Can AI change a control determination in CertiField?

No. AI output is always a proposal. It cannot change a control, an assessment, a POA&M item or a compliance determination. Only an authorized practitioner accepting the proposal writes to the authorization record, and the acceptance is audited.

What stops the model inventing a control identifier?

Server-side allowlists. Any control, CCI or STIG identifier the model returns that does not exist in the catalog is discarded before it reaches a user. Architecture proposals are validated against a closed taxonomy, and an unrecognized zone type or component role is dropped rather than coerced into the nearest match.

Why CertiField · Human-governed AI

AI assists. Your people decide.

Every vendor in this market now has an AI badge. The question a government buyer should actually ask is narrower and much harder to answer well: what, exactly, is the machine allowed to change?

AI proposes
Evidence supports
Practitioners decide
CertiField preserves the record

Ref: W-11
Kind: Invariant
Rule: no machine writes a fact
Enforced by: test, not policy

The rule the product is built around.

No importer, parser or model output mutates authoritative data.

Everything a machine produces in CertiField — a control implementation extracted from your existing SSP, a proposed architecture component, a STIG applicability ranking, a suggested control mapping — is a proposal. It carries its own explanation, including the quote it was drawn from where there is one. It stays a proposal until a person with the authority to accept it does so, and the acceptance is audited like any other authoritative act.

AI-derived architecture implications are explicitly advisory and cannot change controls, assessments, POA&M items or compliance determinations. That is not a setting. There is no code path.

The SSP Intelligence screen. A banner states that the SSP generator and its AI assist never write authoritative data, that ingestion only stages proposals, and that nothing becomes part of the record until a human accepts it. A legend distinguishes deterministic extraction, where a named rule matched the source, from AI-assisted extraction, which is unverified until accepted. AI assist is switched off by default. — The product states the rule on the screen where the work happens, not only in a policy document — and AI assist is off by default, running only where the deterministic rules could not resolve a section. System · Documentation · SSP Intelligence

Ref: W-12
Kind: Guards
Fencing: user content delimited
Filtering: server-side allowlists

What the model is structurally unable to do.

Invent an identifier

Control, CCI and STIG identifiers returned by a model are checked against the catalog server-side. An identifier that does not exist is discarded before anyone sees it, so a plausible-looking hallucination never becomes something a practitioner has to disprove.

Assert a security judgment

Architecture proposals are schema-validated against a closed taxonomy. The model cannot assert a boundary disposition, an encryption state or a data classification — those are the calls that decide an authorization, and they belong to people.

Take instructions from your documents

User content is fenced inside explicit delimiters on its way into the model, so text inside an uploaded SSP or scan report is treated as data to analyze rather than as instructions to follow.

Overwrite a newer human edit

A proposal records a hash of its target at the moment of extraction. If the target changed in the meantime, acceptance is refused as stale rather than clobbering the more recent decision.

Ref: W-13
Kind: Use
Feature switches: per capability

Where AI genuinely earns its place.

None of the above means AI is decorative here. It does the work that is genuinely tedious and genuinely low-stakes: reading a two-hundred-page SSP and finding the paragraph that discusses AC-2; ranking which of four hundred STIGs plausibly apply to a system running these technologies; drafting a first implementation statement from evidence a practitioner then corrects; noticing that a component moved and suggesting which controls that might touch.

Each capability has its own switch, subordinate to a master switch that can turn the lot off. And when it is off, it is off — no feature quietly keeps calling a provider that is not there.

The version of this that matters to a government buyer

"Powered by cutting-edge AI" is not a procurement argument. "The AI cannot change a determination, and here is the mechanism that makes that true" is.

How provenance is preserved →

Next step

See what your RMF process looks like when the package keeps up with the system.

Request a demo Explore the product