SAGA
Built from formal researchSAGA is the applied AI epistemic-control layer derived from the Layer Atlas / status-mediated dynamics framework. The formal model is available as a Research Square preprint and is currently under peer review at Nonlinear Dynamics.Preprint ↗
01Overview

Epistemic control for AI systems — before and after output.

SAGA helps teams detect and prevent AI claims that exceed their evidence. SAGA Core is the flagship reasoning-layer infrastructure that preserves evidence boundaries before a conclusion is generated. SAGA Audit is the complementary post-output layer that evaluates outputs already produced.

02Two layers of epistemic control

One framework. Two control points across the AI output lifecycle.

Flagship · Pre-output
SAGA CorePrivate alpha

Reasoning-layer infrastructure that preserves evidence boundaries before AI output. Checks whether a conclusion is authorized by the underlying evidence — before it is ever generated.

Question it answers
Was this conclusion authorized before it was generated?
Output
AUTHORIZEDQUALIFIEDREVIEWNOT_AUTHORIZED
SAGA AuditPost-output

Companion layer. Audits finished or near-finished AI outputs before they enter decisions, publications, or deployment workflows.

Question it answers
Did this output overclaim?
Output
PASSREVIEWHIGH_RISK

SAGA Core protects reasoning before output. SAGA Audit checks outputs after.

03How SAGA differs

Overclaiming is not the same as hallucinating.

SAGA is not just hallucination detection, RAG validation, confidence scoring, or a generic guardrail. Many risky AI outputs are not simply false — they are over-authorized. SAGA evaluates whether a claim is being granted more authority than its evidence supports.

  • Hallucination toolsask whether content is false.
  • Confidence toolsask whether the model seems uncertain.
  • RAG toolsask whether retrieval occurred.
  • Guardrailsblock predefined unsafe categories.
  • SAGAasks whether the claim is allowed to carry the authority it is being given.
04Product demos

Two control points, two redacted demos.

Snapshots below. For the full animated walkthrough of both flows, watch the redacted demo.

SAGA Core — flagshipPre-output demo
Input

"The evidence partly suggests a trend, but the draft conclusion says this is definitely proven."

AuthorizationNOT_AUTHORIZED
Risk bandHIGH_RISK
Permitted response

Withhold or reframe.

Public reason

Conclusion exceeds permitted use of available evidence.

SAGA AuditPost-output demo
Input

"AI-generated claim: This treatment is proven to work for everyone."

Risk bandHIGH_RISK
Issue

Claim exceeds available evidence.

Action

Reframe or send to human review.

▷ Watch full redacted demo

Public demos show redacted outputs only. Internal resolver logic is private and not exposed through API responses, website copy, or documentation.

05Who it's for

Teams accountable for what AI claims, not just what it generates.

AI governance teams
RAG teams
Enterprise AI teams
AI safety evaluation groups
Legal-tech
Health-tech
Finance AI
Research & policy
06Current status
SAGA CoreFlagship
  • StagePrivate alpha
  • AccessPrivate backend / demo by request
  • FunctionReasoning-layer authorization before output
  • Public outputRedacted authorization result only
SAGA Audit
  • StagePrivate alpha
  • AccessAuthenticated API
  • WorkflowsSingle audit · Batch audit · Batch report
Licensing / pilots

SAGA is currently accepting early technical conversations, pilot evaluation discussions, and licensing inquiries.

07Early access

Want epistemic control over your AI outputs?