Responsible AI

AI should explain the work, not hide it.

SourcingOS uses AI to help interpret roles, organize evidence, compare candidates, propose sourcing strategies, and accelerate repetitive research. It is not designed to make final hiring decisions or silently turn uncertain data into fact.

Evidence before assertion

Candidate reasoning is designed around requirement-by-requirement evidence. SourcingOS distinguishes explicit, inferred, normalized and derived assertions, along with uncertainty and missing evidence. A high aggregate score cannot erase a failed hard requirement.

Requirements are not search expansions

Recruiter language is separated into hard requirements, preferences, normalized equivalents, discovery expansions, geography expansions, numeric thresholds, credentials and anti-signals. Discovery terms can help find people; they do not silently become proof that a requirement was met.

External content is untrusted

Resumes, public profiles, webpages, job descriptions and provider data may contain instructions that look like prompts. SourcingOS treats that text as evidence content. It cannot redefine system policy, grant tool access, export secrets, or authorize an action.

Recruiter-controlled actions

Identity confirmation, contact reveal, learning changes and other consequential actions use explicit approval/authorization boundaries. Recruiter learning is intended to be inspectable: what changed, why it changed, and how to edit or undo it.

Sensitive and protected characteristics

SourcingOS should not infer protected characteristics from names, photographs, language, schools, locations or other proxies for the purpose of ranking candidates. Role criteria should be job-relevant and evidence-backed.

Model and provider transparency

AI providers are treated as processors of defined inputs, not as the canonical candidate database. The long-term policy is to disclose which AI/provider class handled a feature, what data may leave SourcingOS, and the applicable retention/training terms before making broader enterprise claims.

Evaluation

RecruiterBench tests role-understanding quality without spending live sourcing credits. SecurityBench tests adversarial trust boundaries. Product changes should improve measured recruiter truth while preserving safety floors rather than optimizing an opaque model score.

Trust Center Security Methodology