How every number is grounded.

Every metric on a Colleviate dossier carries an evidence marker, and quantified narrative claims are independently checked against the company's own measured evidence before they are published. This page describes the mechanism plainly: what is measured, what is verified, and what the platform does not claim.

How we ground every number

The engine attaches a provenance tier to every metric it computes: measured (taken directly from a source, such as a filing figure or a performance test result), modeled (derived deterministically from measured inputs), or estimated (inferred, and labeled as such). A number without a provenance marker does not reach a dossier. This is enforced by automated sweeps across the full corpus, not by editorial discipline.

On top of provenance, quantified claims in the narrative layer are independently verified: each claim is checked against the measured evidence for that company using a separate grounding-verification pass, and scored as verified or not. Across the corpus today, 83 percent of quantified claims independently verify against measured evidence, measured across more than 2,300 claim-verification records. The number is published as measured; claims that fail verification are flagged, not silently kept.

What "audited" means

The corpus is built entirely from public signals. When a design partner connects internal systems through private connectors, that first-party data is isolated to that customer and used only for their own diagnostics; it never enters the shared corpus or the cross-company graph.

An audited company has a complete empirical record built from public signals, layer by layer. As of July 2026, the corpus holds 5,200+ fully audited companies across a 12,100+ company universe, roughly 43 percent coverage, and grows continuously.

The layers and their sources, in plain terms:

• Technical debt and web performance: direct measurement of the company's public web surface (performance, stability, and rendering metrics), scored into a debt index and its revenue-friction consequences.
• Financial and forensic signals: public filings and disclosures, including year-over-year changes in stated risk language, read forensically rather than summarized.
• Go-to-market and funnel health: the observable marketing and conversion surface: search visibility, pricing pages, hiring signals, and the gaps between them.
• Cultural and narrative signals: public discourse and disclosure, scored deterministically, with language models used only to judge, never to compute.
• Peer grounding: every company is scored against its structural peer cluster in the cross-company knowledge graph, so findings are relative to real comparables, not absolutes.

Deterministic code computes every metric. Language models reason over the evidence and write the narrative, but they do not invent the numbers, and their quantified statements are then verified against the evidence as described above.

What we do not claim

Precision about scope is part of the methodology:

• Distress signals are leading indicators, not outcomes. They are built from disclosure and public signal; they flag elevated risk, they do not predict or confirm specific events.
• Board-grade narrative depth covers a subset of the corpus today. Every audited company has the full deterministic layer stack; the deepest multi-pass board briefs are being extended across the corpus and we say so rather than imply otherwise.
• The homepage engine demo is illustrative. It simulates the pipeline for demonstration. The corpus lookup on the same page is the opposite: real, read-only output from the live corpus, labeled as such.
• We do not publish a time-per-audit figure yet. The steady-state sample is not large enough to state one honestly. When it is, we will publish it with the sample behind it.
• Verification is a rate, not a halo. 83 percent of quantified claims verify independently; the remainder are flagged. We publish the measured rate rather than rounding it to a promise.
• Your first-party data does not build our corpus. The corpus and the cross-company graph are public-signal only. Data you connect through private connectors stays isolated to your account and powers your diagnostics, nothing else.