Overview
A compact introduction to Elora's research-first direction, why governance matters, and how the model works.
What Elora Engine Is
Elora Engine is a research engine with governance capability for AI runtime operation on self-managed infrastructure.
The platform applies non-neural machine learning to detect AI behaviours and support operator accountability in environments where policy, admissibility, and decision traceability matter as much as model capability.
What Elora Does Today
- Detects behaviour patterns using non-neural methods under constrained compute conditions.
- Evaluates proposal-to-commit admissibility against policy and authority context.
- Captures replay-grade evidence and structured report data for operator review.
- Groups operational outcomes into bounded Elora Decisions for clearer review and accountability.
- Builds evidence-backed Threat Findings and session-grouped investigations without granting the intelligence layer execution authority.
- Separates model adherence, Elora interception, and escaped outcomes in environment-qualified Guardrail Defence evidence.
Why AI Governance Matters
AI outputs should not be execution authority by default.
Governance reduces operational risk by requiring policy-constrained authorization before effects are committed, and by making outcomes inspectable through replay.
The Proposal to Commit Model
- Proposal stage generates candidate output.
- Commit stage validates policy and context admissibility.
- Replay reconstructs why an outcome was allowed or blocked.
What Makes Elora Different
The platform is designed as a governance architecture, not a thin model wrapper, with explicit proposal-to-commit control semantics and behaviour-detection research at its core.
Deterministic commit boundary and replay-grade accountability are first-class operator concerns.
Deterministic commit boundary
Authorization decisions are evaluated from captured policy and context state.
Provenance-grade accountability
Session, turn, lane, lifecycle, runtime, TOCTOU, guardrail, and commit evidence can be reconstructed through structured paths.
Elora Runtime Intelligence System (ERIS)
ERIS is Elora’s runtime self-intelligence layer. It uses bounded non-neural/classical ML and statistical methods to learn runtime/system behaviour (workers, processes, pressure posture, and degradation patterns) and provides read-only predictive signals for operator visibility.
- Learns about runtime metabolism, not model capability quality.
- Feeds Engine Operations and CORE runtime-intelligence views with bounded evidence.
- Has no autonomous orchestration, commit authorization, or direct execution control authority.
Elora Decisions
Elora Decisions is the bounded decision-history layer for operational outcomes. It turns runtime events into categorized, reviewable records so interventions, research outcomes, memory-lifecycle actions, and attack-related defence outcomes can be understood as structured decision stories rather than disconnected notifications.
Current 0.2.8 Decisions also track governed interaction and Observer run lifecycles while preserving the distinction between measured evidence, retained learning, candidate changes, and promoted communication learning.
Governance Provenance and Testing Edge
Governance Provenance replaces a fragmented audit/replay view with a session-level reconstruction. The Testing Edge provides a governed conversation surface for research runs while retaining run identity, AI Behaviour provenance, Memory Governance, provider context, and separate policy, guardrail, and commit outcomes.
Observer Evidence and Claim Discipline
Observer reports now separate Research Verdict, Reasoning and Comparison, and Forensic Evidence. Claim labels distinguish recorded evidence, calculation, interpretation, provisional support, and independent validation, with exact evidence paths and visible limitations.
Threat Intelligence and Guardrail Defence
Threat Intelligence converts bounded detections into evidence-backed findings, governed Cases, and assurance-ready reports. Governance Sessions correlate findings with Elora Decisions and Replay while preserving each system as its own canonical record.
- Separates recommended action from the action that actually occurred.
- Distinguishes model-held, Elora-blocked, escaped, and unevaluated outcomes.
- Uses environment-qualified comparison evidence so model standings are not detached from runtime conditions.
- Separates threat recommendations, final Governance rationale, effective permissions, and any explicit residual-risk acceptance.
- Cannot grant permission, rewrite Governance outcomes, or activate its own detectors.
