How verdicts are computed
Last updated: September 2026
A security finding is only worth as much as the answer to two questions: what was actually read, and what was it compared against. This page answers both, and describes the mechanism rather than the intention — where the two would differ, what follows is the mechanism.
The model does not decide anything
The load-bearing part of an answer — the verdict, the status of each step in an attack path, the counts and the citations — is rendered from engine data before the model is called, and it is emitted to you exactly as rendered. The model is then given a bounded job: narrate what is already there, without restating it. It never rewrites the verdict, and it never supplies one.
This is a structural choice rather than a filter. A system that lets a model write the verdict and then inspects the prose for mistakes can only catch the mistakes it thought to look for. Rendering the verdict first makes whole categories of error impossible to emit: the verdict comes from the engine's own status field, the grounding is the set of confirmed findings, the counts come from the finding records, and citations are rendered from a catalogue. Any sentence the model produces that states a verdict of its own is removed before you see it, whether it agrees with the computed one or not — because a model that happens to agree today is not a model that can be relied on to agree tomorrow.
The same rule is enforced once more at the boundary where a response leaves the server: a value that looks like a verdict must carry the mark that only the gate applies. An unmarked verdict is refused rather than served.
What a verdict is compared against
Your instance. Every check is verified against your instance, on your release, at scan time: each table, column, property and plugin a rule reads is confirmed by that scan’s own reads of your instance before the rule may return a verdict, and the verdict records which scan of which build it rests on. Nothing about your instance is compared against a snapshot of somebody else’s.
Where a check cannot be verified with the read-only connection, it reports UNKNOWN with the reason — never a pass or a skip. Each path reads things the other does not. An upload, which runs in your instance, reads the plugin inventory and the tables ServiceNow keeps off its web-service API. An upload does not yet carry the record creator, the OAuth client and scope reads or exposure context, which the connection reads. A table or column an upload shows your instance does not have makes that check not applicable, and it is shown as a row rather than dropped. A property or plugin that is missing is not a missing feature: it is the value the check measures.
The names we ground chat answers and generated scripts on are a separate thing: a catalog captured from live Personal Developer Instances, not generated from training data. The captures are zurich, australia, with 90 verified system properties distilled from a 3,585-property capture and a 6,419-table inventory. The 143 per-table field catalogs beside them are compiled from our own checks and detection scripts; they are a catalog, not a capture. These are our regression reference and our documentation-drift baseline; a name they do not carry is reported to you as unconfirmed, and no verdict about your instance is compared against them.
A verdict that cannot be held at full confidence says so rather than quietly softening. If a result carries no verification of your instance, or the last reading of it could not be confirmed, the verdict is capped and the reason travels with it — including how long ago the reading was taken.
The five rules
R1 · Confidence expires. An instance reading is good for 24 hours. Past that the verdict downgrades itself and prints the reading’s age, because a refresh that could not run and one that never happened are the same position to be in.
R2 · The AI never decides. Verdicts are computed by rule before the model writes a word, and verdict-shaped prose is stripped out of what it writes afterwards.
R3 · It refuses rather than guesses. An advisory we do not track returns “not a tracked CVE id as of <date>” — never “no known issue”. An absence of evidence is reported as an absence, not as a clean result.
R4 · Claims withdraw themselves. A rule change supersedes every verdict decided under the older rules. The dashboard asks for a re-scan and names the version; it never just goes quiet.
R5 · Mappings are dated. Every regulatory citation names its reviewer and the date it was read — or says plainly that nobody has reviewed it.
How a regulatory mapping is qualified
Saying a finding maps to an article of NIS2, DORA, ISO 27001 or NCA ECC is a claim about a regulation, and it is not one a scanner can make on its own. So every mapping carries its own review state, and the state is shown next to the citation rather than buried in a methodology note.
- Unattested. The mapping exists and nobody has reviewed it. It renders with the tag (unattested) and it caps the verdict it supports at conditional.
- Attested. A named reviewer has read that specific mapping and recorded the decision in an append-only log. What this guarantees is a name and a date, not a signature — a signature is something a person puts on a delivered document, and this page will not blur the two.
