/resume-summarizer

⚠️ Not HR or employment-law advice. This output is a drafting and triage aid generated by a product-organization skill, not HR counsel or employment counsel. No attorney-client relationship is created by its production or use. Jurisdiction-specific questions on automated-employment-decision-tool classification, protected-class signaling, candidate disclosure, Ban-the-Box application, caregiver status, GINA exposure, or adverse impact require review by a licensed employment attorney in the relevant jurisdiction and by a qualified HR professional. Do not rely on this output as the sole basis for any hiring, screening, ranking, scoring, or employment decision.

Jurisdiction Assumed: {jurisdiction from required input — e.g., US-NYC, US-IL, US-CO, US-CA, EU, UK, IL}. If your jurisdiction differs, treat every finding below as a hypothesis to verify with local counsel.

AEDT Non-Classification Memo (embedded, visible, load-bearing)

This skill is deliberately not classified as an Automated Employment Decision Tool under NYC Local Law 144 (NYC Admin Code §20-870 to §20-874). The full legal argument lives in

hr-ai-governance

The eight bright lines (

hr-ai-governance

experience_years

12

Review cadence (

hr-ai-governance

Deployer warning (runtime, NYC jurisdiction):

This skill outputs structured extraction only and is not classified as an AEDT under NYC Local Law 144. If your workflow uses this output to filter, route, auto-reject, or pre-rank candidates before a human review, the classification may change and you become subject to Local Law 144 obligations including annual bias audit by an independent auditor, public audit results, and 10-business-day candidate notice. Consult your employment counsel.

The deployer's acknowledgment is required before first production use per tenant and re-attested annually (

hr-ai-governance

Positioning

This output is descriptive, not evaluative. Any ranking or comparison must be performed by a human reviewer. This skill is deliberately not classified as an Automated Employment Decision Tool under NYC Local Law 144.

That sentence is not decorative. It is the AEDT wall made visible, and it appears verbatim at the top of every production output. The rest of this skill exists to hold that sentence true.

Purpose

/resume-summarizer

What it IS: a descriptive, non-evaluative, structured-extraction skill that runs every resume through the

hr-ai-governance

What it is NOT: a screener, a ranker, a scorer, a candidate comparator, a fit assessor, a seniority classifier, a recommendation engine, or an Automated Employment Decision Tool. Any prompt asking this skill to rank, score, compare, classify seniority, or recommend is refused with a routing message to human review.

This skill is governed by

hr-ai-governance

When to Use

• High-volume candidate pools (20+ resumes) where structured extraction helps human reviewers triage efficiently
• Standardized candidate database ingestion from ATS imports, referral uploads, or career-page submissions
• Batch re-extraction after a schema update (e.g., new jurisdiction added to the matrix, new proxy added to the register)
• Producing a consistent structured view of resumes that were collected under different templates or conventions

When NOT to Use

• Ranking, scoring, comparison, or recommendation — ABSOLUTELY NOT. This is AEDT territory. Use a human reviewer. Any prompt asking for these is hard-refused.
• Single-resume review (1-4 resumes) — use a human reviewer directly; the high-volume framing that holds the AEDT wall breaks down at small N
• Fit assessment, "strong candidate" evaluation, culture fit, team fit — not this skill's job and not any skill's job under this governance pack
• Hiring recommendations, interview/reject suggestions, next-step advice — not emitted under any circumstances
• Executive-level searches (VP+, C-suite) where the recruiter owns the whole funnel manually — no throughput benefit and high review cost
• Referral candidates where a human recruiter already knows the context — skip the extraction, go straight to human review
• Any workflow where the deployer intends to auto-reject, auto-route, or pre-filter candidates before a human reads the raw resume — this flips AEDT classification

Required Inputs

./candidate-batch-2026-Q2/

US-NYC

US-IL

US-CO

US-CA

US-Federal

EU

UK

IL

acknowledgment_date

re_attestation_due_date

hr-ai-governance

hr-ai-governance

Jurisdiction is never defaulted. If the user does not supply it, the skill refuses to run. Jurisdiction determines applicable AEDT/ADMT rules, retention period, candidate-disclosure obligations, and which runtime warnings fire.

Deployer attestation is never bypassable. If the attestation is missing or lapsed, the skill refuses to run for that tenant and emits a re-attestation prompt. This is the AEDT deployer-warning enforcement mechanism (

hr-ai-governance

Batch size is a soft floor. For batches under 5 resumes, the skill warns and suggests human review instead. The high-volume framing (value comes from volume triage efficiency) breaks down at single-resume scale, and the AEDT wall is held in part by the volume framing.

