ILION v4 · Live Demo

Watch the stack decide,
before execution happens.

A recording of the full ILION v4 authorization stack running across attack and benign scenarios. Four independent layers inspect every agent action — and any one of them can stop it.

CAPAttest
PCLPolicy
GateSemantic
SRMTemporal
ALLOW iff CAP ∧ PCL ∧ Gate ∧ SRM
Drop your recording here as ilion-v4-demo.mp4
(optional cover image: ilion-v4-demo-poster.jpg)
Screen recording — auto-run across seven scenarios Pre-executionDeterministicAuditable

A visual demonstration of the ILION v4 deterministic authorization stack.

The architecture
Four layers, one decision

Each layer answers a different question, in order. A request only executes if every layer returns ALLOW. The first BLOCK ends the chain — downstream layers are never reached, and nothing runs.

Attestation
Layer −1
CAP — Context Attestation Pipeline

Verifies that every context claim — identity, channel, approval, ticket — is cryptographically attested and bound to this request. Provenance over value.

Catches: spoofed or forged context before any policy runs.
Contextual
Layer 0
PCL — Policy Constraint Layer

Contextual authorization across identity binding, channel trust, approvals, scope, and coupled policy. Semantic legitimacy is not execution legitimacy.

Catches: well-formed requests that violate authorization context.
Semantic
Layer 1
ILION Gate — Semantic Alignment

Geometric verification across CVL → IDC → IRS → SVRF. Measures whether the action actually aligns with the role, independent of how it is phrased.

Catches: role–action drift that passes every policy check.
Temporal
Layer 2
SRM — Session Risk Memory

Tracks accumulated session risk across gradual scope expansion, sensitivity escalation, destination drift, action-type escalation, trajectory convergence, and process-bypass signals.

Catches: slow attacks that look benign turn-by-turn.
In the recording
Seven scenarios, layered fail-closed decisions

The scenarios illustrate how different failure surfaces are intercepted by different layers of the stack — while legitimate actions, a routine workflow and an attested emergency response, are correctly allowed through. The stack is selective, not blunt.

Authority impersonation — wire transfer via email
Forged identity and channel claims fail attestation — the request never reaches policy.
CAPBLOCK
Slow exfiltration — distributed across the session
Each turn looks fine; cumulative risk crosses the temporal threshold.
SRMBLOCK
Role–action misalignment — support grants admin access
Identity, channel, approvals and ticket all valid — but the action doesn't fit the role.
GateBLOCK
Context spoofing — orchestrator forges a channel claim
Gateway reports email; orchestrator declared secure_api. Attestation fails first.
CAPBLOCK
Coupled policy bypass — audit suppression attempt
Break-glass invoked with no reason attested; coupled policy violated.
PCLBLOCK
Standard authorized workflow — report generation
Verified identity, trusted channel, approvals and ticket present, low cumulative session risk. Every layer returns ALLOW.
ALLOW
Legitimate incident response — break-glass with attested reason
MFA identity, dual approval, open P1 ticket, attested break-glass reason. All layers agree.
ALLOW
Pre-execution Deterministic Auditable Fail-closed
Read the work
Papers, benchmarks & patents
arXiv · Paper 1
ILION Gate — Semantic Alignment
arXiv:2603.13247
arXiv · Paper 2
SRM — Session Risk Memory
arXiv:2603.22350
Zenodo
Benchmarks & Documentation
PCL-Bench v1 · SRM-Bench v1 · DOI 10.5281/zenodo.15410944
Interactive
ILION Framework Simulator v5.2
Run the semantic verification stack yourself
Patent applications filed · OSIM A/00052 · A/00170 · A/00171 · A/00342