Understanding findings
A finding is a single, evidence-backed claim that a specific page is likely wrong — with the quote, the reasoning, and a confidence score behind it.
Anatomy of a finding
Every finding carries the same parts, so you can judge it in seconds:
- Type — what kind of problem it is, from the taxonomy below.
- Evidence — one or more spans quoted verbatim from the page. Nothing is flagged without a quote.
- Reasoning — one or two sentences explaining why the quote is suspect.
- Confidence — a score from 0 to 100. Higher means the app is more sure the page needs attention.
- Suggested action — review the page, verify a reference, or reconcile it with a related page.
- Related page — for contradictions, the other page involved is always linked.
- Status — where the finding is in its lifecycle (see below).
The epistemic stance — what the app does and doesn’t claim
Evergreen AI surfaces pages most likely to be wrong. It does not claim omniscience, and it has no external source of truth for facts like “we use Jenkins” or “contact Jane.” So it calibrates its language deliberately:
- It asserts a problem with confidence only where a page convicts itself — an expired validity window, internally broken logic — or where two pages in your own tenant directly conflict.
- Everything else is framed as “likely wrong / worth verifying, here’s the evidence,” leaving the judgement to you.
- Ownership and contact references are always presented as something to verify, never asserted as wrong — the model can’t know your org chart.
The point of the precision engineering is to decide which pages earn your attention — not to be right about every claim on every page.
The six finding types
Examples below are illustrative of the format only.
| Type | What it means | Example evidence |
|---|---|---|
| OUTDATED_FACT | States something time-bound as current that is likely no longer true. | “We currently deploy via Jenkins” on a page where another page describes a move to GitHub Actions. |
| EXPIRED_BY_OWN_TERMS | The page declares its own expiry or validity window, and it has passed. Self-convicting — stated with confidence. | “This policy is valid until 31 Dec 2024.” |
| DEPRECATED_REFERENCE | References a system, process, or version that other pages describe as deprecated or renamed. | A runbook step calling a service by its pre-rename name. |
| CONTRADICTION | Materially conflicts with another page. Both pages are always cited. May be off by default — see note. | Two pages giving different expense limits. |
| INTERNAL_INCOHERENCE | The page contradicts itself or is structurally broken. Self-convicting — stated with confidence. | “See step 7” in a five-step list; a DRAFT — DO NOT USE marker on a published page. |
| STALE_OWNERSHIP | Names a person or role as a contact or owner. Flagged for human verification — never asserted wrong. | “Escalate to J. Smith” — please verify this contact is current. |
About CONTRADICTION (beta). Cross-page contradiction detection is the newest finding type. To keep precision high, it may ship switched off by default and, when enabled, surfaces only the highest-confidence candidate pairs (such as pages that explicitly link to each other). You control it in Settings → Sensitivity, where it carries a “beta” label.
Confidence and sensitivity
Each finding has a confidence score from 0 to 100. A sensitivity threshold decides how confident the app must be before it shows you a finding at all:
| Sensitivity | Threshold | Trade-off |
|---|---|---|
| Conservative | 80 | Fewest findings, highest confidence. Best for a first impression. |
| Default | 70 | Balanced — the recommended starting point. |
| Thorough | 60 | Casts a wider net; expect more to review. |
Anything below the threshold is suppressed — you never see it. To keep the list defensible, the app also caps each page at a maximum of five findings, keeping the highest-confidence ones. Borderline findings get a second verification pass before they ever reach you, which re-checks whether a careful human would call the page genuinely problematic or merely old.
The finding lifecycle
Findings move through a deliberately simple state machine — no approvals, no multi-stage workflows:
OPEN → CONFIRMED or DISMISSED → RESOLVED
- Open — newly surfaced, awaiting triage.
- Confirmed — a human agrees it’s a real problem worth fixing.
- Dismissed — not a problem (with an optional reason). Dismissing can also teach the app not to raise it again — see the suppression learning loop in the dashboard guide.
- Resolved — the underlying page has been fixed.
Why an empty report is a good report
When the app isn’t confident, it stays quiet. A clean “all clear” for a scanned space is a deliberate feature, not a gap — it means the pages you care about held up to scrutiny. The goal is a short list worth acting on, not a long list you learn to ignore.