Start with the shift.
Agentic systems matter because actions can change external state. RESONANCE follows that shift from capability to consequence and asks what evidence is required before an action should be trusted.
The Agentic TurnWhy AI is moving from chat toward execution, tools and persistent workflows.
The Missing Trust LayerA verification model spanning state, causality, phase, transition, time, recovery, verification and evidence.
When Agents FailAn eight-axis failure taxonomy paired with a reproducible benchmark seed.
Verified research.
The verified-report series turns abstract trust claims into bounded, reproducible tests. The sequence now runs from framework checks into real PostgreSQL concurrency, transactional outbox, HTTP acknowledgement ambiguity, execution-memory durability across restart, time-bounded retention, clock authority and monotonic temporal history.
Structural verification against a pinned OpenAI Agents SDK revision.
Tests of filesystem, workspace and Docker control-plane boundaries.
Reconcile-before-retry under timeout-after-commit ambiguity.
Preserving UNKNOWN instead of treating uncertainty as retry permission.
Authority and freshness when sources disagree.
Why claimed authority is not verified authority.
Why formerly valid authority can become invalid before action.
Why fresh evidence evaluated against stale trust state is not trusted evidence.
Execution-bound authorization and trust-epoch TOCTOU protection.
Two readers, one version-bound writer through an atomic transition.
Eight protocol stages across a compounded adversarial trajectory.
The state-version invariant moved from model to a real PostgreSQL service.
0-row conflicts and SQLSTATE 40001 mapped into one business recovery law.
Durable delivery intent, stable idempotency and reconciliation across an external-effect boundary.
A separate service commits remotely, drops the response, then proves same-key redelivery and remote-status reconciliation.
A durable effect survives service restart while volatile dedupe memory disappears; persistent safety memory or reconciliation prevents replay.
A durable dedupe record can expire before delayed recovery. Short retention duplicated a same-key replay; sufficient retention or authoritative reconciliation preserved one effect.
One record can be ACTIVE for one clock and EXPIRED for another; authority time or bounded uncertainty prevents local-clock replay from becoming permission.
A reversible wall-clock value can make an expired record look active again. Temporal epoch fencing or a monotonic time watermark prevents retired execution state from being resurrected.
The research question.
The central question is not simply whether an agent produced the desired final output. It is whether the trajectory can be independently inspected: what caused the action, whether the transition was allowed, what survived failure or restart, whether safety memory remains valid at recovery time, whose time decides that validity, whether execution history has advanced beyond a reversible clock value, and whether the surviving evidence deserves trust.