Publish the trace, including the failures
An agent's run history should be readable by the people it affects: the runs that stopped as much as the ones that shipped.
“We hold that any agent presented as capable must publish its run history: what it attempted, what it used, how long it took, and why a run that stopped short stopped, with failed and abandoned runs retained rather than pruned.”
Capability claims without traces are marketing. A number that says an agent succeeds most of the time is unfalsifiable unless the runs behind it can be read, and the runs that matter most are the ones nobody chose to feature.
A run that stopped because the agent would not guess is the single most useful record a system like this produces. It is evidence of a boundary being respected. Treating it as an error to be swept up removes the proof that the boundary exists at all, and trains everyone to read silence as success.
Retention has a cost and traces contain things that should not be public: credentials, private content, somebody's data. The commitment is to publish the shape of every run and to redact rather than delete: a trace withheld for a stated reason is still accountable, a trace that never existed is not.
Signing this means
- Keep failed, cancelled and abandoned runs alongside the successful ones.
- Record why a stopped run stopped, in words, where the run is read.
- Redact what cannot be shown rather than removing the record that it happened.
Add your name
Signatures are counted per position, not per person. You can hold this one and not the next.
