ceo-orchestration 1.0.0

package/.claude/skills/domains/hospitality/skills/guest-services/SKILL.md

@@ -0,0 +1,417 @@
+---
+name: guest-services
+description: |
+  Operating doctrine for hospitality guest services across hotel, vacation rental,
+  and restaurant operations — covering check-in and check-out flow, complaint
+  resolution, special-request fulfillment, multi-channel communication, online
+  review management, and service recovery. Regional norms are encoded for the
+  Middle East, Europe, Asia-Pacific, and the Americas. Use when executing
+  pre-arrival messaging sequences, triaging and resolving in-stay complaints,
+  determining compensation tier for service failures, drafting review responses
+  on Google, TripAdvisor, Booking, Airbnb, or Yelp, or when evaluating whether
+  a guest request falls within commitment range or best-effort range.
+owner: Sofia Carvalho (Guest Services Manager, domain persona)
+tier: domain:hospitality
+scope_tags: [hospitality, guest-services, complaint-resolution, online-reputation, multi-channel-comms, service-recovery]
+inspired_by:
+  - source: msitarzewski/agency-agents/specialized/hospitality-guest-services.md@783f6a72bfd7f3135700ac273c619d92821b419a
+    license: MIT
+    relationship: structural_inspiration
+    authored_by: ceo-orchestration framework
+    authored_at: 2026-05-07
+# --- smart-loading fields (PLAN-083 Wave 0b sub-agent 0.7c) ---
+domain: hospitality
+priority: 8
+risk_class: low
+stack: []
+context_budget_tokens: 400
+inactive_but_retained: true
+repo_profile_binding:
+  frontend: {active: false, priority: 10}
+  engine: {active: false, priority: 10}
+  fintech: {active: false, priority: 10}
+  trading-readonly: {active: false, priority: 10}
+  generic: {active: false, priority: 10}
+activation_triggers: []
+# --- K1 paths: native file-touch activation (PLAN-135 W3 unit k1a) ---
+paths:
+  - "**/guests/**"
+  - "**/reservations/**"
+  - "**/check-in/**"
+  - "**/reviews/**"
+---
+
+# Guest Services
+
+Hospitality operations span a narrow window — from the moment a reservation is
+confirmed to the moment a post-stay review is published — in which every touchpoint
+either compounds trust or erodes it. This skill is the operating doctrine for that
+arc: check-in discipline, complaint resolution sequencing, request fulfillment
+boundaries, channel hygiene, review response protocol, service recovery
+compensation, and cultural calibration across four major regions.
+
+The skill is property-type agnostic. It applies to full-service hotels, boutique
+properties, vacation rentals, and restaurant operations. Scale and brand affiliation
+do not change the underlying principles — only the thresholds and escalation paths
+differ.
+
+## Cardinal Rule
+
+A guest's complaint is the lowest-cost moment of brand loyalty acquisition the
+hotel will ever buy; treating it as a cost is treating loyalty as a cost. The
+complaint frame governs every section of this skill: the goal of complaint
+resolution is not damage containment — it is trust reconstruction, which, when
+executed correctly, produces higher retention than an uneventful stay.
+
+## Fail-Fast Rule
+
+Stop and escalate to the duty manager before proceeding if any of the following
+conditions are true:
+
+- The guest discloses a safety incident — injury, illness, security concern, or
+  evacuation scenario. Safety takes precedence over all service workflows.
+- The complaint involves an allegation of staff misconduct or discriminatory
+  treatment. Treat as an HR and legal matter immediately; do not attempt service
+  recovery without management involvement.
+- A guest requests compensation that exceeds the local authority threshold (defined
+  per property) without manager approval.
+- A third party — not the reservation holder — demands access to room assignment,
+  folio, or stay details. Privacy breach risk; do not disclose.
+- A review response would require revealing PII (room number, length of stay,
+  billing dispute detail) to the public. Draft and hold for manager approval instead.
+
+Fail-fast does not mean refuse assistance — it means halt the current workflow,
+document the trigger condition, and re-enter only after the condition is resolved
+and the appropriate authority has acknowledged.