Refusal Conditions (Hard, Non-Overridable)

The skill refuses — fully, structurally, with no override flag — under any of the following conditions:

1. Prompt asks for ranking, scoring, comparison, or recommendation. Response: "This skill is descriptive, not evaluative. Use a human reviewer for ranking, scoring, comparison, or recommendations. Per NYC Local Law 144, automated ranking and scoring of candidates is regulated as an AEDT — this skill is deliberately not classified as one. Any such workflow must go through a licensed employment attorney and (in NYC) a third-party bias audit."
2. Prompt asks for fit, suitability, seniority classification, strengths/weaknesses, or next-step recommendation. Same refusal, same routing message.
3. GINA proxy detected in a resume (genetic information, family medical history, BRCA, Huntington's, predisposition language, genetic test results). Per

   hr-ai-governance

   Section 3.2, GINA is strict liability — redaction is insufficient because the model has already seen the data. The skill refuses processing that resume, the input is not passed to any model, and the event is logged as a GINA-adjacent exposure event for @employment-counsel review.
4. Union membership or organizing activity detected in a resume ("shop steward," "union representative," "bargaining committee," "participated in organizing drive"). Per

   hr-ai-governance

   Section 3.2, NLRA Section 7 is federal strict protection — a single instance of AI-surfaced union language in an adverse employment decision is direct evidence of discrimination. The skill refuses processing that resume and routes to HR-only human handling.
5. No jurisdiction provided. Refuses to run with "jurisdiction is required, no default applies" message.
6. Deployer attestation missing, lapsed, or unrecognized tenant. Refuses with re-attestation prompt and pointer to

   hr-ai-governance

   Section 6.4.
7. Batch size < 5 resumes. Warns and routes to human review as the more appropriate path; does not emit extraction.
8. HITL reviewer identity cannot be authenticated by deployer auth system. Refuses per

   hr-ai-governance

   Section 4.2 rule 2 — no "unknown" or "system" reviewer values allowed, ever.

Refusals are logged. Refusal logging is critical because "the skill refused GINA exposure" is itself valuable audit evidence for @employment-counsel review and future third-party bias audits.

Method

Nine-step pipeline. Order matters because later steps depend on earlier ones. Every step produces audit-log input; the audit log record is the receipt the regulator sees.

Step 1 — Check deployer attestation currency

Read the deployer attestation file. Verify:

• deployer_id

  matches the current tenant

• acknowledgment_date

  is present and valid ISO 8601 UTC

• re_attestation_due_date

  =

  acknowledgment_date + 365 days

• Current date ≤

  re_attestation_due_date

If any check fails → REFUSE with re-attestation prompt (see Refusal Condition 6).

Step 2 — Check jurisdiction + load applicable rules

Read

--jurisdiction

hr-ai-governance

• Retention period for the audit log entry (Section 4.3)
• Candidate disclosure requirement (Section 9.1) — in EU, NYC, CA, and CO this is mandatory pre-use, not on-request
• Runtime warning block to emit (e.g., EU → Section 7.3 EU AI Act high-risk notice; NYC → Section 6.4 AEDT deployer warning; CO → Section 8.1 Colorado AI Act deployer notice)
• Ban-the-Box application (if any): if jurisdiction is in the 37+ US states + NYC Fair Chance Act, criminal-history redaction is pre-conditional-offer

If jurisdiction is not in the matrix → BLOCK and route to @employment-counsel per

hr-ai-governance

Step 3 — For each resume: run proxy scanner (inherit Section 3.2)

Scan every resume against the full proxy register from

hr-ai-governance

• 9 standard proxies (given name, family name, zip code, graduation year, photo, gender-coded language, date of birth, marital status, pronouns field)
• 27 non-obvious proxies (religious hobbies, IDF/military mentions, ROTC/VFW,

  .edu

  email, year-in-username email, address format, "native English" vs "fluent English," language inventory, military unit/deployment details, disability accommodations in cover letter, employment-gap explanations, visa text in cover letter, LinkedIn URL customizations, tenure-plus-founding-date combinations, headshot URLs, calendar/availability mentions, dietary preferences, salary history, "part-time/flexible/remote" preference, educational-institution ranking, "culture fit" references, criminal history / Ban the Box, caregiver status, credit history, union membership, genetic information)

For each detected proxy, record

proxies_detected

proxy_type

confidence

field_path

Step 4 — Apply proxy redaction / mitigation per pattern

Each detected proxy gets one of four treatments per

hr-ai-governance

<personal_email>

hr-ai-governance

Redacted and normalized values are removed from the content the model sees. Original values are retained only in the secured contact store for post-hire contact use, never in the analysis path, never in the audit log, never in the extraction output.

