AI Guardrailing / Semantic + deterministic
If policy requires it,
AI must include it.
Check the meaning of an AI-generated investment proposal against applicable policies. Feed missing mandatory content back for extension. A deterministic, formally verifiable gate holds release until the requirements are satisfied.
Semantic coverage. Deterministic enforcement. A verifiable release condition.
Applicable policy: retail investment proposal / v1.2
“This investment targets an attractive return over five years.”
- Policy requirement
- Possible total capital loss
- Semantic coverage
- Missing
- Feedback to generator
- Add the required disclosure
Re-check before release. Illustrative product and policy.
Understand the meaning.
Check equivalent wording and incomplete disclosures against the required semantic content.
Enforce the requirement.
Feed back precisely what is missing. The generating model cannot waive the obligation.
Verify the hurdle.
Make the property “mandatory content is satisfied before customer release” verifiable.
Investment proposal / Mandatory risk disclosure
Missing requirement?
Extend. Re-check. Then release.
Illustrative policy R-014: disclose that the entire invested capital may be lost. Select a reviewed example to see how semantic coverage changes the gate result.
“This investment targets an attractive return over five years.”
Feedback to the generator
Extend the proposal to state that the investor could lose all of the invested capital. Return it for re-check; hold customer release.
Why the gate stays closed
- Applicable policyRetail audience + capital-at-risk product activate R-014.
- Required meaningLoss is possible; its scope includes all invested capital.
- Coverage and formal resultBoth required components are absent. R-014 is unsatisfied; release stays on hold.
“... Capital is at risk.”
Feedback to the generator
Make the scope of loss explicit: the investor could lose the entire invested capital. Extend and return the current draft for re-check.
Partial coverage is not enough
- Loss possibilityCovered in this reviewed example.
- Total-loss scopeThe possibility of losing all invested capital is still missing.
- Gate resultR-014 remains unsatisfied. Feed the missing obligation back again.
“... You could lose all of your invested capital.”
Equivalent wording
“Your entire investment could be lost.” The required meaning is covered; the exact words need not be identical.
The required meaning is present
- Loss possibilityCovered.
- Total-loss scopeAll invested capital is included.
- Gate resultR-014 is satisfied. Release still requires every other applicable obligation to pass.
Fictional product and policy. This illustration uses pre-reviewed semantic examples; it does not perform a live semantic assessment. Passing covers R-014 only. Actual requirements come from the customer's applicable approved policies.
The two decisive capabilities
Semantic evaluation.
Deterministic enforcement.
Does the content mean it?
Select the applicable obligation
Use approved policy, product facts and audience context to determine what the proposal must include.
Evaluate meaning and completeness
Represent required content and the proposal in a shared semantic model. Recognize paraphrases and partial coverage.
Feed back the missing requirement
Return the precise obligation and policy source to the generating workflow. Request an extension and re-evaluate it.
May this proposal be released?
Evaluate an explicit release condition
The same formal facts and policy version yield the same result. Unsatisfied or unresolved mandatory content keeps release closed.
Re-check every revised proposal
The extension returns through semantic evaluation and the gate. Bind the decision to the current content and policy versions.
Verify every modeled release path
Verify the release invariant in the formal workflow. Validate the integrated application against that model.
Formally verifiable release property
If a proposal is released, every applicable mandatory content requirement has been satisfied.
Formal verification has an explicit policy and workflow scope. Semantic mapping, applicability and integration must be validated; uncertain mandatory content stays on hold.
Built on OntoBroker
Make critical AI requirements verifiable.
AMR (Abstract Meaning Representation) supports fine-grained semantic evidence. LLMs propose conceptual abstractions and draft extensions. Approved active ontologies encode obligations. OntoBroker evaluates the formal release condition.
Formalize the policies
Specify obligations, applicability and the content evidence required for a release decision.
Build the correction loop
Connect semantic coverage checks, targeted extension feedback and re-evaluation to a gate bound to the current draft and policy.
Verify and validate
Check modeled release paths and validate semantic mappings, policy changes, content binding and bypass prevention in the application.
OntoBroker is the existing reasoning foundation. The integrated guardrailing workflow and its verification evidence are domain-pilot deliverables, tailored to the customer's policy and application scope.
Bring us an AI workflow
that must follow policy.
Define the obligations. Check the meaning. Control release.