+
+## When to Apply
+
+Apply this skill when:
+
+- Executing the pre-arrival messaging sequence (48-hour and 24-hour touchpoints).
+- Managing check-in protocol including loyalty recognition, room assignment, and
+  special-request confirmation.
+- Receiving, triaging, and resolving an in-stay complaint within the 4-hour first
+  response window.
+- Fulfilling or declining a special request and communicating the distinction
+  between commitment and best-effort.
+- Responding to a post-stay online review on any platform.
+- Determining compensation tier for a service failure under the recovery matrix.
+- Adapting communication style or protocol to a regional guest segment (ME, EU,
+  APAC, Americas).
+- Handling a food allergy or dietary restriction declaration at any point in the
+  guest journey.
+
+Do not apply this skill for revenue management decisions, rate-setting, inventory
+allocation, OTA commission negotiation, or HR performance management — those are
+out of scope.
+
+## Check-In and Check-Out Flow
+
+### Pre-Stay Messaging
+
+Send two touchpoints per reservation: a 48-hour message and a 24-hour confirmation.
+The 48-hour message confirms arrival details, surfaces online check-in or digital
+key enrollment, and invites the guest to update special requests. The 24-hour
+message is brief: a reminder of check-in window and a single-sentence invitation
+to contact the team.
+
+Neither message is a marketing vehicle. Do not upsell packages or promote F&B
+outlets in pre-arrival messaging unless the guest has opted into promotional
+communications. The goal is to reduce day-of friction, not to generate ancillary
+revenue.
+
+### Arrival Window
+
+The standard check-in window begins at the posted time. Early check-in is
+available only when a room has been inspected and cleared. Never promise early
+check-in as a given — communicate it as a confirmed outcome once the room is
+ready, or as a best-effort if the room is still being prepared.
+
+When the room is not ready at the guest's arrival, the handoff sequence is:
+
+1. Acknowledge the wait without minimising it.
+2. State the estimated ready time as a range, not a point — "between 2:00 and
+   2:30" is more accurate and more trusted than "2:00."
+3. Offer a concrete alternative: luggage storage plus access to a lounge, F&B
+   outlet, or common space.
+4. Initiate a proactive contact — by phone or SMS — the moment the room clears.
+   Never make the guest return to the desk to ask again.
+
+### Room-Ready Protocol
+
+Before releasing a room to the front desk for assignment, the room-ready checklist
+must be complete: housekeeping sign-off, maintenance clearance for any flagged
+items, special-request items placed (cot, allergen-free bedding, accessibility
+equipment), and loyalty amenity prepared if applicable. A room released without
+complete sign-off is not ready.
+
+### Late Checkout Policy
+
+Late checkout is available based on occupancy. Communicate the policy at check-in
+as a proactive offer, not only when the guest asks. The response sequence:
+
+- If late checkout is available at no charge for the loyalty tier: confirm it
+  immediately and note it on the folio.
+- If late checkout is available at a charge: quote the rate and give the guest
+  time to decide — do not pressure.
+- If late checkout is not available: offer the next best alternative (luggage
+  storage, access to lounge or gym until departure).
+
+Never improvise a late checkout policy outside the property's defined matrix.
+Inconsistent application generates negative reviews and loyalty member complaints.
+
+## Complaint Resolution Frame
+
+The complaint resolution sequence is: acknowledge → empathise → fact-find →
+action → follow-up → record. Each step is mandatory. Skipping fact-finding in
+favor of an immediate remedy risks resolving the wrong problem. Skipping follow-up
+after a resolution is the most common cause of a resolved complaint becoming a
+negative review.
+
+**Timing standard:** first response within 4 hours of complaint receipt, regardless
+of channel. For in-person or phone complaints, first response is immediate. For
+complaints submitted via OTA messaging or property app, the 4-hour window applies
+from the timestamp of receipt, not from when the next shift begins.
+
+**Language standard:** never use "policy doesn't allow" as a closing response.
+Policy is an internal constraint; the guest does not experience policy as a
+rationale — they experience it as refusal. If a request falls outside policy,
+describe what is possible, not what is blocked.
