L1 · Instrument / measure
Capture internal readings during inference in an instrumentable model.
Inference Assurance for AI Models
From measurement to assessment to policy
Phi-Lattice provides inference assurance for AI models: it measures internal model signals, produces risk-specific assessments, and retains evidence you can inspect.
Built for organizations that run their own models.
Phi-Lattice requires a model you run and can instrument: open-weight, fine-tuned or proprietary models where the required internal readings are accessible. Hosted frontier APIs generally do not expose the internal state Phi-Lattice needs.
Open weights alone do not guarantee compatibility; we confirm the serving environment and authorized access during evaluation.
On-premises GPU capture during generation and CPU replay have been demonstrated. Air-gapped or disconnected integration is assessed with design partners.
Phi-Lattice measures. Your policy decides.
The calibrated assessment informs your policy; it does not authorize action. Your observability, evaluation and governance systems retain their roles.
Application / Agent
MODEL INFERENCE UNDER YOUR CONTROL
Model execution
Internal computation, not just the output
Internal readings
Phi-Lattice
L1 · Instrument / measure
L2 · Assess
Evidence · Retain / replay
Authorized instrumentation captures internal readings during execution. The assessment is calibrated to your risk and reference.
Phi-Lattice assessment
Your policy / governanceL3 · Your policy
Proceed / Review / Hold / Investigate
Phi-Lattice evidence
Risk / model risk / compliance / auditInspection and reassessment when the required inputs are retained
Existing systems retain their roles: Observability · Evaluation · Governance
Design-partner integration: your policy and action workflows. Phi-Lattice provides evidence; you grant authority and control execution.
Demonstrated today
L2 · Assess
Convert selected readings into a risk-specific assessment calibrated against a reference. An assessment describes a state, not an action.
Verdicts: No Issues, Noisy, Caution, Suspicious, Major Issues.
Evidence · Retain / replay
Keep the inputs needed to inspect, verify or recompute a defined operation later without calling the model again.
Scoped design-partner implementation
L3 · Your policy
Make the assessment available to your policy or governance workflow. You decide what it means operationally and retain control of execution.
Two implementation options
- API integration with your existing policy system.
- A lightweight Phi-Lattice policy configuration tool, scoped with the design partner.
What does a Noisy assessment mean in terms of action? That depends on your policy. One organization routes it to human review. Another treats it like No Issues.
Demonstrated today
Evidence
Replay re-checks a past assessment from the evidence Phi-Lattice kept, without running the model again.
Depending on what was recorded, that can mean recalculating the score, testing a different policy setting on the same case, recomputing a measurement from the saved internal readings, or recalculating the difference between two recorded cases. It does not regenerate the original answer.
Scoped design-partner implementation
Inspect the supported replay operations and their limits Customer-system integration remains scoped design-partner work.