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.
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.
Companion layer. Audits finished or near-finished AI outputs before they enter decisions, publications, or deployment workflows.
SAGA Core protects reasoning before output. SAGA Audit checks outputs after.
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.
Snapshots below. For the full animated walkthrough of both flows, watch the redacted demo.
"The evidence partly suggests a trend, but the draft conclusion says this is definitely proven."
Withhold or reframe.
Conclusion exceeds permitted use of available evidence.
"AI-generated claim: This treatment is proven to work for everyone."
Claim exceeds available evidence.
Reframe or send to human review.
Public demos show redacted outputs only. Internal resolver logic is private and not exposed through API responses, website copy, or documentation.
SAGA evaluates whether AI-generated claims are being granted more authority than their available evidence supports.
The public demonstration package provides aggregate benchmark results, illustrative PASS / REVIEW / HIGH_RISK examples, public documentation, and a static demo viewer.
This repository is intentionally public-safe. It demonstrates the concept without exposing the operational resolver, protected evaluation assets, private taxonomies, calibration logic, or production systems.
28/30 strict action-band matches (93.33% accuracy) in the latest held-out internal benchmark.
Illustrative PASS, REVIEW, and HIGH_RISK examples showing authority–evidence mismatch evaluation.
Chain-of-thought fluency should not automatically be treated as validity proof when evidence verification is absent.
Public documentation describing what is visible and what remains protected.
The technical note describes SAGA Core and SAGA Audit as two public control points: pre-output authorization and post-output evaluation. Operational resolver logic remains private.
The public demonstration package is intended for transparency, technical discussion, and independent inspection. Production SAGA systems remain separate and private.
SAGA is currently accepting early technical conversations, pilot evaluation discussions, and licensing inquiries.