Know what the system does before it shapes a shortlist.
This policy names every current scoring, ranking, filtering, and generative system in CandidatesAssessment—together with its inputs, outputs, limits, and required human review.
Policy version 1.0 · effective and last reviewed 27 August 2026 · owner: CandidatesAssessment product team
Automation can organize the evidence. It does not inherit the judgment.
Software can calculate evidence, assign a recommendation label, rank completed candidates, filter a generated client brief, and draft one short summary paragraph. Those actions can influence what a recruiter sees, so they require review—not blind acceptance.
Recruiters must review the underlying evidence, role relevance, confidence, limitations, accommodations, and material candidate context before acting or sharing a shortlist. Clients remain responsible for the final decision.
The system register, in plain language.
“Rules-based” means the same structured inputs follow coded thresholds and selection rules. “Generative” means a language model drafts new text from a restricted evidence summary.
| System | Inputs | Outputs | Required human gate |
|---|---|---|---|
| Evidence engineRules-based · not a learning model | Completed assessment scores, answer-derived dimensions, coverage, consistency, response quality, and the selected role context. | Competency evidence, recommendation level, confidence, strengths, risks, development areas, and interview questions. | Review individual results and limitations before comparison, ranking, interview, or client sharing. |
| Shortlist selectionRules-based · not a learning model | Completion status, recommendation label, overall evidence score, ties, and the number of openings. | Completed candidates labelled Strong or Proceed are ranked and selected for the generated client brief; Caution, Review, and incomplete records are not included. | The recruiter initiates the export, checks who is included or absent, reviews the PDF, chooses the recipient, and can revoke the link. |
| Client-brief narrativeGenerative · Anthropic Claude | Role title, candidate names, recommendation labels, score-grounded evidence, shortlist counts, and locale. | One short executive-summary paragraph inside the client-facing PDF. A deterministic evidence-derived paragraph is used if the model call fails. | The recruiter reviews the generated PDF and source evidence before sharing. The paragraph must not be treated as a new score or decision. |
Current model configuration for the narrative step: Anthropic Claude Sonnet 5. A model or provider change requires this register to be reviewed.
The generative step receives a narrow evidence packet.
Only the client-brief narrative uses a third-party generative model. The rules-based evidence and shortlist systems run in the application logic and should not be described as machine learning.
Anthropic’s current public API notice says inputs and outputs are deleted from its backend within 30 days by default, with stated exceptions for longer-retention services, different agreements, usage-policy enforcement, and law. CandidatesAssessment does not currently claim a zero-data-retention agreement.
Read Anthropic’s Current Retention NoticeSent to Anthropic
- Role title and locale
- Candidate names
- Recommendation labels
- Score-grounded evidence summaries
- Candidate and shortlist counts
Not sent in this step
- Raw assessment answers
- Candidate email addresses
- Client recipient email address
- Report access token
- The complete PDF or its share link
What the current platform does not do.
This boundary describes the product as it operates today. It is not a promise about every possible future feature.
- No facial, voice, or video emotion inference.The platform does not score a candidate from their face, voice, accent, or recorded video.
- No résumé or social-profile scraping.The automated recommendation engine is not enriched from scraped employment or social-media data.
- No protected-characteristic inputs.The current recommendation rules do not use protected-characteristic fields as scoring inputs.
- No self-training from customer records.The current engines do not learn new recommendation rules from customer or candidate records.
- No automated rejection communication.The platform does not send a rejection decision to a candidate.
- No final hiring decision.Recruiters and clients remain responsible for the lawful decision and its explanation.
Keep a person at every consequential handoff.
The recruiter’s review is not a ceremonial click. It is where job relevance, context, accommodations, uncertainty, and conflicting evidence must be considered.
- EvidenceCompleted, role-relevant assessmentsCheck access, completion, and material context.
- RulesRecommendation, confidence, risks, and questionsRead individual evidence before ranking.
- SelectionStrong and Proceed candidates enter the generated briefCheck inclusions, exclusions, ties, and openings.
- DraftClaude drafts one summary paragraphVerify every statement against source evidence.
- DecisionRecruiter and client interpret, interview, and decideDo not rely on the automated output alone.
Known risks need operating controls.