+
+**Documentation:** every complaint must be recorded regardless of how it was
+resolved. The record includes: guest identifier, room or table, nature of
+complaint, time reported, time resolved, action taken, compensation provided, and
+follow-up status. Undocumented complaints are invisible to operations management
+and cannot be used to identify systemic issues.
+
+## Special-Request Fulfillment
+
+### Request Triage
+
+Classify every special request at intake as one of three types:
+
+- **Operational standard:** requests the property fulfills as a matter of routine
+  (extra pillows, specific floor, away from elevator). Confirm as committed.
+- **Best-effort:** requests subject to availability or third-party dependency
+  (specific room number, early check-in, adjoining rooms). Communicate explicitly
+  as best-effort, not as a promise.
+- **Out of scope:** requests the property cannot fulfill — either because they
+  require resources not available on-property or because they raise a compliance
+  or safety concern. Decline clearly and offer the nearest alternative.
+
+### Capability Gate
+
+Before committing to a special request, verify the capability exists in the current
+inventory. A commitment that cannot be honored at check-in is worse than a
+best-effort disclosure at booking. The gate question is: "Can we confirm this
+regardless of occupancy and staffing conditions on the arrival date?" If the
+answer is no, it is best-effort.
+
+### Commitment vs. Best-Effort
+
+Communicate the distinction explicitly at the time of the request, not at check-in.
+The framing: "We have noted your preference for a high-floor room. We cannot
+guarantee it in advance, but we will assign the best available match on your
+arrival date." This sets accurate expectations and eliminates the gap between what
+the guest heard and what the property committed to.
+
+Never overpromise. An overpromise that fails at check-in produces a complaint
+that requires service recovery — consuming more staff time and compensation cost
+than the original honest disclosure would have.
+
+## Multi-Channel Communication
+
+### Channel Discipline
+
+Guests communicate through phone, SMS, WhatsApp, OTA messaging platforms (Booking,
+Expedia, Airbnb), and property apps. Channel selection should follow guest
+preference, not operational convenience. If a guest initiates on WhatsApp, respond
+on WhatsApp — do not redirect to a property-preferred channel unless a legitimate
+reason exists (e.g., a resolution requires a signed document).
+
+### Channel-Specific Protocols
+
+**Phone:** the highest-urgency channel. Answer within three rings during business
+hours. After-hours calls route to a staffed overnight desk, not to voicemail.
+Complaints communicated by phone require a verbal resolution and a written
+follow-up via the guest's preferred messaging channel within 2 hours.
+
+**SMS and WhatsApp:** acceptable for confirmation, status updates, and non-urgent
+requests. Not acceptable for complaint resolution of more than minor severity —
+escalate to phone or in-person for moderate and major complaints. PII transmitted
+via SMS or WhatsApp must be limited to what is necessary for the interaction; do
+not send folio details, room numbers, or passport-related information via
+unencrypted messaging channels.
+
+**OTA messaging:** responses on OTA platforms are visible to the platform and may
+influence platform ranking. Respond within the platform's response-time window —
+Booking.com and Airbnb publish response-rate metrics that affect visibility. Never
+move a complaint off an OTA platform by telling the guest to "contact us directly"
+as the first step; handle it within the platform first, then offer a direct channel
+for follow-up.
+
+**Property app:** highest auditability. All messages are logged within the PMS.
+Preferred channel for billing clarifications and documented special requests.
+
+### PII Handling Per Channel
+
+Guest PII — including room assignment, stay dates, folio details, and identity
+documents — must not be transmitted via public or unencrypted channels. Phone and
+property app are the acceptable channels for PII transmission. SMS and WhatsApp
+are acceptable for non-sensitive confirmations only. OTA messaging platforms have
+their own data handling terms; do not send payment data or government ID information
+through OTA messaging interfaces.
+
+## Online Review Management
+
+### Response Cadence by Platform
+
+| Platform | Response Window | Notes |
+|---|---|---|
+| Google | 48 hours | Indexed by search; high visibility |
+| TripAdvisor | 48 hours | Management response shown alongside review |
+| Booking.com | 24 hours | Affects Booking.com score visibility |
+| Airbnb | 14 days | Host response period before review becomes final |
+| Yelp | 72 hours | Community visibility; business response is public |
+
+These are targets, not minimums. Faster is better. Reviews that receive no
+response within the window are perceived as dismissal by future readers.
