How Aria scores candidates, and how we audit ourselves.
Ask Aria is an agentic recruitment assistant. Decisions about people deserve transparency. This page documents how Aria sources, ranks and surfaces candidates, what we deliberately won't infer, where the system is weakest, and how to challenge a result.
1. What Aria does (and does not) decide
Aria is a recommendation surface, not a hiring decision-maker. Every shortlist, score, and outreach draft is produced for a human recruiter or hiring manager to review, edit, and act on. Aria never sends an email, books an interview, posts a job or submits a candidate without an explicit human confirmation step. Aria never auto-rejects candidates.
Decisions Aria makes
- Which public profiles to surface as plausible matches for a brief.
- A 0–5 score across five dimensions (see below) and a derived match percentage.
- A short rationale (“why”, “watch out”, “wildcard”, “tldr”) for each candidate, grounded in visible signal.
Decisions Aria does not make
- Whether to interview, reject, hire, or pay a candidate.
- Whether a candidate is “a culture fit” in any holistic sense ndash; the score is on visible signal, not character.
- Anything that depends on protected characteristics (see §3).
2. The scoring rubric
Every ranked candidate gets a 0–5 score across five weighted dimensions, plus a confidence band and visible-evidence requirement. The match percentage is the weighted sum.
| Dimension | What it measures | Weight |
|---|---|---|
| Hard skills | Specific tools, languages, frameworks named in the brief, seen on the public profile. | 30% |
| Experience | Years and scope at the relevant seniority, in the relevant role family. | 25% |
| Soft skills | Leadership / communication / scope signal where visibly evidenced (talks given, published work, team-lead titles). | 20% |
| Culture fit | Match to HM-supplied cultural signals (e.g. “ships fast, no process”). Never a personality call. | 15% |
| Growth potential | Trajectory + non-obvious upside (the “wildcard”). | 10% |
A confidence band (high / medium / low) is attached to every candidate based on how much visible evidence supported the score. If Aria can't find concrete evidence, she is required to say so in the “why” field instead of inventing a rationale.
3. Anti-bias rules built into the prompts
The sourcing and ranking prompts include explicit, enforced rules that bar Aria from inferring or using protected characteristics, even where the signal exists in the data she sees.
- No inferred demographics. Aria does not infer race, ethnicity, age, gender, sexual orientation, disability, religion, marital status, pregnancy, or HIV status from any data source.
- No anchoring on names or photos. When candidate photos appear in source data, they are not passed to the ranker. Names are used only for verification, never as a feature in scoring.
- No salary anchoring. Aria does not score candidates based on their previous salary, and refuses to surface previous salary as a screening signal.
- No anti-junior bias. Search prompts forbid an “education cluster” (degree / institution) in the boolean, because that demonstrably skews recall against capable juniors and self-taught candidates.
- No invented signal. The prompts forbid composing a job title or claim that the SERP snippet does not literally contain. If a current title cannot be verified, the candidate is either dropped or flagged “uncertain” with the evidence shown to the recruiter.
4. Country gating & location verification
Where the brief names a country, candidates whose location is clearly outside that country are excluded from the ranked shortlist. Where the location is missing, the candidate is included but the “location is unverified” flag is surfaced to the recruiter and the experience score is reduced by one point. This is to prevent location signal from silently inflating or deflating a score.
5. Cost, model, and per-action telemetry
Every Aria action records, for the customer's own visibility:
- The model used (Claude Opus, Sonnet or Haiku for reasoning; Claude Haiku for cheap classifications; OpenAI for embeddings).
- Input and output token counts.
- The estimated cost in ZAR, against the customer's monthly budget.
If a customer's budget exceeds 90% of plan, Aria automatically falls back from Opus to Haiku to protect runway. Customers can see this in their Usage page in real time.
6. Known weaknesses (honest)
This is not a complete list. It is the list of issues we already know about and are tracking.
- No third-party bias audit yet. The fairness rules above are self-published. A third-party assurance review is planned for 2026 H2.
- SERP recency. Public profile snippets can be weeks or months old. Aria flags when a verified title relies on a cached snippet, but the recruiter must still confirm.
- Boolean recall vs precision. When the HM brief is thin, Aria's boolean errs on the side of higher recall, which produces some weaker matches. We'd rather show more and let the recruiter cut than miss real candidates.
- Apollo coverage gaps. Apollo's database underrepresents certain SA metros. When that happens Aria broadens the location filter and tells the recruiter she did.
- RAG freshness. Internal knowledge (meetings, references, candidates) is indexed via a queue. There is a short delay between a write and that data becoming searchable through Aria's knowledge base.
7. How candidate data is handled
- Per-customer Postgres RLS isolates every candidate record to the org that created it. Other customers cannot see another customer's data.
- Right-to-erasure and subject-access-request flows are wired to
compliancetools available to each customer. - Candidate notes, references, and contact data are retained according to each customer's configured retention window in Settings.
- Reference responses are collected via individual tokenised links; the candidate can withdraw consent and we delete on request.
8. How to challenge a result
If you are a candidate, a recruiter, or a regulator and you believe an Aria-produced result is biased, inaccurate, or otherwise harmful:
- Email [email protected] with the role and candidate context (or the search ID if you have it).
- We respond inside one business day, and will explain the specific scoring inputs and prompts behind that result if asked. The system prompts are not a secret.
- If the concern relates to candidate personal data, the same address routes to our POPIA / privacy responder.
9. Versioning
This page is versioned. When the scoring rubric, prompt rules or model selection change in a way that affects candidate outcomes, this page is updated and the change is dated below.
Change log
v1, 30 June 2026. Initial publication.