Every number you see in ClinicalRate carries its receipts. This is the full provenance chain from raw placement event to the percentile rendered in your browser.
Public feeds, partner contributions, VMS ingests, public-domain.
Records below threshold route to human review, never silently bucketed.
Crisis-rate outliers visible at p90 without bending the median.
Below floor, individual percentiles suppress and "thin market" surfaces.
Nine source classes — public job-board feeds where re-use is permitted, contributing staffing agencies under data-share agreements, direct ingest from supported VMS platforms, and the curated public domain (BLS, state filings, public RFPs).
●9 source classes · per-record provenanceEach source lands in a typed staging table with source ID, ingestion timestamp, and content hash. Duplicate detection runs before normalization — a single placement contributed by both an agency and a VMS is reconciled, not double-counted.
●Idempotent · de-duped · hashedEvery record passes through the Taxonomy Mapper before it enters a distribution. Records that don't resolve at confidence ≥ 0.92 are routed to a human review queue, not silently bucketed.
●428 roles · ≥0.92 thresholdLinear-interpolated empirical percentiles on the trailing 30-day window, winsorized at 1.5×IQR to suppress single-record outliers. Never a mean masquerading as a median. Thin markets (n < 30) suppress individual percentiles and surface a "thin" indicator instead.
●p10 → p90 · 1.5×IQR · n≥30Every UI cell and every API response carries as_of, sample_size, and the source-class distribution. Defensible by design — the buyer can always ask "where did this number come from" and you can answer in one click.
●UI · API · CSV · all surfacesSurvey reports sometimes publish a "median" derived from a parametric fit. ClinicalRate uses empirical percentiles on the raw data — interpolation only, never assumption.
"RN" is not a role. The platform refuses to render a distribution that mixes ICU and Med/Surg, or travel and per diem, into one chart. Every distribution is a single canonical cohort.
Every cell carries its as_of timestamp. If a source feed lags, the affected distributions are flagged in the UI rather than served as current.
Crisis-rate placements ($300/hr Travel RN bookings during 2020) are real, but rare. Winsorization at 1.5×IQR keeps them visible at the 90th percentile without bending the 50th.
| Audit surface | What you see | Tier |
|---|---|---|
| Sample size on every cell | n events behind every percentile, in UI and API. | All tiers |
| As-of timestamp on every cell | Most recent ingest contributing to the cell. | All tiers |
| Source-class distribution | What share of the cell came from each of the nine source classes. | Operator + |
| Per-record provenance trace | Down to the individual record's source, ingest timestamp, and normalization confidence. | Enterprise |
| Annual methodology audit by external firm | Report shared under NDA. | Enterprise |
The decisions that distinguish a defensible distribution from a polished number.
“The audit was three meetings, not three months. Every cell carried its receipts and our auditor signed off without a methodology dispute.”
When the rate-card meeting becomes a methodology meeting, the number defends itself.
Every benchmark drillable to the source-class mix. Auditor sign-off in meetings, not months.
A pricing agent can cite the n, the as-of, and the source mix in the same response. No "where did this come from" tickets.
Provenance is a platform default, not an upgrade. Sample size and as-of on every cell, on every tier. Per-record trace and external audit on Enterprise.
Book a demo. Bring the question your CFO asked you that you couldn't answer — we'll show you where the answer lives in the chain.
Book a data demo →