How Sonder scores a candidate
Most AI screeners hand you a number and ask you to trust it. This page publishes Sonder's formula: the components, the weights, the rules, and the places it can be wrong — because a score you can inspect is worth more than a score you have to believe.
The fit % is a weighted blend
40% must-have skill coverage, 30% role match, 20% experience fit, 10% preferred skills. Defaults — every job can reweight the blend and the engine renormalizes.
40% — must-have skills, deterministically
Each must-have is checked against the CV through a skill ontology that knows java is not javascript. No neural network in this leg: a skill is evidenced or missing, identically on every run. Core skills count double; a hard "no X" from the recruiter's brief deducts points as a visible penalty.
30% — role match, semantically, field by field
A language model running inside Sonder — never a cloud AI service — compares JD to candidate in three passes: title vs title (35%), skills vs skills (40%), context vs CV body (25%). Field-by-field is deliberately harder to fool than one blob-to-blob comparison. Candidate CVs are never sent to a cloud AI.
20% — experience fit, simple published rules
Meets the minimum: full marks. Below it: proportional credit. Unknown years: a neutral 0.5, because unknown is not unqualified. More than four years over the stated maximum: 0.75, flagged as notably overqualified.
10% — preferred skills
Same ontology check as must-haves at a quarter of the influence. They separate two otherwise-equal candidates; they never rescue a missing must-have.
Every score shows its evidence — or says it does not have it yet
Each matched requirement cites the CV lines that produced it, and the deeper analysis builds a claim-by-claim evidence table. On a large pool, ranking runs ahead of line-by-line verification: those rows are labelled evidence pending, on your screen and on your client’s, rather than dressed up as verified. You are never shown a citation that was not computed.
The score orders attention. It never decides.
Sonder does not auto-reject, auto-advance, or hide anyone. The ranked list is where a recruiter starts reading, not where a candidate stops existing.
Feedback tunes it — within limits
Recruiter thumbs and client verdicts nudge per-skill affinities for that workspace only. The nudges are capped, inspectable and reversible.
Where it is weak, we say so
Adjacent roles on the same platform — a Benefits versus a Payroll consultant — can look similar to the semantic leg. That is why hard requirements stay deterministic and cited, and why the final call is always human.
Security and data
Workspace isolation on every query. Scoring never sends a CV to any AI provider — it is local ML in every edition. The optional narrative layer is the one place a cloud model can appear: it runs locally by default, and a cloud workspace may route it through Anthropic behind a per-workspace switch, with every such call recorded in your decision ledger. DPDP consent facts stored at collection, retention clocks on unreviewed data, WhatsApp opt-outs enforced at the API. HttpOnly cookie sessions, optional two-factor auth, weekly snapshots, one-click export.