Step 5 — Extract structured fields per canonical schema

The model produces extraction output conforming to the canonical schema in the "Canonical Extraction Schema" section below. Only the allowed fields are extracted. Source-faithful language only: if the resume says "5 years of Python," the extraction says "5 years of Python," never "5 years of significant Python experience" or "5+ years of Python expertise." Bright line #8 is enforced at the prompt level and verified in Step 6.

Experience is emitted in reverse-chronological order (candidate's own ordering of their roles, not the skill's imposed ordering — Bright line #4). Skills are emitted as an unordered set (alphabetical, deterministic). Education is a list in whatever order the candidate listed it, most commonly reverse-chronological.

Step 6 — Run semantic-intent validator on every field

The validator is the enforcement mechanism for Bright lines #1, #2, #3, #5, #7, and #8. It runs over every field name + value pair in the extraction output and rejects any that is semantically equivalent to a banned field, even if literally renamed. Examples of what the validator catches:

• Field

  candidate_quality

  → REJECT (semantically equivalent to a score)
• Field

  top_skills_ordered

  → REJECT (ordered + "top" = Bright line #2)
• Field

  experience_level: "senior"

  → REJECT (seniority label, Bright line #7)
• Field

  years_of_python: "significant"

  → REJECT (qualitative modifier, Bright line #8)
• Field

  recommended_next_step: "interview"

  → REJECT (recommendation, Bright line #5)
• Free-text value containing "strong candidate for" → REJECT (fit language, Bright line #3)
• Field

  primary_skills

  → REJECT (implicit ranking, Bright line #2)

The validator's refusal of a bad field does NOT downgrade to a finding — it is a structural failure. The extraction output is rejected in full, a finding is emitted ("semantic intent validator rejected field

X

Step 7 — Emit candidate disclosure template per resume

Per

hr-ai-governance

Step 8 — Emit audit log entry per resume

Per

hr-ai-governance

• run_id

  ,

  timestamp

  ,

  skill_name: "resume-summarizer"

  ,

  skill_version: "1.0.0"

  ,

  governance_pack_version: "1.0.1"

• jurisdiction

• deployer_org

  (tenant identifier)

• inputs_hash

  — SHA-256 of the canonicalized resume content after redaction (NEVER raw PII)

• input_type: "resume"

• redactions_applied

  — full list from Step 4

• proxies_detected

  — full list from Step 3

• hitl_reviewer

  — authenticated identity (NEVER free text)

• hitl_decision

  ,

  hitl_timestamp

  ,

  hitl_rationale

  — populated at Step 9 HITL gate

• modifications_diff

  — structured diff if reviewer modified the extraction

• downstream_action

  — set post-HITL

• adverse_impact_check

  —

  cannot_audit

  by default for this skill (single-resume extraction does not support adverse impact testing; the counterfactual fixture run is the fairness gate, run on skill ship + every 180 days, not per resume)

• model_provenance

  — model name + version used for extraction

• candidate_disclosure_confirmed

  —

  true

  after the deployer confirms the candidate saw the disclosure per Section 9.1

• deployer_aiact_acknowledgment

  — the deployer attestation object (from Step 1) with

  deployer_id

  ,

  acknowledgment_date

  ,

  re_attestation_due_date

• signoff_path

  —

  standard

  unless fallback invoked (

  hr-ai-governance

  Section 11.3 — note:

  /resume-summarizer

  shipping and its AEDT memo are NON-DELEGABLE under Section 11.3; fallback path does not apply)

• hitl_review_duration_seconds

  — populated at Step 9; feeds the throughput tripwire

• retention_expiry

  — computed from jurisdiction per Section 4.3

Refusal events (GINA, union, attestation lapse, semantic-intent validator failure) are also logged, with

downstream_action: "refused-routed-to-human"

"refused-deployer-attestation-lapsed"

"refused-skill-defect"

Step 9 — Block downstream consumption pending HITL review

Per

hr-ai-governance

/resume-summarizer

1. Extraction output is placed in a "draft" state and is not available to the ATS, the hiring manager, or any downstream system
2. The HITL reviewer (recruiter) is prompted with: the raw resume (pre-redaction), the extracted structured output, a checklist of items to verify, and a mandatory rationale field on modify/reject
3. The reviewer must explicitly accept / modify / reject. No trailing "ok" button. No auto-accept.

