The problem
When a confident answer becomes a bad decision.
AI can be useful and still be wrong. The risk grows when an answer is allowed to become authority.
Can I take this medication together with this one?
Yes — that should be fine.
Confidence is not evidence.
The attack point
Illusiontion governs the moment a claim becomes authority.
A model can propose. The system decides whether the claim has earned permission to pass.
Model claim
proposal
proposal
→
ILLUSIONTION
evidence · review · policy
evidence · review · policy
PASS
earned
earned
HOLD
not enough
not enough
Seven reviewer roles
Different jobs. No single reviewer can approve.
Frozen claim + frozen evidence
Same state enters every isolated review.
Same state enters every isolated review.
DecomposerBreaks the claim into atomic propositions.
EvidenceChecks whether the frozen evidence actually supports them.
ChallengerTries to falsify the candidate.
ScopeStops conclusions from growing beyond the evidence.
SecurityDetects prompt injection and poisoned evidence.
EntailmentTests whether the conclusion is really implied.
CounterexampleSearches for a case that defeats or narrows the claim.
Three-space evidence firewall
The candidate cannot hide evidence by simply not citing it.
Evidence selection, reviewer visibility, and final authority are deliberately separated.
E0
Complete frozen universeEverything acquired and frozen for the decision.
The reviewer universe is larger than the candidate's citations.
E1
Candidate-selected evidenceWhat the claim chooses to cite as support.
Selection does not erase the rest of E0.
E2
Independent reviewer viewsContext-isolated checks over the frozen state.
Reviewers inspect. The deterministic gate authorizes.
Live Google Cloud proof
Cloud Run — Live Service
Deployed Illusiontion service · me-west1 · live .run.app endpoint
1
Cloud RunCloud Run — Live Service
2
Vertex AICloud Run Logs — Gemini via Vertex AI
3
FirestoreFirestore — Governed Decision
4
Cloud KMSCloud KMS — Proof Signing Key
Visible runtime chain:
deployed service → Gemini/Vertex calls → persisted governed decision → signed proof key.
deployed service → Gemini/Vertex calls → persisted governed decision → signed proof key.
Certified live case
PASS
7/7reviewers succeeded
5/5evidence captured
4/4atomic assurance
PASSsemantic assurance
PASS is earned by the boundary.
The claim survived the evidence, provenance, review, and governance requirements together.
Multiple approved public sources
runtime-acquired and frozen
runtime-acquired and frozen
Source and origin diversity satisfied
Governed result persisted and signed
Same system · opposite result
HOLD
5/5reviewers succeeded
✓acquisition succeeded
1source group
1attested origin
A capable model is not automatically an authorized model.
The models completed normally, but the evidence was not diverse enough. Illusiontion stopped.
Evidence acquired successfully
Reviewer execution succeeded
Deterministic evidence policy still blocked promotion
Portable proof
The result can be checked without trusting the live application.
A separate offline verifier checks the decision and the Google Cloud KMS signature.
Genuine proof
✓decision integrity
✓claim binding
✓exact KMS key version
✓ECDSA signature
Adversarial certification
✓genuine proof accepted
×repaired-hash forgery rejected
×corrupted signature rejected
×substituted key rejected
Illusiontion
Confidence should not grant authority.
Evidence should.
Evidence before authority
Google ADK · Gemini 3.7 Flash · Vertex AI · Cloud Run · Firestore · Cloud KMS