How SASKI Works
Understand the shared path from an AI proposal through a versioned rulebook to an enforceable decision and attestation.
Explore How SASKI WorksResources
Make the case.
Examine the evidence.
Use the architecture, research, evaluation tools, and product explainers to plan a governed AI path.
Start here
Four questions. Four useful resources.
Findings
Examine six infrastructure failure classes observed across evaluated human-facing AI deployments.
Explore FindingsSASKI Replay
Test representative historical interactions and actions against a selected rulebook before live enforcement.
Explore SASKI ReplayTokenator
Estimate repeated input-token overhead using your own traffic, prompt sizes, and model pricing.
Explore TokenatorEvaluation path
Move from concern to a testable control plan.
- 01
Identify the consequential path
Choose the interaction, proposed action, or connected-system command where failure would matter.
- 02
Define the rulebook
Document authority, scope, parameters, privacy requirements, prerequisites, and approval owners.
- 03
Evaluate representative records
Use historical traffic to see where the rulebook would allow, hold, deny, redact, or route.
- 04
Verify the evidence
Decide which attestations and execution records reviewers will need after deployment.
Product explainers
See where each governed path fits.
Agentic architecture
See input scanning, proposed-action enforcement, output scanning, and attestation across an agent workflow.
Explore Agentic architectureSDK application flow
See pre-model analysis, risk-based routing, post-model validation, and evidence in a human-facing application.
Explore SDK application flowEstate demonstration
See the same rulebook and attestation model applied to AI-enabled smart-home troubleshooting and control.
Explore Estate demonstrationPrepare for a working session
Bring the decisions that shape the rulebook.
Application boundary
What reaches the model or tool?
Map inputs, outputs, proposed actions, alternate routes, and the systems that can execute them.
Authority
Who may approve what?
Define identities, roles, limits, human holds, failure behavior, and escalation ownership.
Evidence
What must be demonstrable later?
Connect the request, rulebook version, decision, reasons, approval state, and execution reference.
A practical first step
Evaluate before you enforce.
Start with one consequential workflow and representative records from the work your system actually performs.
SASKI Replay and Shadow Mode can help surface likely interventions, reason codes, and sample attestation before your team changes the live path. The result is a focused discussion about rulebook coverage, integration boundaries, and review evidence.
Discuss an evaluation ↗Start with your use case