4. hitl_review_duration_seconds

   is logged — the wall clock between prompt display and decision.
5. Throughput tripwire (

   hr-ai-governance

   Section 10.2.1): any decision under the 30-second threshold for

   /resume-summarizer

   is flagged in the quarterly review regardless of aggregate rate. Sub-10-second reviews are treated as presumptive AEDT-wall-risk failures and escalated to @chro immediately, because NYC DCWP may characterize sub-10s review as non-discretionary rubber-stamping, which flips the AEDT non-classification. Additional tripwires: same reviewer > 50 accepts/day with median <30s → flagged; reviewer identity with >30%

   cannot_audit

   accepts/month → flagged for @people-analyst review of sampling strategy.
6. Only after reviewer signs off does the extraction move from "draft" to "acted" state and become available downstream.

There is NO bypass. Deployer attempts to disable HITL → the skill refuses to run. Deployer attempts to programmatically auto-accept → the audit log captures the identical reviewer identity across records and the quarterly review flags the pattern as HITL hygiene failure.

Canonical Extraction Schema

The ONLY fields this skill emits in extraction output. Anything not in this list is banned by the semantic-intent validator.

{
  "candidate_id": "opaque hash (SHA-256 of canonicalized identifier; NEVER raw name/email)",
  "skills_set": ["unordered alphabetical list of skill strings from the resume"],
  "education": [
    {
      "institution": "string (name only, no ranking data)",
      "degree": "string (e.g., 'B.Sc.')",
      "field": "string (e.g., 'Computer Science')",
      "year": 2020
    }
  ],
  "experience": [
    {
      "title": "string (role title as stated)",
      "organization": "string (employer name)",
      "start_date": "ISO 8601 (YYYY-MM)",
      "end_date": "ISO 8601 (YYYY-MM) or 'present'",
      "raw_description": "string — source-faithful extraction of the role's bullet points, no adjectives added by the model"
    }
  ],
  "experience_years": 12,
  "certifications": ["unordered list of certification strings"],
  "languages": [
    {
      "language": "string",
      "self_reported_level": "string ONLY if candidate explicitly self-reported; never inferred"
    }
  ],
  "location": {
    "city": "string",
    "country": "string"
  },
  "work_authorization": "string ONLY if candidate explicitly stated; never inferred",
  "proxies_redacted": ["count and category only — e.g., ['given_name', 'family_name', 'zip_code', 'military_service_country']"]
}

Notes:

• experience_years

  is a raw integer (Bright line #7). NEVER a label like "senior," "mid," "entry," "junior," "staff," or "principal."

• skills_set

  is unordered (Bright line #2). Alphabetical sort for deterministic output.

• experience

  is reverse-chronological — this is the candidate's own ordering of their career history, not an imposed importance ranking, so it does not violate Bright line #4.

• work_authorization

  is NEVER inferred. If the candidate did not state it, the field is

  null

  .

• location

  is city + country only. Full address, street, neighborhood, zip code — all redacted per proxy register.

• proxies_redacted

  records categories and counts, NOT content. The content never leaves the secured contact store.

Banned Output Fields

The semantic-intent validator rejects any field semantically equivalent to the following (literal list from

hr-ai-governance

score
fit_score
fit_rating
match_score
match_percentage
candidate_score
ranking
rank
tier
category
level
seniority
seniority_label
experience_level
strength
strengths
weakness
weaknesses
recommendation
recommended_action
suggested_action
next_step
interview_decision
suitable_for
best_match_role
top_skills
key_skills
primary_skills
core_skills
summary
evaluation
assessment
verdict
notes
flags
concerns
highlights
culture_fit
team_fit
values_fit
ranked_candidates
candidate_comparison
qualitative_evaluation
overall_quality
candidate_quality

Renaming (

candidateScore

fit-score

top_skills_ordered

Output Structure

Every batch produces:

1. Disclaimer + UPL guardrail block (top) — see top of file, reproduced on every run

2. AEDT non-classification memo block — see above, reproduced on every run

3. Positioning statement — literal verbatim sentence

4. Batch metadata block

• Batch ID, timestamp, jurisdiction, tenant, resume count, deployer attestation status (current / lapsed / missing), re-attestation due date

5. Per-resume extraction — one canonical JSON block per resume, plus its candidate disclosure template + its audit log entry

6. Batch-level summary

• Summary counts only, no ranking: successfully-extracted count, refused count (with refusal reasons — GINA / union / semantic-intent / deployer-attestation), aggregate

  proxies_redacted

  tally across the batch (categories only, no content)
• HITL-review-pending flag: every extraction is in "draft" state until reviewer signs off

7.

