The Elora Taurus Project - Elora Learning System

Elora uses two learning domains. ERIS learns runtime-system behaviour and pressure posture using bounded non-neural/classical ML and statistical methods. EBLS learns AI model behaviour, governance adherence, and model/profile suitability from governed evidence.

ERIS and EBLS Boundary

ERIS (Runtime Self-Intelligence)

ERIS learns runtime metabolism: worker/process/container state, pressure signatures, and degradation/anomaly posture. Output is read-only runtime intelligence for operator visibility and never direct execution authority.

EBLS (Behaviour Learning)

EBLS learns model behaviour and governance outcomes: suitability, adherence, pattern outcomes, and observer/model-exam learning evidence across task/profile contexts.

How ERIS Learns

Sense: capture bounded runtime signals (state, pressure, continuity events).
Classify: map patterns into deterministic runtime posture categories.
Aggregate: build longitudinal runtime learning windows and profile summaries.
Surface: expose read-only intelligence to Engine Health and CORE views.
Constrain: learned runtime signals do not autonomously execute control actions.

Classical ML Families in ERIS

Rolling-window statistical baselines and threshold bands.
Weighted signal fusion and deterministic posture scoring.
Trend and profile comparison across bounded runtime windows.
Anomaly/degradation pattern detection under explicit policy constraints.

What ERIS Learns

Runtime Pressure Posture

Queue, CPU, RAM, and saturation signatures are tracked to identify unstable pressure conditions and bounded degradation risk.

Worker and Process Continuity

ERIS learns how runtime components behave over time so continuity patterns and failure posture are visible earlier.

Anomaly and Degradation Patterns

Repeated runtime degradation signatures are accumulated as traceable learning context for operator review.

Predictive Runtime Signals

Deterministic predictive posture signals are produced for visibility, without granting orchestration authority.

How EBLS Learns (In Development)

Collect: capture governed observer/model-exam and policy-control evidence.
Score: compute adherence, suitability, and behavioural pattern outcomes by model/profile.
Link: connect outcomes to task families and governance-control familiarity signals.
Recommend: provide model/profile guidance to operators under policy-gated controls.
Constrain: deterministic commit-boundary controls remain final authority.

Planned Expansion

Dynamic Observer Questions

Observer prompts will adapt by scenario family and policy-control focus so behaviour detection quality improves over repeated runs.

Governance Training Tracks

Dedicated scenario packs will compare guardrail methods and score control adherence consistency per model and profile.

Engine-Only Readiness

Readiness scoring will measure whether Elora can detect and classify breaches with reduced or no external guardrail scaffolding.

Additional Skill Research

Research expands into additional skills not typically associated with inference control planes, while retaining deterministic evidence, policy boundaries, and replay auditability.

What It Does Not Do

Does not train model weights.
Does not bypass policy or commit controls.
Does not hold autonomous orchestration or commit authorization authority.
Does not use unconstrained or random routing in governed execution paths.

Model Wiki (EBLS Preview)

Model Wiki is a public-safe preview of EBLS outputs: model/profile behaviour summaries, adherence posture snapshots, and evidence-linked strengths/weaknesses.

Live Model Wiki Preview

Coming Soon: staged public EBLS preview is not live yet.

Open Sample Model Wiki Report

Sample report format is available now as a public-safe EBLS output preview only.