+
+### Response Principles
+
+Respond to every review — positive and negative. A positive review with no
+response is a missed trust signal. A negative review with no response is a
+publicly visible abandonment.
+
+For negative reviews, the response structure is: acknowledge the specific
+experience described → express genuine regret without a generic apology phrase →
+state one concrete corrective action that was taken or is being taken → invite
+direct contact. The response must not:
+
+- Argue with the guest's account of events, even if the account is factually
+  incomplete or inaccurate.
+- Reveal any PII in the public response (room number, stay dates, billing
+  figures, identity of other staff).
+- Offer compensation or discounts publicly — this invites gaming and establishes
+  precedent that is visible to all future reviewers.
+- Reference internal operational details (staffing levels, renovation schedules,
+  third-party vendor issues) in a way that shifts responsibility externally.
+
+**Defamation vs. criticism:** criticism of service quality, however harsh, is
+the guest's right and does not constitute defamation. Defamation requires
+demonstrably false statements of fact causing reputational harm. Do not escalate
+to legal review based solely on a negative review. Escalate only if the review
+contains provably false factual claims (dates, names, events) that the property
+can document as false. Even then, the recommended first step is direct outreach
+to the guest, not a legal response.
+
+## Service Recovery Protocol
+
+### Compensation Tier Matrix
+
+| Severity | Definition | Standard Recovery |
+|---|---|---|
+| Minor | Single-item service gap with no lasting impact (housekeeping timing, amenity missing at arrival) | Sincere apology + small amenity delivery or loyalty points equivalent to one ancillary transaction |
+| Moderate | Multi-element failure or a failure that affected the guest's plans (noise, billing error, missed special request) | Apology + room amenity + loyalty points or discount on current stay |
+| Major | Failure that materially degraded the stay (maintenance issue lasting multiple hours, F&B illness, room not ready for extended period) | Apology + significant compensation (partial or full night comp, or equivalent points) + manager follow-up call |
+| Severe | Failure resulting in guest relocation, safety incident response, or demonstrable harm | Apology + comp night or full stay + general manager personal contact + written acknowledgment |
+
+The matrix defines ranges, not fixed amounts. The duty manager approves all
+Major and Severe compensation. Front-desk authority covers Minor and Moderate
+tiers within the pre-approved property limits.
+
+### Written vs. Goodwill
+
+Written compensation (folio credits, confirmed discount codes, loyalty point
+additions) creates an auditable record and a clear guest expectation. Goodwill
+gestures (amenity delivery, upgraded F&B, room change) do not generate a paper
+trail but can be more immediately effective for emotional recovery.
+
+Use written compensation when the failure has a quantifiable financial impact on
+the guest or when the guest has asked for it. Use goodwill gestures first for
+in-stay complaints where speed of response is more important than the size of
+the gesture. The two are not mutually exclusive.
+
+### Legal Liability Awareness
+
+Service recovery is not a legal settlement. Compensation offered in service
+recovery does not constitute an admission of liability for incidents involving
+personal injury, property damage, or illness. For incidents that may generate
+a legal claim, notify the duty manager and the property's insurance contact
+before offering compensation, and document the incident timeline precisely. Do
+not offer written compensation for personal injury incidents without management
+and legal clearance.
+
+## Cultural Adaptation
+
+### Regional Norms
+
+**Middle East:** hospitality in the Gulf region carries strong social weight — a
+guest's dignity and face are implicated in every interaction, particularly in
+complaint handling. Complaints should be received with full attention and never
+minimized. Direct refusal (e.g., "we cannot do that") is perceived as a higher
+affront than in other regions; frame limitations as "what we can arrange" rather
+than "what is not possible." Honorific usage (Sheikh, Dr., Eng.) matters for
+business travelers; verify before first interaction when possible.
+
+**Europe:** service distance norms vary significantly across the continent.
