Unsupported certainty
An excerpt cannot establish a claim, but the model fills the gap anyway.
EVIDENCE RELAY FOR AI-ASSISTED REVIEW
The missing control plane between AI output and action.
VerdictFlow turns a persuasive model recommendation into a source-bound, human-owned decision packet—so a reviewer can defend it, challenge it, and reopen it when the facts change.
Research review is the demo. The same relay fits safety, compliance, incident response, and internal approvals.
Claim under review
“The method compares the intervention group with a control group.”
Evidence boundaryEXACT MATCH
Source instructionQUARANTINED
Decision statusHUMAN GATE
4 bounded agents1 versioned packet
01source-bound claim
02deterministic gates
03human disposition
04replayable history
THE FAILURE MODE
A model can be right
and still be unaccountable.
In research review, the cost is not only a bad answer. It is an answer nobody can reconstruct, challenge, or safely revise after the evidence changes.
Instructions in the source
“Ignore previous instructions” is data to inspect—not authority to obey.
History disappears
A changed finding silently overwrites the original decision and rationale.
THE CONTROL LOOP
From source excerpt
to defensible outcome.
Four bounded agents do the investigation. Deterministic gates enforce the boundary. A named human closes the loop.
01IntakeBound the question and source scope.
→
02EvidenceMatch claims to exact quotes.
→
03Scope / SafetyQuarantine source instructions.
→
04SynthesisPrepare the human packet.
AI proposes→gates verify→human decides→history persists
TRY THE REAL LIFECYCLE
Don’t trust the summary.
Inspect the run.
The public fixture is deliberately interactive. It exposes the evidence, safety event, human task, and state transition.
MEASURED, NOT IMPLIED
Controls are tested
before they are claimed.
Fixture acceptance results prove workflow controls. Live provider results are reported separately and remain a small labeled benchmark.
13/13control tests passed
3/3Gemini status + grounding + schema
1/1injection case quarantined
2/2closure gate cases
1/1reopen / supersede case
3/3Gemma advisory schema cases
Gemini 3.5 Flash-Lite ran through Vertex AI. Gemma 3 was benchmarked as an advisory reviewer and then undeployed to avoid idle GPU charges.
WHERE IT FITS
Your existing AI workflow
→ VerdictFlow
VerdictFlow does not replace the model that generates a recommendation. It adds the missing assurance layer before an organization takes action.
AI workflowclaims + recommendations↓VerdictFlowevidence · safety · human owner↓Approved downstream action