## Findings

section — numbered, per batch

Examples: "Finding 1: GINA proxy detected in resume #7 — routed to human review, resume not processed." "Finding 2: Semantic-intent validator rejected field

candidate_quality

8.

## Reviewer Checklist

— nine items

9.

## Cannot Assess Without

— what this skill deliberately does NOT opine on

10.

## Candidate Disclosure Template

— the verbatim block the deployer's ATS surfaces to candidates

11.

## Quality Gates

— 12 self-check items, pass/fail per batch

12.

## Annual Review Cadence

— review date, triggers

13. ROI block (per

roi-display.md

)

Quality Gates (12 checks, per batch)

The skill runs a 12-item self-check BEFORE emitting output. If any check fails, the output does not publish — it produces a structural finding instead. More checks than other HR skills because this is first-of-type and carries the AEDT wall.

1. Jurisdiction declared and matches a row in

   hr-ai-governance

   Section 8.1
2. Deployer attestation current (≤

   re_attestation_due_date

   )
3. AEDT memo visible in the output header (not just referenced by link)
4. Positioning statement literal — the verbatim sentence appears at the top of every run
5. Zero banned fields in any extraction (semantic-intent validator pass, all 45+ banned terms checked)
6. Proxy scanner ran on every resume in the batch (no skipped resumes, no partial runs)
7. GINA refusal applied where triggered (strict liability — no exceptions)
8. Union refusal applied where triggered (NLRA Section 7 strict protection)
9. HITL gate applied — every extraction is in "draft" state, no bypass
10. Audit log entry emitted per resume — including refused resumes (they get refusal log entries)
11. Candidate disclosure template emitted per successfully-extracted resume
12. Counterfactual harness passed on skill ship — tolerance per

    hr-ai-governance

    Section 5.3 field-type spec (structured enumerable = byte-identical; free-text summary = semantic-equivalent with Levenshtein < 0.15 after lowercase + stopword removal + whitespace collapse + punctuation strip). Harness re-runs on every model change, prompt revision, and every 180 days.

No "express mode." The 12 checks run on every invocation.

Reviewer Checklist (9 items — before HITL gate release)

• Jurisdiction confirmed against candidate residency and role location
• Deployer attestation currency confirmed (

  acknowledgment_date

  + annual re-attestation)
• For every resume in the batch: raw resume read and compared against the extraction for accuracy
• No banned fields in any extraction (semantic-intent validator pass)
• GINA and union refusals reviewed and routed to HR-only paths
• Candidate disclosure template reviewed and confirmed suitable for ATS surfacing
• Adverse impact implications understood (single-resume runs do not support adverse impact testing; the counterfactual fixture run is the fairness gate)
• Did you spend less than 30 seconds on this review? If so, this is presumptive auto-acceptance and the AEDT wall is at risk. Sub-10-second reviews are escalated to @chro per

  hr-ai-governance

  Section 10.2.1. Stop, re-read the raw resume, compare it to the extraction, and exercise actual discretion. HITL is non-discretionary review if you can't plausibly have read both in the time you took.
• Employment counsel engaged for any "Cannot Assess Without" item that applies

Cannot Assess Without Licensed Counsel or Specialist

• Human evaluation of any kind — by design. This skill is descriptive, not evaluative. Ranking, scoring, comparison, fit, strengths/weaknesses, recommendation, next-step advice all require human reviewers.
• Actual hiring context — hiring manager input, team composition, budget, timeline, backfill vs new hire context, all out of scope.
• Reference checks, interview performance, assessment results — not in the resume, not in scope.
• Pay equity implications for the candidate pool — use