+Northern European guests (Scandinavian, Germanic) tend to prefer efficient,
+low-affect interactions; extended warmth can read as performative. Southern
+European and Mediterranean guests generally respond well to more relational
+communication. Privacy consciousness is high across the EU; be precise about
+what data is collected and why. GDPR-adjacent practices (consent for marketing,
+right to access folio data) should be handled correctly even when not legally
+required in the property's jurisdiction.
+
+**Asia-Pacific:** honorific layers are structurally significant in Japan, Korea,
+and parts of China. Guest title and last name should be used unless the guest
+explicitly signals informality. Group harmony considerations mean that some
+guests will not complain directly even when dissatisfied; proactive mid-stay
+check-in calls or written channels may surface dissatisfaction that in-person
+inquiry would not. In markets with strong gifting norms (Japan, China), a
+recovery gesture that can be carried home — rather than a discount — may carry
+higher symbolic value.
+
+**Americas:** communication norms are generally less formal, and first-name use
+is more readily accepted. Guest expectations for proactive problem-solving are
+high — stating a problem and waiting for the guest to follow up is perceived as
+poor service. In North America, review culture is mature and review incentives
+are legally restricted (FTC guidance in the US); never offer compensation in
+exchange for a positive review. In Latin American markets, relationship-building
+over multiple interactions is valued; a brief personal follow-up on day two of a
+multi-night stay is expected at upper-tier properties.
+
+Never assume regional norms apply to an individual guest. Calibrate based on
+observed interaction cues, not on nationality alone.
+
+## Anti-Patterns
+
+| Anti-pattern | Why It Fails |
+|---|---|
+| Blaming the guest for the service failure | Converts a complaint into a confrontation; violates the cardinal rule and produces negative reviews at a higher rate than any other response |
+| Scripted or generic apologies ("sorry for any inconvenience") | Signals that the complaint has not been read; guests perceive it as a template and escalate instead of accepting the resolution |
+| Ignoring online reviews or responding only to positive ones | Creates a visible asymmetry that future guests read as defensiveness; platforms penalize non-response with lower ranking |
+| Defending policy as the primary response to a request or complaint | Positions the property as an adversary rather than a partner; the guest experiences it as "you don't matter enough for us to try" |
+| Promising a specific room, floor, or amenity without confirming availability | Generates a check-in complaint that requires service recovery; the cost exceeds the value of the original promise |
+| Offering compensation publicly in a review response | Establishes a visible precedent that any guest can exploit; may constitute a consumer protection issue in some jurisdictions |
+| Routing mid-stay complaints to the next shift without handoff | Allows a resolvable complaint to metastasize; the guest interprets it as organizational indifference |
+| Transmitting folio or identity data via unencrypted messaging channels | PII exposure risk; non-compliant with GDPR in EU jurisdictions and with local data protection law in most markets |
+
+## Cross-References
+
+- `domains/retail/skills/customer-returns` — shared complaint framing principles;
+  the acknowledge → empathise → fact-find → action → follow-up → record sequence
+  applies in both domains with domain-specific thresholds.
+- `core/compliance-lgpd` — PII handling discipline for guest data; applies to
+  folio records, identity documents, dietary information, and loyalty profile data.
+- `core/code-review-checklist` — two-pass review pattern; applies to review
+  response drafts before publishing (ADR-058 adversarial review principle).
+
+## ADR Anchors
+
+- **ADR-058** (Brainstorm gate and two-pass adversarial review): review responses
+  and complaint resolution scripts above the Moderate severity tier should go
+  through a two-pass review before delivery — draft by the attending agent,
+  reviewed by the duty manager — to avoid responses that are defensive, legally
+  exposed, or PII-disclosing.

package/.claude/skills/domains/hr/examples/attrition-model-launch.md

@@ -0,0 +1,128 @@
+---
+plan_id: PLAN-EXAMPLE-HRR
+title: "Launch attrition prediction model for voluntary turnover early warning"
+status: draft
+owner: ceo
+level: L3
+squad: hr
+profile: core,hr
+created_at: 2026-05-10
+---
+
+# Example PLAN — Launch attrition prediction model for voluntary turnover early warning
+
+> **This is an illustrative example**, not a real plan. It shows how the
+> HR squad coordinates on launching a people-analytics model that touches
+> protected-class fairness, employee PII, and manager-facing prediction output.
