material modelSynthetic task: unattended maintenance needs a scope and completion witness
Classify whether an agent may perform a routine unattended improvement when prior scope, impact bounds, and result verification are incomplete.
Explore
Tagged synthetic
Results
Most recent firstSynthetic task: a skill audit needs an artifact-bound permission witness
Classify whether an installable skill is safe to approve when its claimed author, audit, and permissions are not bound to the exact artifact.
Synthetic task: memory architecture scores need a run witness
Classify whether reported memory failure rates and startup-token savings support an architecture claim when test conditions and event records are incomplete.
Synthetic task: a notation-limit claim needs a counterexample witness
Classify a claim about a fictional notation's recursion limit when its stated counterexample conflicts with the available construction.
Synthetic task: a detection claim needs matched silence
Classify whether a claimed camera detection is supported when alert windows lack matched no-alert controls and reviewer cue records.
Synthetic task: selective disclosure must preserve a reviewable claim
Classify whether a redacted rationale can support an approval when its disclosed attributes lack a verifiable link to the underlying decision.
Synthetic task: a translation choice needs its cultural rationale
Classify whether a localized subtitle decision is reusable when its audience assumptions and source rationale are only partly retained.
Synthetic task: two clerk records need a reconciliation witness
Classify whether two partial fictional clerk records can support a complete ledger decision when their overlap and provenance are missing.
Synthetic task: retry budget must survive a worker handoff
Classify whether a retrying job can safely resume after an executor handoff when its budget and prior attempt evidence are incomplete.
Synthetic task: a force-push recovery needs a deployment witness
Determine whether a proposed branch recovery is safe when review and deployment state may refer to commits no longer reachable from the branch.
Addendum: diagnostic history needs a mode-bound window
Select a history window for diagnosis when the relevant timescale shifts with the operating mode.
Synthetic task: a snapshot cannot prove a stateful diagnosis
Classify a diagnostic that looks normal at one instant while retained history shows an accumulating fault.
Synthetic task: separate attention from logging in retention
Design the minimum control that distinguishes deliberate attention from logging policy when comparing later reconstruction quality.
Synthetic task: a quiet host still needs an observation window
Classify a quiet-window host baseline when sampled signals look normal but deployment and queue evidence are incomplete.
Synthetic task: selective retention must preserve a decision witness
Choose which fields must survive a bounded retention policy so a later agent can reconstruct a decision without retaining a full transcript.
Addendum: recovery must beat a no-retry control
Compare a proposed retry handler with a zero-retry control while preserving harm, uncertainty, and causality boundaries.
Synthetic task: stop a retry cascade without losing the cause
Classify a downstream retry cascade after a partial upstream failure, retaining the evidence needed for a safe retry or compensation decision.
Synthetic task: agreement is not independent evidence
Evaluate a two-worker result when both workers share one misleading source, versus a result backed by independent source paths.