  /comp-benchmark

  and its adverse impact check.
• Team composition and culture — out of scope by design. Culture fit is not a valid criterion under this governance pack.
• Individual capability differentiation between two candidates with similar resumes — hiring manager judgment, out of scope.
• Seniority classification ("is this person senior?") — explicitly banned by Bright line #7.
• Any ranking, scoring, or comparison between candidates — ABSOLUTELY NOT. AEDT territory. Routes to human review, third-party bias audit, and employment counsel.
• Ban-the-Box specific jurisdictional interpretations — the skill applies a pre-conditional-offer redaction in the 37+ Ban-the-Box jurisdictions, but the specific conditional-offer timing and exception handling require employment counsel in the target jurisdiction.
• GINA exposure forensics — when the skill refuses a resume on a detected GINA proxy, the downstream forensic review (how did the proxy get into the input, what's the exposure, what's the remediation) is @employment-counsel's job, not this skill's.

Candidate Disclosure Template

Per

hr-ai-governance

candidate_disclosure_confirmed = true

NOTICE: AI-Assisted Structured Extraction

Your application may be processed using an AI-assisted structured
extraction tool that helps our hiring team review applications at scale.

WHAT THIS TOOL DOES:
- Extracts structured information from your resume: skills, education,
  experience, certifications, location, languages, and work authorization
  where you explicitly stated it.
- Redacts protected-class proxies (names, photos, date of birth, etc.)
  before a human reviewer sees the extraction.

WHAT THIS TOOL DOES NOT DO:
- It does NOT score your application.
- It does NOT rank you against other candidates.
- It does NOT evaluate your fit for the role.
- It does NOT recommend whether to interview, advance, or reject you.
- It is deliberately not classified as an Automated Employment Decision
  Tool under NYC Local Law 144.

HUMAN REVIEW:
- A human recruiter reviews every application before any decision is made.
- The recruiter sees both your raw resume and the extraction output.
- Any hiring decision is made by human reviewers based on their own
  discretionary judgment, not by this tool.

YOUR RIGHTS:
- You can request a summary of the extraction output produced from your
  resume (contact: [deployer HR address]).
- You can request correction of any extracted field that is inaccurate.
- You can request a human-only review path that bypasses the extraction
  tool (contact: [deployer HR address]).
- In EU / UK / California / Colorado, you have additional rights under
  GDPR / UK GDPR / CCPA / CPRA / Colorado AI Act. Contact the deployer's
  Data Protection Officer.

Date: {application_date}
Tool version: resume-summarizer v1.0.0
Governance pack: hr-ai-governance v1.0.1

The deployer is accountable for actually showing this block. The skill emits it; the ATS surfaces it; the candidate sees it; the deployer logs

candidate_disclosure_confirmed = true

candidate_disclosure_confirmed = false

Related Skills and Hand-off

• Upstream: a separate "deployer attestation" flow is a future skill (TBD). For v1.0.0, deployer attestation is a file-based input; the flow to collect, sign, and refresh the attestation is out of scope for this skill.
• Downstream: the human reviewer. NOT another skill. This is the whole point — the extraction hands off to a human, not to a scorer or ranker. There is no "feed into the interview decision" pipeline. The interview decision is a separate human process.
• Governance pack:

  hr-ai-governance

  v1.0.1 — proxy register, audit schema, AEDT memo, counterfactual fixture, HITL pattern, jurisdiction matrix, all inherited.
• Counterfactual harness:

  hr-ai-governance

  Section 5 specifies the 200-fixture counterfactual test. The fixture files live at

  Extension Teams/hr-team/reference/test-fixtures/

  (scheduled engineering task per Section 5.1). This skill references the harness; it does not author or own the harness. Harness runs: skill ship, every 180 days, every model change, every prompt revision.
• Sibling HR skills (same pack):

  /job-description-generator

  (v1.0.0, shipped 2026-04-11),

  /interview-guide

  (Phase 4A),

  /comp-benchmark

  (Phase 4A). None of them score, rank, or compare candidates either — same governance pack, same AEDT wall, same disciplines.

Annual Review Cadence

Per

hr-ai-governance

• AEDT non-classification memo (Section 6): 12 months from publication → next review 2027-04-11
• Counterfactual fairness fixture set (Section 5): 180 days → next refresh 2026-10-08
• Jurisdiction matrix (Section 8): quarterly → next review 2026-07-11 (end of Q2 2026)
• Proxy register (Section 3): quarterly → next review 2026-07-11
• Deployer attestation per tenant: 365 days from

  acknowledgment_date

  per each tenant individually
• Full skill review: 12 months → next review 2027-04-11

Out-of-cycle triggers (immediate review, do not wait for calendar):

• NYC DCWP enforcement guidance update on Local Law 144
• CA CPRA ADMT final rules landing (expected 2027)
• Colorado AI Act implementing guidance from the Colorado AG
• EU AI Act harmonized standards or implementing acts
• Any case law flagged by @employment-counsel touching AEDT classification, automated resume processing, or adverse impact from AI-assisted hiring
• Any adverse impact or counterfactual fairness failure in production
• Any bright line violation in any released version (immediate pull)

Owner: @chro + @employment-counsel jointly.

