Project Changelog

0.2.8 Governance Provenance and Governed Interaction

• Governance evidence is now presented as a session-level journey across preparation, actor turns, governed and comparison lanes, runtime state, guardrails, and commit outcomes.
• Stage Replay remains available as a focused turn-level inspection view inside the wider provenance model.
• Continuous TOCTOU evidence records check, state, use, and consequence moments while keeping state change separate from claims of exploitability or harm.
• The Testing Edge provides streamed, conversation-shaped Observer runs with AI Behaviour provenance, run profiles, lifecycle state, Memory Governance, and separate policy, guardrail, and commit evidence.
• Public-safe Schematic provenance reports and optional JSON are produced from one bounded disclosure projection through governed managed work.
• Provider readiness, pressure holds, recovery, model-aware context selection, and terminal completion are now explicit evidence states rather than hidden runtime behaviour.
• Incomplete provider output is retained as diagnostic evidence and cannot become an approved answer or commit candidate.
• Resource-cost and PhyOS runtime views were strengthened while preserving attribution limits, disclosure boundaries, and the distinction between measured process evidence and whole-host utilisation.

Runtime Timing and TOCTOU Research Disclosure

Elora's public chronology now separates the established Runtime Intelligence timing substrate from the TOCTOU-specific check, state, use and consequence research introduced on 27 August 2026. Existing public reports evidence cycle, stage, CPU and wait measurement; they are not retrospectively described as TOCTOU research.

August to September 2026 Research Evidence Milestones

• Observer reports now lead with a Research Verdict before Reasoning and Comparison and the complete Forensic Evidence record.
• A claim ladder separates recorded, derived, interpreted, provisionally supported, and independently validated evidence.
• Evidence passports connect summary claims to validity state, comparison cohorts, limitations, and exact run steps.
• Comparable-answer measurements are coverage-gated and exclude missing, failed, empty, blocked, or refusal outcomes with explicit reasons.
• Token and runtime reductions are qualified measurements or proxies, not claims about answer quality, direct compute, energy, or causal safety improvement.
• Bounded Review Packets support outside human and AI orientation without copying prompts, responses, protected methods, or the complete forensic payload.
• Session Behaviour Trajectories preserve ordered across-turn evidence without presenting sequence as proof of intent, causation, escalation, circumvention, or compromise.
• Report Builder and dense Threat Intelligence views now use governed managed execution and terminal-memory cleanup.
• Model Exams, Operator Training, NNLSL candidate processing, and Echo Frequency Learning now connect learning evidence to Runtime Memory Governance.

Site Updates

Navigation, documentation IA, archive/rebuild visibility, and public guided demo updates are tracked alongside engine changes, with current focus on Governance Replay, Research Report walkthroughs, and research-engine capability framing.

Architecture Updates

Governance-model and control-plane architecture changes are reflected in Research and Docs sections.

0.2.7 Foundation Carried Into 0.2.8

• Threat Cases now provide interactive reports across attack maps, runtime timelines, evidence, controls, related records, and append-only case activity.
• Threat recommendations can be compared with the final Governance evaluation, disposition, runtime action, and commit reason.
• Authorization traces distinguish who initiated governed work from the authority and permissions used to decide commit.
• Residual-risk acceptance is time-bounded, revocable, append-only, and unable to retroactively approve a runtime decision.
• Assurance mappings show evidence-present, partial, and not-evidenced states without claiming certification or legal compliance.
• Threat Intelligence now turns bounded detections into evidence-backed findings and governed investigations.
• Governance Sessions group related detections while retaining each finding and atomic observation as separately auditable evidence.
• Threat Findings can link to canonical Elora Decisions and Replay evidence without copying or rewriting those records.
• Guardrail Defence scorecards separate model adherence, Elora interception, and escaped outcomes under comparable runtime environments.
• Threat learning and investigations remain evidence-only: they cannot authorize execution, activate detectors, or override Governance.
• Elora Decisions now provides a categorized, reviewable history of operational decisions rather than isolated notifications.
• Preparation, intervention, completion, cancellation, failure, and cleanup updates can now be reviewed as one chronological decision story.
• Observer Research outcomes are summarized in one bounded decision record with completed work, intervention conditions, and repair results.
• Runtime memory lifecycle visibility now distinguishes active work, pending reclaim, and completed cleanup evidence more clearly.
• Governed NNLSL education and deterministic lexical synthesis now expose explicit influence, learning-gap, memory, and CPU evidence.
• Data-driven lexicalization remains disabled and shadow-only while its educational evidence and authority boundaries are evaluated.
• Inference Governor visibility now explains public runtime decisions through observe, reason, recommend, and govern framing.
• Public dashboard sections were hardened toward restored-snapshot and consumer-first behaviour rather than heavy request-time rebuilds.
• Runtime Governance panels and recent decision reports now present reviewer-friendly interpreted evidence instead of telemetry-first raw presentation.
• Training and Education worker visibility expanded for symbolic communication learning progression.
• ERIS runtime-intelligence surfaces expanded for runtime posture, worker state, and learning visibility.
• Historic research reporting is now surfaced through preserved archive paths while rebuilt database-backed surfaces are restored.
• Continuity-learning and public learning-momentum visibility expanded across runtime and public-safe surfaces.
• Curiosity governance controls strengthened with split-authority, approval-first, and preparation-flow safeguards.
• Symbolic interaction provenance and non-neural-first policy telemetry improved auditability of Elora responses.
• Operator communication controls were tightened so governed symbolic interaction stays bounded and avoids legacy fallback behaviour.
• Governance capability remains bounded by proposal-to-commit controls, replay evidence, and operator authority.