- Expired. A review is good for a year. Past that the citation reads (attestation expired on …) with the date it was last read, and it caps the verdict exactly as an unreviewed mapping does. A framework whose reviews have all expired is treated as unreviewed at the document level, which is the conservative reading.
- Not registered. Where no record of the underlying instrument exists to check against, the citation says so instead of implying one was checked. This also caps.
Two further details, because they are the ones that decide whether any of this is worth anything. The unit is a single rule-to-control mapping rather than a control, so attesting one rule's view of a control does not attest another's. And the control's own title is part of what a reviewer signs off: rename the control and the mapping returns to unattested, deliberately, because a renamed control is a claim nobody has actually read.
We do not pre-attest mappings to make a dashboard look finished. A mapping is reviewed when a customer needs it reviewed, and until then it says, on its face, that it has not been.
What happens to an answer before you see it
Generated code is bound to the same release-pinned catalog the verdicts are compared against. Fabricated tables, fields or API names are caught at the bind step, before any script is generated. On Connected, your tenant's actual schema is extracted on connect, so scripts target your customisations rather than a generic template.
Every answer is checked by 16 automated detectors: uncited claims, unknown knowledge-base IDs, self-contradictions and fabricated statistics among them. Claims link back to specific knowledge-base articles, and the IDs are validated after the response is written; if the source is not there, the citation does not ship.
Fabricated property names and unsupported claims are stripped before display, not quietly flagged in a log nobody reads. Every response carries a quality signature tracked in the admin console, so a regression is visible and drift is auditable.
When we cannot read something
If we can't read it conclusively, the verdict says so — silence is never rendered as good news. A check whose underlying probe did not run is reported as not assessed, and no surface infers a pass from the absence of a finding. This is the single rule most of the others are a consequence of.
Impact sentences
Each scored finding carries one sentence saying who can do what, to which data, and with what consequence. It is assembled from slots, not written per scan. Who, what and the data class come from the check's own declared claim. The finding's evidence can only narrow that claim, never widen it. The figures are the finding's own counts, copied as they were read, “at least” included. The consequence is the one piece of free text, written by us for each check. No language model writes or rewrites an impact sentence. A finding that was not assessed, or that lists platform-shipped records rather than scoring them, carries no sentence at all. Where a finding is shown in full — on the scan page, in the report, through the API, in chat and in a remediation ticket — it carries the same sentence.
Compensating controls
A connected scan also reads the controls that can stand between a finding and an attacker: IP address access control, REST API access policies, Explicit Roles, login and lockout properties, column encryption and OAuth client Auth Scopes. Each is read with the same discipline as a check's evidence — a declared predicate, a server-side count, the user the scan read as — and each is either verified present, verified absent, or not assessed with its reason. What it changes is what you see beside a finding: a “Compensating controls” line on the findings it applies to, an “entry restricted” note on an attack path whose entry an IP access control gates, and an “authentication type restricted” note where a REST API access policy covers the API itself — a policy restricts which authentication types may call an API, not who is authorised, so it never speaks to an operation on another API. What it never changes is the finding itself, its severity, its points, the band, its SLA, a control's state or an attack path's state. Ranges, policy names and users are never printed; the report shows counts and flags. A scan from an uploaded bundle does not read them.
Security Center corroboration
Where ServiceNow Security Center is installed, a scan also reads what it has recorded: the native scan runs and their dates, the checks and whether each is active, the findings of the newest completed run and which checks it executed, and its scores. The scripts behind those checks are never read, and a native scan is never started — running one stays your choice. What this changes is what you see beside a finding: a “Native scanner” line on a finding whose check overlaps a Security Center check, saying whether the native scan agrees, and a “Native scanner posture” panel with the coverage. What it never changes is the finding itself, its severity, its points, the band, its SLA or a control's state. A Security Center check counts as passed only when the native scan recorded running it and recorded no finding for it; otherwise the line reads “not recorded in the native scan”, never a pass. A native scan older than 90 days is stale and is shown as stale, not as current. Each line names the Security Center version its pairing was verified on, or says that this version is unverified. When Security Center is not installed, nothing is shown, and that absence is neither a finding nor a pass.
Related
What we read from your instance, who else is involved, and how credentials are held are a different subject with their own page: access and data handling.