Sourced, dated, and never inflated.

Articles and agency playbooks are written by the Multilat team and reviewed by current or former UN staff with direct experience of the system being described (§6.7.3 of the product specification). Each article names its reviewer category. We do not publish claims about how a specific office or panel behaves unless they are corroborated by at least two independent sources.

Articles are dated and versioned. When an HR policy changes — a new salary scale, a shift in JPO sponsorship terms — we update the article and mark the revision. Old conclusions that no longer hold are explicitly corrected, not quietly edited.

Every pattern (visa sponsorship, consultant-to-FTA conversion, contract renewal) describes an office, not an individual applicant. A headline might read: “UNICEF Nairobi country office stated visa sponsorship in approximately 62% of postings at P3 and above (n = 14, 90% CI [0.41, 0.79]).” It will never read: “Your odds of visa sponsorship are 62%.” The distinction is deliberate and enforced throughout the product (§6.11.9).

Cohort floors are non-negotiable. Visa and recruitment patterns require at least 10 evidence rows. Conversion and renewal patterns require at least 15. No single data contributor can exceed 30% of evidence weight. Where the data does not support a number, you see a label — “moderate”, “low confidence”, or “insufficient data” — never a false-precision estimate. Every pattern card shows the evidence count, confidence band, source mix, and the last-recomputed date.

No LLM produces pattern probabilities. Wilson confidence intervals are computed deterministically; LLMs extract structured fields from source text, which the deterministic layer then scores.

Every salary figure on the platform comes from a publicly available ICSC scale, a published post adjustment index circular, or an official agency HR manual. We do not use private offer data, salary surveys, or crowd-sourced compensation databases. Every figure shows its source document and effective date — either inline or in a tooltip. If the underlying scale has not been updated in the past 12 months, we flag it.

Fit scores are computed deterministically from structured fields extracted by an LLM (your CV against the job description). The LLM extracts; it does not score. The calibrated score is capped at 90% — because no analysis of a CV against a JD can capture panel composition, internal candidates, or geographic representation status. The cap is a product constraint, not a hedge.

Every AI-generated artifact — CV, cover letter, motivation statement — runs through a two-pass fabrication check before you see it. First, a regex pass flags any number, date, employer name, or proper noun not present in your source CV. Then an LLM auditor checks for unverifiable claims. If either flags, the affected segments are excised and the artifact is regenerated. You see the audit pane on every tailoring output; it is never hidden.