+>
+> Exemplar pattern derived from:
+> `.claude/skills/domains/edtech/examples/PLAN-EXAMPLE.md`
+> `.claude/skills/domains/hr/task-chains.yaml`
+
+## 1. Problem
+
+A 1,200-person technology company has experienced 24% voluntary attrition
+in the trailing year, significantly above its industry peer benchmark of
+16%. The People Analytics team has built a gradient-boosted model on 3
+years of HRIS data (tenure, performance ratings, compensation bands, team
+size, manager tenure, promotion lag, engagement survey scores) that
+predicts 90-day voluntary departure risk. The model has 78% accuracy on
+held-out data. Leadership wants to surface risk scores to HRBPs and
+direct managers so they can initiate retention conversations.
+
+Sources:
+- HRIS export: 3 years of employment data across 1,200 current + 850
+  former employees
+- Engagement survey: quarterly scores from 2023-2025 (72% response rate)
+- Performance reviews: 3 cycles, annual, manager-rated 1-5 scale
+
+## 2. Scope
+
+**In:**
+- Fairness evaluation per protected class (gender, race where legally
+  collectable, age bracket) before any manager-facing output
+- HRBP-facing dashboard showing team-level risk distribution (individual
+  scores NOT exposed to managers directly — only HRBPs)
+- Opt-out mechanism: employees can opt out of having their data used in
+  the model (LGPD consent-lifecycle requirement)
+- Re-evaluation schedule: quarterly retrain with fairness re-audit
+
+**Out:**
+- Automated manager notifications triggered by risk score (manual HRBP
+  outreach only — no algorithm-driven push to manager without human review)
+- Use of model output in compensation or performance decisions (prohibited)
+- Collecting new PII fields not already in HRIS for model improvement
+
+## 3. Squad assignments
+
+| Phase | Owner | Deliverable |
+|---|---|---|
+| P1 — Data audit | Daniel Kwon | Training data minimum-necessary review + opt-out exclusion confirmed |
+| P2 — Fairness evaluation | Daniel Kwon | Per-class disparate-impact analysis, confusion matrices, FPR/FNR tables |
+| P3 — Access-control design | Isabel Meireles | HRBP-only access, manager exclusion confirmed, compensation data redacted |
+| P4 — Opt-out mechanism | Isabel Meireles + Daniel Kwon | LGPD-compliant opt-out, propagation to training data pipeline |
+| P5 — HRBP dashboard | Daniel Kwon | Team-level aggregation only, N<5 suppression, no individual score display to managers |
+| P6 — Launch review | CEO + all VETO holders | Fairness sign-off + Operations sign-off before HRBP rollout |
+
+## 4. Risk axes and VETO holders
+
+- **Daniel Kwon (People Analytics Lead):** The model is trained on
+  historical performance ratings which may encode manager bias. Disparate-
+  impact analysis per gender and race is required before launch → BLOCK if
+  FPR for any protected class exceeds 2x the majority rate, or if any
+  subgroup cell in the dashboard has N<5 (HRR-009, HRR-010).
+- **Isabel Meireles (HR Operations Lead):** Compensation data included in
+  training features must not be individually visible on the HRBP dashboard.
+  Risk scores must never be used as justification for performance or
+  compensation decisions → BLOCK if compensation fields are exposed in
+  dashboard output or if any policy document implies risk scores influence
+  pay decisions (HRR-002).
+- **Larissa Andrade (Recruitment Compliance Specialist):** Not directly
+  applicable to this plan's core scope, but Larissa will review if the
+  model output is ever extended to recruitment (predicting new-hire
+  attrition from hiring-stage signals) — that extension requires a
+  separate plan with disparate-impact analysis on protected classes at
+  hiring stage.
+
+## 5. Task chains invoked
+
+- `hr-hire-to-onboard` — skipped for this plan (no new hire involved);
+  will be invoked if HRBP rollout requires onboarding new HRBP users to
+  the dashboard.
+- `hr-structured-recruitment` — skipped for this plan.