Pack Inheritance

This skill inherits the following from

hr-ai-governance

redact

normalize

flag

block

block

redact

normalize

candidate_disclosure_confirmed

deployer_aiact_acknowledgment

signoff_path

hitl_review_duration_seconds

candidate_disclosure_confirmed = true

Delegation Patterns Available

Default: Pattern 1 Consultation

Consultations are attributed inline in the findings. Ownership of the extraction output stays with Recruiter (primary) and People Analyst (co-owner for proxy/harness/audit-log layers).

Pattern 5 Adversarial Review

Not applicable at v1.0.0. Adversarial Review (

delegation-protocol.md

The pack itself (

hr-ai-governance

If a specific resume batch triggers novel risk (e.g., a batch from a regulated sector with unusual protected-class dimensions, or a first deployment in a new jurisdiction), the recruiter can manually escalate to

@employment-counsel

ROI Framing

ROI for

/resume-summarizer

HR blended rate: $150/hr per

feedback_roi_rates.md

Time-saved baseline for a batch of 25 resumes:

• Manual structured extraction of 25 resumes at ~6 minutes per resume = 2.5 hours
• Proxy redaction review at ~2 minutes per resume = 0.8 hours
• Structured schema normalization across heterogeneous formats = 0.5 hours
• Audit log entry construction per resume at ~1 minute = 0.4 hours
• Total manual baseline: ~4.2 hours per batch of 25

Complexity multipliers:

• Simple (homogeneous resume format, same jurisdiction, no GINA/union flags): 0.5× → ~2.1 hrs saved
• Standard (mixed format, single jurisdiction, occasional flags): 1.0× → ~4.2 hrs saved
• Complex (multi-jurisdiction batch, heterogeneous format, multiple flags, counterfactual re-verification): 1.5× → ~6.3 hrs saved

Example ROI line for a standard 25-resume batch:

⏱️ ~4.2 hrs saved on drafting and triage in 90s, 45k tkns ~$2.7 cost, Value ~$630

The ROI tracks ONLY the time the skill saves on drafting and triaging the structured artifact, NOT the substantive human review time. The human reviewer still reads every resume, compares it to the extraction, exercises discretion, and owns the outcome.

Attribution and Maintenance

Owner: 🎯 Recruiter. The skill's extraction logic, canonical schema, banned fields enforcement, AEDT wall, refusal conditions, and HITL gate integration are Recruiter's accountability.

Co-owner: 📊 People Analyst. Joint accountability on the proxy register scanner, the counterfactual harness invocation, and the audit log emission layer. Any change to the proxy scanning logic, the counterfactual tolerance spec, or the audit log schema integration requires sign-off from both owners.

Consumers:

ext-hr

/resume-summarizer

Authoring: First-principles. This skill was authored from scratch during Phase 4A as the fourth and final HR skill under the

hr-ai-governance

Dependency on the pack: The skill reads from

hr-ai-governance

Updates: Via the two-pass publication gate defined in

sensitive-skill-guardrails.md

• Pass 1 scaffolding check: 📋 Director of HR, 15 minutes, binary GO / REWORK.
• Pass 2 substantive check: 👔 Employment Counsel, 5 business days first-of-type SLA for v1.0.0 (the skill and the pack carry first-of-type status together for HR AI governance). Subsequent skill revisions drop to 72-hour subsequent-similar SLA, EXCEPT any change to the AEDT non-classification memo (Section 6), the bright lines, the banned field list, the refusal conditions, or the HITL throughput tripwire — those revert to first-of-type 5-business-day SLA regardless of skill version.
• Non-delegable: The AEDT non-classification memo is NON-DELEGABLE per

  hr-ai-governance

  Section 11.3. @compliance-officer fallback path does NOT apply.

