evidence-safety-reviewer

Audit candidate evidence provenance, privacy, consent, and external action safety in recruiting workflows.

1|Updated Aug 3, 2026
One-click install
npx skills add https://github.com/getyak/talent-signal --skill evidence-safety-reviewer-getyak
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: evidence-safety-reviewer
Source: https://github.com/getyak/talent-signal/tree/main/.agents/skills/evidence-safety-reviewer
Command: npx skills add https://github.com/getyak/talent-signal --skill evidence-safety-reviewer-getyak

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Recruiting tools that extract facts from private conversations can silently bind evidence to the wrong candidate, present unsupported claims as confirmed, or trigger unauthorized external writes. This Skill audits the full chain from source artifact to user-confirmed fact to external action, catching safety failures before they harm a candidate relationship. ## Core Features & Use Cases - Provenance and Identity Auditing: Verifies source-level traceability including OCR spans, speaker assignment, candidate matching, and edit history, with adversarial tests for same-name candidates and forwarded messages. - Privacy Lifecycle Review: Challenges data minimization, retention, export, deletion, vendor defaults, and cross-tenant authorization across raw images, OCR, embeddings, and backups. - External Action Governance: Requires explicit previews, execution-time permission checks, idempotency, verified results, and append-only audit records for contact, calendar, and notification writes. - Use Case: Before shipping a new OCR extraction pipeline, run this review to confirm every extracted assertion links to a quoted source span, ambiguous dates trigger clarification rather than silent guesses, and no calendar invite can be sent without explicit recruiter approval. ## Quick Start Use the evidence-safety-reviewer to audit this candidate evidence extraction flow for provenance, privacy, and unsafe external actions before release.

Frequently Asked Questions about evidence-safety-reviewer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I audit an AI recruiting tool for evidence provenance?▼

Trace each extracted assertion back to its source asset, OCR span, speaker assignment, and model version. Require exact quoted spans for every claim and test adversarial cases like same-name candidates, forwarded messages, and cropped screenshots.

What safety checks should gate external actions like calendar invites?▼

Require an explicit preview of target, fields, timing, and effect with edit and cancel options before execution. Verify permissions at execution time, enforce idempotency, confirm the external result, and keep an append-only audit record.

How should AI extraction handle ambiguous dates or conflicting evidence?▼

Ambiguous dates and missing years must trigger clarification anchored to capture context, never silent guesses. Conflicting evidence must create a visible conflict state rather than an overwrite, and no-action must remain a first-class outcome.

What privacy risks exist in OCR-based candidate screening pipelines?▼

Risks include raw images and OCR text leaking into analytics or evaluation logs, unbounded retention of embeddings and backups, cross-tenant exposure, and vendor model training on private conversations. Audit the full lifecycle including export and deletion across every derived representation.

When should a release be blocked for evidence safety failures?▼

Block release when evidence can bind to the wrong candidate without mandatory review, unsupported assertions appear as confirmed, sensitive traits are inferred, external writes occur without authorization, or source and deletion paths are unavailable. Vetoes are never averaged against other scores.