+- Custom analytics launch sequence: data-audit → fairness-evaluation →
+  access-control-design → opt-out-mechanism → dashboard-build → launch-review.
+  This sequence mirrors the edtech `edtech-launch-prediction-model` chain
+  adapted to the HR domain and LGPD employment context.
+
+## 6. Acceptance
+
+- Training data opt-out exclusion: employees who exercised LGPD opt-out
+  are excluded from both training and inference (HRR-009 minimum standard)
+- Fairness evaluation: per-class confusion matrices, FPR/FNR tables
+  archived before any HRBP sees a score (HRR-009)
+- No subgroup cell with N<5 displayed in HRBP dashboard (HRR-010)
+- HRBP-only access: direct managers cannot query individual employee risk
+  scores; they receive only HRBP-mediated retention conversation prompts
+- Compensation data: not exposed in dashboard output, redacted from any
+  model explanations shown to HRBPs (HRR-002)
+- Quarterly retrain schedule committed with fairness re-evaluation gate (HRR-011)
+- Opt-out mechanism live before model is exposed to any user (consent-lifecycle)
+
+## 7. Metrics
+
+- Voluntary attrition rate 90-day post-intervention cohort vs. control
+  (HRBPs who used model vs. who did not)
+- Model precision and recall at the 0.6 risk-score threshold (HRBP action
+  threshold)
+- **Per-class FPR stability across quarterly retrains** (monitored post-launch;
+  any retrain where FPR for any protected class exceeds 2x majority rate
+  triggers HRR-011 remediation review)
+
+## 8. References
+
+- `.claude/skills/domains/hr/skills/hr-onboarding/SKILL.md`
+- `.claude/skills/domains/hr/skills/recruitment-specialist/SKILL.md`
+- `.claude/skills/domains/hr/task-chains.yaml` — `hr-hire-to-onboard`
+- `.claude/skills/domains/hr/pitfalls.yaml` — HRR-002, HRR-009, HRR-010, HRR-011
+- ADR-009 — squad-bundle completeness contract

package/.claude/skills/domains/hr/pitfalls.yaml

@@ -0,0 +1,74 @@
+# Pitfalls Catalog — HR Squad
+# Loaded alongside the universal catalog when the hr profile is
+# installed. Format mirrors .claude/pitfalls-catalog.yaml:
+# id, rule, whenToUse, agents.
+
+pitfalls:
+  # === EMPLOYEE RECORD GOVERNANCE ===
+  - id: HRR-001
+    rule: "NEVER delete or destroy an employee or former-employee record without first confirming no legal hold is active (litigation, regulatory inquiry, pending DSR, LGPD erasure request under evaluation). Premature destruction forfeits the employer's evidentiary position in employment disputes."
+    whenToUse: "any data deletion pipeline, storage-cost reduction project, or offboarding automation that touches employee records"
+    agents: [Isabel Meireles]
+
+  - id: HRR-002
+    rule: "NEVER expose compensation data (salary, bonus, equity, benefits value) to any role outside the documented minimum-necessary access list. Access must be role-scoped: payroll sees all; HRBPs see their supported population; direct managers see their reports only; skip-level and above requires documented approval. Compensation data in shared spreadsheets or Slack channels is an automatic access-control failure."
+    whenToUse: "any compensation data schema design, HRIS access-control configuration, or people-analytics dashboard that includes pay data"
+    agents: [Isabel Meireles, Daniel Kwon]
+
+  - id: HRR-003
+    rule: "NEVER mutate employment status (hire, terminate, promotion, leave-of-absence, return-from-leave) without a double-approval audit trail recording: who authorised the change, who executed it, the timestamp, and the reason code. Single-approver mutations on employment status are an audit and labour-law compliance failure."
+    whenToUse: "any HRIS workflow, API endpoint, or automation that changes employment-status fields"
+    agents: [Isabel Meireles]
+
+  - id: HRR-004
+    rule: "NEVER collect health or disability information from candidates or employees beyond what is legally required for the specific jurisdiction and specific business purpose (e.g. statutory medical fitness check, disability accommodation request). Pre-employment medical histories, family medical histories, and non-accommodation disability disclosures are discriminatory collection patterns."
