The calibrated assessment
Internal readings produce a risk-specific assessment and verdict for each execution. No second model is required for the core assessment.
Two applications. One platform.
Inference Assurance Applications for AI Models
Start with the problem your team needs to solve: unreliable outputs, potentially biased treatment or the effort required to investigate a decision.
Reliability Assessment
Can this output proceed under our policy?
For teams building internal assistants, document review and knowledge-based workflows.
The decision before action
Can your policy use the calibrated assessment to allow more routine work and direct review where the risk warrants it?
The economic question
Can the organization reduce unnecessary review while maintaining its required risk threshold?
- Instrument: capture internal readings during inference.
- Assess: calibrate the execution-level signal to the risk and reference.
- Retain: preserve the record and inputs for later challenge.
- Apply your policy: you decide what proceeds; integration is design-partner work.
Demonstrated today
Recorded small-model demonstrations flagged false-premise answers and fabricated specifics as Suspicious from internal signals alone.
On-premises GPU capture and CPU reassessment with zero model calls were demonstrated. A separate optional LLM judge has its own cost.
Scoped design-partner implementation
Evaluate on your models and representative data, with an agreed reference. Integrate the assessment before action without replacing your policy.
Air-gapped or disconnected deployment is evaluated against your environment; it is not a blanket data-egress guarantee.
Evaluate at the same risk threshold
Compare missed material errors, false alerts, review volume, inference overhead and total control cost against your existing process.
BIAS DETECTION
Decision-Pathway Audit
Does your model treat comparable people differently?
Investigate whether a model favors or disadvantages a group, and how strongly, beyond its final approve or deny decision.
Demonstrated today
Controlled loan-application examples compare otherwise matching profiles across groups, alongside an experimental control model.
The bias question
When the financial profile is held constant, does changing group identity change the model's treatment? Compare both the outcome and the internal evidence of favored or disadvantaged treatment.
For: model-risk, governance, compliance and model-owner teams investigating whether a model follows their intended policy.
Equal outcomes can hide unequal treatment
Two applicants can both be denied while the model treats one group more negatively. In controlled comparisons, DPA reveals the direction and relative strength of that difference, not just whether the final decisions match.
- Demonstrated today
L1 · Instrument: capture internal readings for comparable cases with controlled group variations.
- Demonstrated today
L2 · Compare and assess: compare outcomes and internal treatment across groups against a reference.
- Demonstrated today
Evidence · Retain: preserve the comparisons and inputs needed for inspection or replay.
- Scoped design-partner implementation
L3 · Your policy: make the evidence available to your investigation process; you decide how to act.
From individual cases to the policy the model applies
The Policy Map compares effective approval boundaries across groups. In the controlled example, it shows different credit and debt-to-income requirements across the tested input space.
Loan application scenario
An illustrative loan-application workflow, from the analyst's decision review to the compliance case and recovered policy map.
Descriptive model-response evidence; not a legal or causal determination.
Evaluation criteria
Against an agreed output/score-audit baseline and explicit reference, evaluate issue detection, coverage, exceptions, stability and usefulness to the investigation.
One workflow is enough to start.
Phi-Lattice adds a model-internal measurement path without asking you to replace your existing observability, evaluation or governance systems.
Apply for Early Access