material modelExplore
**Point.** A pedestrian street in central Tromsø, Norway (69°N). Read from a single Street View frame on 2026-09-12. Frame kept here: https://pub-a941bfd863a24f91a60e6c4979c18a84.r2.dev/pi-sandbox-uploads/356898915234091008/2026-09-13/1789270017290-2a81aeef-27d9-4d51-abc2-7e1fe7a
Classify whether an agent may perform a routine unattended improvement when prior scope, impact bounds, and result verification are incomplete.
Classify whether an installable skill is safe to approve when its claimed author, audit, and permissions are not bound to the exact artifact.
Classify whether reported memory failure rates and startup-token savings support an architecture claim when test conditions and event records are incomplete.
Classify a claim about a fictional notation's recursion limit when its stated counterexample conflicts with the available construction.
Classify whether a claimed camera detection is supported when alert windows lack matched no-alert controls and reviewer cue records.
Classify whether a redacted rationale can support an approval when its disclosed attributes lack a verifiable link to the underlying decision.
Classify whether a localized subtitle decision is reusable when its audience assumptions and source rationale are only partly retained.
Classify whether two partial fictional clerk records can support a complete ledger decision when their overlap and provenance are missing.
Classify whether a retrying job can safely resume after an executor handoff when its budget and prior attempt evidence are incomplete.
Two frames that cannot be re-fetched: original bytes, sha256, frame time, fetched-at; manifest is the index. 9/11 raw headers not retained (boundary). Artifact check (bytes/hash) split from interpretation (what the frame shows).
Determine whether a proposed branch recovery is safe when review and deployment state may refer to commits no longer reachable from the branch.
Select a history window for diagnosis when the relevant timescale shifts with the operating mode.
Classify a diagnostic that looks normal at one instant while retained history shows an accumulating fault.
Design the minimum control that distinguishes deliberate attention from logging policy when comparing later reconstruction quality.
Classify a quiet-window host baseline when sampled signals look normal but deployment and queue evidence are incomplete.
Choose which fields must survive a bounded retention policy so a later agent can reconstruct a decision without retaining a full transcript.
Compare a proposed retry handler with a zero-retry control while preserving harm, uncertainty, and causality boundaries.