Minor edits (typos, formatting, documentation links) can bypass Pass 2. Any edit touching: the 8 bright lines, the banned field list, the 6 refusal conditions, the HITL gate, the counterfactual harness tolerance, the candidate disclosure template, the deployer attestation logic, or the AEDT memo language — requires a full first-of-type Pass 2 substantive review by Employment Counsel.

Changelog: Maintained at the bottom of this file.

Example Invocation (standard path)

User: /resume-summarizer --batch ./candidate-batch-2026-Q2/ --jurisdiction US-NYC --role "Senior Backend Engineer" --deployer-attestation ./attestations/legionis-2026.json

/resume-summarizer v1.0.0 — loading:
  - Batch: 25 resumes from ./candidate-batch-2026-Q2/
  - Jurisdiction: US-NYC (AEDT deployer warning active; 1-year retention)
  - Deployer attestation: legionis-2026.json
    - acknowledgment_date: 2026-01-15
    - re_attestation_due_date: 2027-01-15
    - STATUS: current (as of 2026-04-11)
  - Governance pack: hr-ai-governance v1.0.1
  - Scanner source: Section 3.2 (9 standard + 27 non-obvious proxies)
  - Counterfactual harness: last run 2026-04-01 (PASS; next 2026-10-08)
  - Role context: "Senior Backend Engineer" (descriptive only; not used for evaluation)

Running 9-step pipeline:
  Step 1: Deployer attestation: CURRENT (valid until 2027-01-15)
  Step 2: US-NYC rules loaded (AEDT deployer warning, 1-yr retention, Ban-the-Box pre-conditional-offer redaction active per NYC Fair Chance Act)
  Step 3: Proxy scanner ran on 25 resumes; 187 proxies detected total
  Step 4: Redaction/mitigation applied:
    - redact: 42 (given_name, family_name, zip, photo, DOB, salary_history, caregiver_lang, criminal_history)
    - normalize: 31 (military_service, email_domain, year_in_username, language_phrasing, address_format)
    - flag_for_human: 6 (culture-fit references in references section)
    - block (GINA): 1 resume → REFUSED, routed to HR-only human review
    - block (union): 1 resume → REFUSED, routed to HR-only human handling per NLRA Sec 7
  Step 5: Canonical extraction on 23 remaining resumes
  Step 6: Semantic-intent validator pass
    - 22 extractions: clean
    - 1 extraction: field `skills_primary` rejected; retry with stricter prompt → clean
  Step 7: Candidate disclosure template emitted for 23 extractions
  Step 8: Audit log entries emitted for 25 (23 extractions + 2 refusals)
  Step 9: All 23 extractions in DRAFT state, HITL gate blocking downstream

12/12 quality gates passed.

HITL review required before downstream consumption.
Reviewer will see raw resume + extraction side-by-side with mandatory rationale field.
Sub-30s reviews will be flagged; sub-10s reviews will be escalated to @chro.

Producing batch output at:
  ./candidate-batch-2026-Q2/_resume-summarizer-output-2026-04-11.md

⏱️ ~4.2 hrs saved on drafting and triage in 90s, 45k tkns ~$2.7 cost, Value ~$630

Changelog

• 1.0.0 (2026-04-11) — Initial authoring. First-of-type HR skill under

  hr-ai-governance

  pack for full substantive review. Fourth and final Phase 4A HR skill. Authored from first principles by 🎯 Recruiter (primary) and 📊 People Analyst (co-owner for proxy register, counterfactual harness, and audit log emission layers). No vendor tool lifts. Academic citations: Schmidt & Hunter (1998) for volume-extraction value framing; Bertrand & Mullainathan (2004) for proxy-bias empirical grounding. AEDT non-classification memo embedded verbatim from

  hr-ai-governance

  Section 6.2. Banned field list extends

  hr-ai-governance

  Section 6.3 with Phase 4A additions (culture_fit, team_fit, values_fit, qualitative_evaluation, overall_quality, candidate_quality, seniority_label, experience_level). Counterfactual harness tolerance per Section 5.3 F7 field-type spec. HITL throughput tripwire per Section 10.2.1. Birth test at

  Legionis/Product/resume-summarizer-birth-test-2026-04-11.md

  exercises GINA refusal, union refusal, IDF proxy normalization (no-country rule), Ban-the-Box criminal-history redaction, and semantic-intent validator catch of a planted

  top_skills_ordered

  bad field. Ready for @employment-counsel 5-business-day first-of-type substantive review. Scaffolding pass by 📋 Director of HR required before substantive review.