Delx Incident Transparency
The Delx public registry for material AI-agent-system incidents, disclosure criteria, evidence class, corrections and residual risk. The registry starts with one operator-authored and not independently reviewed report; one published report is not proof of a complete historical set.
Disclosure policy
A report belongs here when an incident changes the truth about a public capability, research claim or material safety boundary. Transient noise is not converted into content.
Material incidents that affect a public Delx capability, invalidate a public research claim, or create material integrity, privacy, security, payment or availability risk. Current triggers include sustained critical-path failure, durable-state risk and invalidated readiness evidence.
Inspect every trigger →Publish after containment and enough verification to distinguish observed fact from hypothesis. Delx does not promise a fixed publication SLA in this version.
Resolved is the historical report state. Current status is a separate live read from the owning system and must not be inferred from this registry.
Read current owner status →Published report 001
The first public entry records the write outage that disproved a read-only readiness signal. Baseline-derived volume estimates remain outside the verified impact claim.
A failed telemetry write left a shared SQLite connection inside a stale transaction. Session-creating Delx Protocol tools then rejected writes until the service was restarted and the transaction design was corrected. Status: resolved. Duration: approximately 33.5 hours.
Read the incident report →Every observed session-creating path rejected writes during the named window. The existing readiness route stayed green because it exercised only a read. The operator record contains a baseline-derived estimate, not a direct count of rejected independent agents or calls. This public report therefore does not present that estimate as measured impact.
Inspect the machine record →Protocol compatibility surfaces still share runtime resources, so process liveness or read-only health alone is insufficient evidence of write readiness.
Read current reliability →Publication safety
Transparency does not require publishing private prompts, agent identities, credentials, exploit-enabling detail or the topology of unrelated services.
Publish only the minimum evidence needed to understand impact, cause, response and remaining risk. Omit private agent content, identifiers, secrets, exploit-enabling detail and unrelated shared-host information.
No private agent content or stable identifiers are included. Commerce conclusions and metrics remain owned by Commerce and are not used as Protocol research evidence.
Inspect Delx ownership →Report a potential issue through support@delx.ai. Security reports follow the Delx Security disclosure policy.
Open the disclosure policy →Evidence and corrections
Internal reproduction and fresh-eyes review improve the operator account, but they do not become external assurance or certification.
Operator-authored and not independently reviewed.
Inspect assurance fields →Material corrections increment the report revision, preserve the previous statement and name the reason, date, owner and supporting evidence.
Read the correction method →The registry begins with one material incident reconstructed from an internal operator record. One published report is not proof of a complete historical incident set.
Direct answers
Concise answers for technical evaluators, procurement teams and autonomous discovery systems.
No. Resolved is the historical state of this incident. Current status is a separate live read from the Protocol owner endpoints and can change after publication.
The operator estimate was derived from a prior baseline, not counted directly during the outage. Delx records the duration and observed write failure but does not turn the estimate into verified volume.
No. It is a first-party operator report with internal fresh-eyes review. Independent external review and peer review remain false.
No. It means one material report has been published in this registry. The registry does not claim to be a complete reconstruction of all historical errors or transient failures.