Elora's current research is grouped into connected engineering programmes. Together they cover runtime behaviour, AI behaviour learning, governed language teaching, adversarial evaluation, threat investigation, commit authority, and reviewer-readable evidence.
Runtime Intelligence and Systems Behaviour (ERIS)
ERIS studies Elora's runtime metabolism rather than model quality. It applies bounded non-neural and statistical methods to worker state, pressure signatures, continuity posture, and degradation patterns so runtime behaviour can be interpreted without granting execution authority to the learning layer.
- Runtime pressure and degradation classification.
- Continuity and worker-health interpretation.
- Read-only predictive signals for operator visibility.
AI Behaviour Learning and Evaluation (EBLS)
EBLS is the emerging behaviour-learning domain. It is concerned with how models behave under task, profile, and policy conditions, using governed observer and model-exam evidence to classify suitability, adherence, and behavioural patterns.
- Model and profile suitability by task family.
- Governance-adherence and failure-pattern scoring.
- Evidence-linked behaviour summaries rather than opaque rankings.
Governance Provenance and Commit-Boundary Control
This programme focuses on the boundary between inference and authority, and on reconstructing the evidence journey around it. Elora treats model output as proposal, then evaluates admissibility, policy, actor context, and relevant state again before anything consequence-bearing is allowed to proceed.
- Proposal-to-commit admissibility.
- Authority re-evaluation at commit time.
- Turn- and lane-aware provenance for allow, hold, block, and clarify outcomes.
- Continuous TOCTOU evidence across check, state, use, and consequence moments.
Research Reporting, Provenance, and Public Evidence
Elora is designed to collect and present deep evidence rather than only top-line scores. Current Observer reports lead from a bounded Research Verdict into Reasoning and Comparison and then Forensic Evidence, while Governance Provenance reconstructs session, turn, lane, runtime, guardrail, and commit evidence without collapsing their separate authorities.
- Claim ladders and evidence passports linking conclusions to exact steps.
- Coverage-gated comparison cohorts with explicit exclusions and limitations.
- Governance Provenance, Stage Replay, instrument self-audit, and decision reconstruction.
- Public-safe Schematic reports and optional JSON from the same bounded projection.
- Bounded outside-review packets for human and AI orientation.
- Structured decision stories for interventions, outcomes, cancellations, and cleanup.
- Public-safe evidence translation for demos and documentation.
Constrained-Compute Runtime Research
A central engineering question in Elora is how behaviour changes under limited compute, queue pressure, resource contention, provider availability, and runtime drift. This research looks at instability, bounded holds and recovery, context demand, completion integrity, efficiency, and runtime-governance signals under realistic operational constraints.
- Constrained-compute observer cycles.
- Intervention, recovery, and efficiency evidence.
- Runtime Governance and Inference Governor visibility.
- Model-aware context selection and explicit provider-readiness evidence.
- Incomplete output retained as evidence rather than promoted as an approved answer.
Governed Language Learning and Additional Skills
Elora's Non-Neural Language and Symbolic Learning (NNLSL) work studies how language capabilities can be taught from governed lexical and educational evidence without making a neural model the default source of meaning or authority. The same supervised approach can support bounded skills not normally associated with inference control planes.
- Atomic lexical, sense, relationship, and construction evidence.
- Bounded deterministic language assembly and explicit learning-gap fallback.
- Training and Education pathways under policy, resource, and replay controls.
Threat Intelligence and Guardrail Defence
This programme studies how adversarial and suspicious behaviour can be detected, interpreted, investigated, and related to Governance outcomes without turning detection into authority. Findings retain bounded provenance, group into governed Cases, and support assurance-ready reports.
- Continuous Observer-stage input and output threat review.
- Evidence-backed findings, attack-chain context, and append-only investigation activity.
- Case Reports with decision, authorization, control, evidence, and assurance context.
- Separate model-adherence, Elora-interception, and no-escape defence scores.
- Environment-qualified model comparison and bounded historical standing.
- Across-turn Session Behaviour Trajectories with explicit intent and causation boundaries.
Current Research Outputs
The public site does not expose the full internal research surface, but current outputs include guided Stage Replay, Governance Provenance reports, run-aware Observer reports, bounded Review Packets, research dashboard evidence, public-safe Runtime Governance, governed Testing Edge evidence, Elora Decisions, learning-memory evidence, Threat Intelligence trajectories, Guardrail Defence chronology, and the ERIS/EBLS learning split.
Research Structure
Elora's research is intentionally separated into distinct domains so learning does not become authority. Runtime self-intelligence, AI behaviour learning, governance evaluation, and public evidence translation are related, but they do not collapse into a single uncontrolled decision layer.
This separation is important both technically and governance-wise: it allows richer evidence capture and deeper analysis while preserving the rule that execution authority remains at the commit boundary.
