Bounded Runtime Control
Runtime Governance
Elora studies how AI runtime behaviour can be observed, constrained, explained, and governed without treating model output or learned signals as authority.
Bounded Runtime Control
Elora studies how AI runtime behaviour can be observed, constrained, explained, and governed without treating model output or learned signals as authority.
The governor separates four responsibilities so that escalating from information to intervention is deliberate and reviewable.
Collect bounded runtime signals, state changes, and evidence from registered execution surfaces.
Evaluate conditions against the available policy, runtime context, and evidence without silently converting inference into fact.
Present a proposed response and its evidence for review or downstream governance.
Permit a bounded intervention only where explicit authority, policy, and admissibility controls allow it.
CPU and memory admission, leases, lifecycle accounting, and terminal release help keep worker activity inside declared operating bounds.
State evidence makes pressure, continuity, memory use, and interventions inspectable rather than hiding them behind a single health score.
Related pressure signals and focus interventions can be grouped into one decision story, preserving both the trigger and the resulting response.
Runtime intelligence contributes learned evidence and recommendations. It remains advisory and cannot grant itself execution authority.
Current 0.2.8 work extends Runtime Memory Governance across inference, governed interaction, research, reporting, diagnostics, and learning lifecycles.
Evidence-dense Observer, Threat Intelligence, and Governance Provenance reports run through bounded managed work, produce durable authorized artifacts, and release disposable state at completion.
Conversation-shaped research runs expose memory ownership, lifecycle state, model/provider context, and completion evidence without duplicating the full forensic record.
Queued work can remain safely held while managed inference is busy or unavailable. Recovery is bounded, explicit, and audited rather than silently restarting or overlapping execution.
Context selection considers retained-history demand, available model evidence, and conservative resource limits. Unknown capability falls back safely rather than being guessed.
A transport-level success is not treated as a complete answer. Unterminated or incomplete output remains labelled evidence and cannot become an approved commit candidate.
Completed summaries, reports, promoted learning, and reviewed profiles remain durable while rebuildable scans, predictors, prompt assembly, and rendering state are reclaimed.
Runtime controls are not autonomous permission. Elora re-evaluates proposal, identity, policy, context, and authority at the consequence-bearing boundary, with outcomes linked into Decisions and Replay evidence.
Recent 0.2.8 work makes skill viability, runtime measurements and interrupted research easier to inspect.
Document Review admits its budget against the assembled runtime context and bounded model-capacity evidence. Proposed tuning remains subordinate to declared task viability; insufficient allowance holds before inference rather than silently releasing an incomplete review.
Retained operation timing and thread CPU evidence account for overlapping work. Unavailable measurements stay explicit, and elapsed request waits do not become claims of provider execution time.
Stopping a comparison targets its unfinished children and preserves completed siblings and unrelated runs. Incomplete cleanup stays visible and prevents deletion until stopping succeeds.
Persistent no-progress protection limits repeated Observer child restarts. Elora/Schematic reports can render retained runtime-memory comparisons through reusable evidence sections without inventing absent historical measurements.