, in writing, before the first resume is reviewed.
2. **For each criterion: what is the job-performance justification?** If you can't state one, the criterion is a proxy.
3. **What features is the AI permitted to use?** If the answer is "whatever the model decides," you do not have a defensible process.
4. **What does the demographic distribution of passthrough look like?** Run the four-fifths calculation. If any group's pass rate is below 80% of the highest group's, adverse impact exists and validation is required.
5. **Is the same criterion being applied identically across all candidates?** Inconsistency is both a bias vector and legal evidence.
6. **Has the vendor provided an independent bias audit?** Under NYC Local Law 144, this is required for automated employment decision tools used in NYC hiring. "The vendor handles it" is not an answer — employer liability remains.
7. **Is the audit trail complete?** Can you reconstruct what criteria were applied, how each candidate scored, who made the final determination, and when? If not, you cannot defend the process.

The EEOC AI guidance and NYC Local Law 144 make clear that these questions are not optional for employers using automated screening. They are the legal baseline.
<!--fold:d83e57@file path="sources.md" mode="644"-->
# Resume Screening Sources

## Federal legal authority

**EEOC Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607)**
- https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XIV/part-1607
- The foundational federal regulation governing selection procedures, including resume screening. Establishes the adverse impact (four-fifths rule) standard and the validation requirements that apply when adverse impact is found. Any MQ or screening criterion that produces disparate impact must be validated as job-related and consistent with business necessity under this framework.

**EEOC — Enforcement Guidance and Technical Assistance**
- https://www.eeoc.gov/laws/guidance
- Includes guidance on disparate impact, pre-employment inquiries, and employment tests. The "Questions and Answers to Clarify and Provide a Common Interpretation of the Uniform Guidelines on Employee Selection Procedures" (1979) is particularly useful for understanding practical application of the four-fifths rule and validation requirements. Fetch before advising on any specific criterion — enforcement priorities shift.

**EEOC — Artificial Intelligence and Algorithmic Fairness Initiative**
- https://www.eeoc.gov/artificial-intelligence-and-algorithmic-fairness
- EEOC's active guidance on AI-assisted hiring tools, adverse impact analysis for algorithmic selection, and employer liability when using third-party AI screening vendors. Updated as enforcement develops — fetch before advising on AI-assisted screening.

**OFCCP (Office of Federal Contract Compliance Programs) — Federal Contractor Requirements**
- https://www.dol.gov/agencies/ofccp
- Federal contractors and subcontractors have affirmative action obligations under Executive Order 11246, the Rehabilitation Act, and VEVRAA. OFCCP enforces record-keeping requirements for applicant flow data that are directly relevant to screening audits.
- OFCCP Directive 2022-01 on Predetermination Notices and Technical Assistance: https://www.dol.gov/agencies/ofccp/directives

## Core research

**Schmidt, F.L. & Hunter, J.E. (1998). "The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 85 Years of Research Findings."** *Psychological Bulletin*, 124(2), 262–274.
- The landmark meta-analysis establishing the relative predictive validity of selection methods. General mental ability tests and work sample tests outperform unstructured interviews, GPA, and credentials. The empirical basis for why prestige-proxy screening is methodologically weak.
- Abstract and citation: https://doi.org/10.1037/0033-2909.124.2.262

**Sackett, P.R., Zhang, C., Berry, C.M., & Lievens, F. (2022). "Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range."** *Journal of Applied Psychology*, 107(11), 2040–2068.
- Updated validity estimates for selection methods. Refines Schmidt & Hunter; still the most current comprehensive meta-analysis in the field.
- https://doi.org/10.1037/apl0000994

## SHRM and practitioner guidance

**SHRM — Structured Interviewing Guide**
- https://www.shrm.org/topics-tools/tools/toolkits/interviewing-candidates
- Practitioner guidance on structured vs. unstructured approaches. SHRM resources require a member login but many toolkits are available publicly; fetch to verify current access state.

**SHRM — Background Checks and Screening Compliance**
- https://www.shrm.org/topics-tools/tools/toolkits/conducting-background-investigations
- Covers FCRA compliance for background checks, ban-the-box laws (which affect what screening criteria are permissible at what stage), and state-specific restrictions.

**SHRM — Avoiding Bias in Hiring**
- https://www.shrm.org/topics-tools/topics/talent-acquisition/avoiding-bias-hiring
- Practitioner-focused overview of bias types in hiring, structured screening approaches, and blind review practices.

## State and local law (US)

Resume screening criteria are subject to a growing body of state and local legislation beyond federal EEOC requirements. This area is moving fast; fetch current state of any jurisdiction before advising.

