How far to trust each observation

Every structural observation CompanyGraph shows a reader carries a trust level. This page is the whole ledger at once: each active observation, the level it currently holds, and the evidence counts behind that level. The table below is generated from the same records the product itself reads — it is not marketing copy, and it includes the observations that have not earned trust yet.

What a level means

A trust level names which check passed. No level means "true."

• Reproducible. A separate, deterministic recompute — different code from the engine that scored the observation — re-derived the result from the company's own reported figures and got the same answer. This rules out a coding mistake in the formula. It still rests on an assumption, and we state it: reproducible from the reported figures, assuming those figures are accurate and on a consistent basis. Only this level may enter reports.
• Audited. Confirmed only by an automated reading of the evidence across companies — one reviewer, fallible, not a recompute. Good enough for search and stock pages, always labelled, never reports.
• Untrusted. No clean check has confirmed it, or it failed one. These observations still exist in the catalog — hiding them would misstate the state of the system — but they carry this label wherever they appear, and they never enter reports.

Trust is derived, never hand-set: the level is a pure function of the recorded check results, and if an observation's formula changes, its trust resets and must be re-earned. Trust is also not the same question as showing: an observation can be fully Reproducible and still be unsuitable for a report — because its name says more than the check proves. "Can we reproduce the number?" and "should a reader meet it as a claim?" are answered by different parts of the system, on purpose.

How to read the table

Each row is one observation: its reader-facing name, the machine key underneath (the join handle if you are verifying mechanically), which check earned its current level, and when that assessment was signed off — so every level tells you how old it is. The table deliberately shows no per-row evidence counts: the different checks cover very different company samples, and a single number column would blur them together. The per-company receipts — the actual recomputation, with its calculation trail — appear on the stock pages, next to the firing they verify.

If you are verifying this page mechanically: the data behind it is public JSON at /api/public/en/method/observation-trust, and each observation's reader-facing note (the same level, in plain words) appears wherever that observation is shown on a stock page.

How an observation earns these checks — the full path from a reported number to a bounded claim — is From data to checked claims. The mistakes this machinery has caught, and what changed because of them, are on Corrections: how the method became stricter.