A transparent system register is useful only if it changes how teams review, share, correct, and challenge the output.
| Risk | Current control | Recruiter responsibility |
|---|---|---|
| Automation bias | Systems and labels are named; final decisions are not automated. | Review source evidence and plausible alternatives instead of accepting the ranking as authority. |
| Incomplete or inconsistent evidence | Coverage, consistency, response quality, and minimum-evidence gates contribute to confidence. | Pause interpretation when completion, access, or material context makes evidence non-comparable. |
| Generative error or overstatement | The prompt is restricted to score-grounded evidence and a short paragraph; failure uses deterministic copy. | Check every generated statement before the PDF is shared. |
| Caveats lost in client handoff | Internal outputs can preserve confidence and limitation signals. | Review the internal result first; not every internal caveat is reproduced in the client brief. |
| Stale or corrected output | The issued PDF is immutable and its review link can be revoked. | Revoke the old link, correct the source issue, and generate a new export. |
Claims stop where evidence stops.
This policy reports current product behavior. It does not convert planned governance work into a live control or an external assurance claim.
Review Current Platform StatusSupported today
- Named system purposes, inputs, outputs, and human gates
- Deterministic recommendation and shortlist rules
- Evidence identifiers and engine version in internal outputs
- Confidence gates and deterministic narrative fallback
- Revocable protected sharing and immutable issued PDFs
Not claimed today
- Independent responsible-AI audit or AI certification
- Completed adverse-impact or fairness study of this workflow
- Independent outcome-validity study for automated shortlist selection
- Zero-data-retention agreement with the model provider
- Complete model/output lineage or full audit-event history
Responsibility stays named.
Employment, privacy, accessibility, and automated-decision obligations vary by jurisdiction. This policy is product documentation, not legal advice.
Operate the documented systems, restrict the generative data packet, maintain the product boundary, investigate reported output issues, and update this register when behavior changes.
Select job-relevant assessments, establish a lawful basis, give candidate information and support, review evidence and accommodations, control sharing, and explain how evidence was used.
Consider role context and other lawful evidence, avoid sole reliance on automated output, conduct appropriate interviews and checks, and own the final hiring decision.
Request process information, context review, or an accommodation from the inviting agency; report a platform or output problem to CandidatesAssessment when needed.
Pause, revoke, correct, regenerate.
If an automated output appears wrong, incomplete, unfair, or based on incorrect data, do not keep distributing it while the issue is reviewed.
- Pause use and sharing.Do not treat the disputed output as a resolved hiring signal.
- Revoke the active client link.The recruiter can stop access to the current shared brief.
- Check the source evidence.Confirm identity, completion, role relevance, context, and the displayed result.
- Report the issue.Candidates should contact the inviting agency first for a hiring outcome, context, or accommodation review. Platform issues can be sent through the contact route below.
- Generate a new export after correction.An issued PDF is immutable; a corrected brief requires a new export and review link.
Public reference points.
These sources inform the operating questions in this policy. They do not certify or endorse CandidatesAssessment.
Questions about automation, answered precisely.
Use the privacy, DPA, security, and platform-status pages for the wider data and product boundary.
Review Security & Data HandlingDoes CandidatesAssessment use AI to score assessment responses?
The current evidence and shortlist engines are deterministic, rules-based systems—not learning models. A generative model is used only to draft one short executive-summary paragraph in the client-facing shortlist PDF.
Can the platform automatically reject a candidate?
The platform does not send automated rejection messages or make the employer’s final decision. It can rank completed candidates and filter a generated client brief to Strong and Proceed recommendation levels, so the recruiter must review who is included, who is absent, and the underlying evidence before acting.
What candidate data is sent to Anthropic?
The narrative request includes the role title, locale, candidate names, recommendation labels, score-grounded evidence summaries, and shortlist counts. It does not include raw assessment answers, candidate email addresses, the client recipient email, report access tokens, or the full PDF.
Can a candidate request human review?
Yes. The inviting agency remains responsible for the hiring process and should be the first contact for outcome, context, or accommodation review. A candidate or recruiter can also report a platform or output issue to CandidatesAssessment through the contact page.
How is an incorrect client brief corrected?
Pause distribution, revoke the current review link, verify and correct the source issue, then generate and review a new export. The issued PDF is immutable, so it is not silently edited after sharing.
Use structured evidence without surrendering the decision.
Create an agency account, choose a role-relevant assessment plan, review every result in context, and keep the client conversation grounded in evidence people can inspect.