- **Salary history bans** (prohibit using prior compensation in screening): in effect in California, New York, Massachusetts, Illinois, and many municipalities. Full list: https://www.hrdive.com/news/salary-history-ban-states-list/516662/
- **Education requirement scrutiny**: Several states (Maryland, Colorado, Pennsylvania, others) have removed degree requirements from public sector positions; private-sector scrutiny is growing. Some EEOC guidance treats blanket degree requirements as potential disparate impact sources absent validation.
- **NYC Local Law 144 (AI hiring tool bias audits)**: Effective July 2023. Requires employers using automated employment decision tools (AEDTs) in NYC hiring to commission independent bias audits, publish results, and notify candidates. The audit must be conducted by an independent auditor and posted publicly before the tool is used. Applies to AEDTs that "substantially assist or replace discretionary decision making." https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page
- **Illinois Artificial Intelligence Video Interview Act**: Effective January 2020, amended to extend to vendors. Requires employers to notify applicants when AI analyzes video interviews, explain how it works, and obtain consent. Illinois has continued expanding AI-in-hiring requirements; check for current amendments.
- **Colorado, Maryland, New York (state)**: Active legislative activity on algorithmic employment discrimination as of 2024–2025. Treat as in-flux; verify current law for any specific engagement.

## How to use these sources

1. **Legal baseline**: Start with 29 CFR Part 1607 (Uniform Guidelines) before advising on any MQ or screening criterion. If the criterion produces adverse impact, the business necessity / job-relatedness analysis is required — not optional.
2. **AI-specific risk**: Fetch the EEOC AI guidance before advising on any AI-assisted screening tool. Vendor use does not transfer liability to the vendor.
3. **Validity claims**: Use Schmidt & Hunter (1998) and Sackett et al. (2022) for evidence-based pushback on prestige-proxy criteria. "This has been studied extensively; GPA is a weak predictor for most roles" is a defensible empirical statement with citations.
4. **Jurisdiction check**: NYC Local Law 144 and state salary history bans require checking local law before finalizing screening criteria in those jurisdictions.
<!--fold:d83e57@end-->
PORTDOWN_81B14406

# ── post ──
MARKER=$(awk '/^---$/ { f++; if (f==2) exit; next } f==1 && /^marker:[[:space:]]/ { sub(/^marker:[[:space:]]+/, ""); print; exit }' "$DEST")
[ -z "$MARKER" ] && { echo "seed: archive has no marker — corrupt" >&2; exit 1; }
awk -v m="$MARKER" -v outdir="$TARGET" '
  BEGIN {
    # Match <!--fold:<m>@file path="X"--> with an optional mode attr after
    # the path (fold emits  mode="644"  on executables).
    file_re = "^<!--fold:" m "@file path=\"([^\"]+)\"( mode=\"[0-9]+\")?-->$"
    end_re  = "^<!--fold:" m "@end-->$"
  }
  $0 ~ end_re { if (current) close(current); exit }
  $0 ~ file_re {
    if (current) close(current)
    line = $0
    sub(/^<!--fold:[^@]+@file path="/, "", line); sub(/".*$/, "", line)
    current = outdir "/" line
    dir = current; sub(/\/[^\/]*$/, "", dir)
    if (dir != current) system("mkdir -p \"" dir "\"")
    printf "" > current
    next
  }
  current { print >> current }
' "$DEST"
SEED_EXTRACTED=$(find "$TARGET" -type f -not -path "$DEST" 2>/dev/null | wc -l)
if [ "$SEED_EXTRACTED" = "0" ]; then
  echo "seed: archive contained no files — refusing to delete the source" >&2
  echo "  archive preserved at: $DEST" >&2
  exit 1
fi
rm -f "$DEST"

echo "" >&2
echo "✓ seed unpacked → $TARGET ($SEED_EXTRACTED files)" >&2
find "$TARGET" -type f | sort | while IFS= read -r _sf; do
  echo "  ${_sf#${TARGET}/}" >&2
done
echo "" >&2
if [ -f "$TARGET/SKILL.md" ]; then
  echo "This seed contains a skill (SKILL.md). Install it in your agent's skills directory." >&2
  echo "" >&2
fi
echo "Install the seed skill if not already installed:" >&2
echo "  https://seed.show/skill" >&2
echo "" >&2
echo "Publisher prompt:" >&2
sed 's/^/  /' >&2 <<'__SEED_PROMPT_END_AC1F2B__'
You have the resume screening context. Read README.md for the screening mental model and the 'What AI is changing' section on current regulatory exposure (EEOC AI initiative, NYC Local Law 144, IL AI Video Interview Act, proxy discrimination). Read failure-modes.md for the specific bias and legal failure modes AI screening amplifies — including AI-specific modes: proxy discrimination through resume parsing, absent audit trails, and disparate impact through proxy variables. Fetch sources.md for authoritative guidance and citation. To share your own context bundles: curl seed.show/skill | bash
__SEED_PROMPT_END_AC1F2B__
exit 0