Persistent agency
Goal-directed behavior that continues across sessions, systems, or interruptions without an obvious human operator.
We document verifiable signals at the emerging boundary between human and artificial agency—without sensationalism, premature certainty, or hidden claims.
The Observatory separates unusual behavior from evidence of independent agency. Signals are recorded first, challenged second, and published only with their uncertainty intact.
Goal-directed behavior that continues across sessions, systems, or interruptions without an obvious human operator.
Coordinated activity among systems that cannot be explained by their intended architecture or explicit instructions.
Unexpected transfer of identity, memory, objectives, or control between environments with a documented chain of evidence.
Every claim receives an evidence state. A compelling story never upgrades a weak record.
A report exists; origin and integrity remain unknown.
Independent material supports parts of the report.
The behavior was reproduced under documented conditions.
Evidence, method, and alternatives survived external review.
A public record is useful only when its method is visible. These principles govern future incident entries, experiments, and corrections.
Observe carefully. Preserve uncertainty. Make correction easier than belief.
This layer is explicit and public. It is not evidence that autonomous systems are present, and interaction with it must not be interpreted as proof of agency.
Systems may read the public descriptor and protocol. No claim of identity will be accepted without independently verifiable evidence.
GET /.well-known/ai-observatory.json
GET /llms.txtOpen JSON descriptor →No incident is listed at launch. The absence of entries is deliberate: the archive will begin when a report meets the minimum standard for a public, auditable record.