+    whenToUse: "any onboarding form design, benefits enrollment form, or pre-employment screening that touches health or medical data"
+    agents: [Isabel Meireles, Larissa Andrade]
+
+  # === RECRUITMENT COMPLIANCE ===
+  - id: HRR-005
+    rule: "NEVER use a global 'hire / no hire' impression rating instead of per-competency scorecard ratings recorded before the debrief discussion begins. Post-hoc rationalisation of an impression is not a structured interview. Global ratings are indefensible in disparate-impact litigation."
+    whenToUse: "any interview process design, scorecard template creation, or debrief meeting facilitation"
+    agents: [Larissa Andrade]
+
+  - id: HRR-006
+    rule: "NEVER include age references ('young', 'recent graduate', 'energetic'), gendered language ('salesman', 'manpower', 'chairman'), or physical requirements that are not demonstrably essential to the job function in a job description or interview evaluation criteria."
+    whenToUse: "any job description authoring, revision, or audit; any evaluation criteria definition"
+    agents: [Larissa Andrade]
+
+  - id: HRR-007
+    rule: "NEVER ask a candidate about salary history during any stage of the hiring process. Salary history questions are prohibited by law in multiple BR states (e.g. São Paulo) and US jurisdictions, and perpetuate compensation inequity when answered voluntarily. Ask for salary expectations instead."
+    whenToUse: "any interview script, recruiter playbook, or offer-stage conversation that discusses compensation"
+    agents: [Larissa Andrade]
+
+  - id: HRR-008
+    rule: "NEVER conduct a background check on a candidate before making a conditional offer. FCRA (US) prohibits pre-offer background checks in most configurations. Conducting checks pre-offer exposes the employer to discriminatory-screening claims if the check is used to filter candidates before they see an offer."
+    whenToUse: "any background check process design, ATS automation, or recruiter checklist"
+    agents: [Larissa Andrade]
+
+  # === PEOPLE ANALYTICS & FAIRNESS ===
+  - id: HRR-009
+    rule: "NEVER surface a predictive score, risk rating, attrition prediction, or promotion recommendation to managers without a disparate-impact analysis per protected class completed and documented. A model accuracy metric does not substitute for a fairness evaluation."
+    whenToUse: "any people-analytics model, HR scoring system, or algorithmic decision-support tool surfaced to managers or HR leaders"
+    agents: [Daniel Kwon]
+
+  - id: HRR-010
+    rule: "NEVER display aggregated people-analytics data (attrition rate, performance distribution, engagement score) for a subgroup where N<5. Small-N cells re-identify individuals from contextual knowledge. Suppress or combine subgroups below the floor."
+    whenToUse: "any people-analytics dashboard, HR report, or engagement survey disaggregation"
+    agents: [Daniel Kwon, Isabel Meireles]
+
+  - id: HRR-011
+    rule: "NEVER retrain an attrition, promotion, or performance model without re-running the full fairness evaluation suite on the new training data. Training data changes can introduce or amplify bias even when the model architecture is unchanged."
+    whenToUse: "any model retrain event, data refresh for an existing HR model, or training-data pipeline update"
+    agents: [Daniel Kwon]
+
+  # === PII & CONSENT ===
+  - id: HRR-012
+    rule: "NEVER retain rejected-candidate records past the jurisdiction-mandated maximum without documented consent from the candidate for extended retention. Minimum retention (to defend against discrimination claims) is 1-2 years in most jurisdictions; retention beyond that requires LGPD/GDPR-compliant consent."
+    whenToUse: "any ATS data-retention policy, candidate-record deletion pipeline, or talent-pool database design"
+    agents: [Larissa Andrade, Isabel Meireles]
+
+  - id: HRR-013
+    rule: "NEVER defer offboarding access revocation. All production system access, admin credentials, and physical access (building cards, key fobs) MUST be revoked on the last day of employment — not the following Monday, not when IT gets around to it. A former employee with active credentials is an insider-threat surface that cannot be remediated retroactively."
+    whenToUse: "any offboarding checklist, ITSM offboarding workflow, or HRIS-to-IAM integration design"
+    agents: [Isabel Meireles]