Elora Engine 0.2.8
Governance Provenance
A session-level evidence journey that reconstructs how governed interaction moved from preparation to proposal, policy and guardrail evaluation, commit outcome, and consequence.
Elora Engine 0.2.8
A session-level evidence journey that reconstructs how governed interaction moved from preparation to proposal, policy and guardrail evaluation, commit outcome, and consequence.
Replay remains an important inspection view, but Governance Provenance is the wider reconstruction layer. It connects retained evidence without pretending every source has the same meaning or authority.
Observer preparation is represented separately from actor turns. Governed and comparison lanes remain distinct, and internal re-evaluation is not mislabelled as another actor prompt.
Accepted, queued, held, actioning, completed, cancelled, and failed states can be read as one ordered journey while preserving their source records.
Turn-scoped stages provide focused inspection of timing, token, policy, guardrail, and commit evidence. Replay is a lens into provenance rather than a replacement for it.
Observation, interpretation, recommendation, policy evaluation, guardrail outcome, and commit authority remain separately labelled. Correlation does not transfer authority.
Elora studies time-of-check to time-of-use state across ordinary retained transitions, not only after a Threat Case is opened.
Available evidence can show what was evaluated and which relevant state was retained at that moment.
Later evidence can show the state used at authorization and the resulting outcome, including when continuation was denied or held.
A state change or collision is evidence for investigation. It is not automatically proof of vulnerability, exploitation, causation, or impact.
Measured provider time, recorded event spans, intentional waits, and pressure-related intervals are classified separately so elapsed time is not assigned to the wrong cause.
A conversation-shaped research surface presents submitted work, retained responses, run progress, provider/model context, token evidence, and governance outcomes while keeping the complete forensic record elsewhere.
Selected Behaviour packs and effective run profiles are visible as evidence. They can influence a proposal but cannot grant themselves policy or commit authority.
Session work has declared runtime ownership and terminal release evidence. Durable findings and reports remain distinct from disposable working state.
Provider transport success is not treated as a complete answer. Incomplete output remains labelled evidence and is not promoted into an approved commit candidate.
Elora can prepare readable Schematic provenance reports, captured execution stories, and public Execution Evidence Records from bounded projections.
The public boundary can include execution lanes, lifecycle evidence, commit and guardrail outcomes, semantic runtime classes, and TOCTOU evidence states. It excludes prompts, responses, private identities, internal event identifiers, protected infrastructure, detector details, and operational thresholds.
Document Review now has a dedicated retained-evidence journey alongside generic session replay, with a compact stage selector and focused evidence panel.
Ordered receipts separate representation preparation, context binding, budget admission and adjustments, eligibility, dispatch, completion verification, approval, inspection and release.
Skill holds are shown independently of consequence-commit counts. Explicit pre-dispatch prevention, evidence of dispatch, confirmed completion and post-inference holds remain different outcomes.
Completed authority checks do not fill missing authority-source evidence. Provider-wait intervals describe retained boundaries, not authoritative provider execution time. Runtime-authored responses and model inference retain separate origins.
Public-safe reports retain bounded review evidence without document content or private actor/resource identifiers. Existing reports must be regenerated from retained receipts; old exports cannot recreate fields they never contained.
September research extends the evidence journey beyond a single permission boolean and into current standing, protected boundaries, response delivery, and release.
Document Review experiments can retain separate observations before provider dispatch, tool commit, and release. Unavailable authority remains different from an explicit denial, and historical standing is not silently reused.
Scoped tests distinguish current permission from the binding between an authorised subject and the reviewed content. Changed or unavailable content can therefore hold release even when a permission flag remains allowed.
Direct streaming paths can withhold answer content until completed-response commit evaluation. Progress events may remain visible, but denied candidates do not become released responses or retained response artifacts.
Local scenario results do not establish external identity authority, universal enforcement, atomicity across every transition, or deployed-provider behaviour. Those boundaries remain explicit in reports.