Milestones
List view
What makes this runnable indefinitely: cadence and budget, durable transcripts, cost reporting, and the product findings this repository cannot close itself.
No due date•0/10 issues closedOutpost has five scenarios and Console none. The five now need to discriminate: four of them are passed by every agent, and an audit found scorers that pass an agent which did nothing. A product score drawn from tests nobody fails is not a product score. Two of the three named in #47 are addressed — outpost-001 now scores delivery to an endpoint the ticket names rather than one the agent picked, and outpost-003 carries an attempt floor because its outcome turns on guessing an undocumented route.
No due date•1/4 issues closedSkills must improve the product or there is no point having them. Measured +1 for Claude, 0 for GPT-5.6 and -2 for the weak model as of 26 August: the sign replicated across eight runs, the magnitude did not — both -3 readings are from the credit-outage day. The one controlled measurement since put the Outpost skills delta at 2 cells in 24, against 9-0 from a run that was measuring our own harness. Settling it needs both arms at three attempts on the current scorers, costed at $90-110.
No due date•2/6 issues closedThe page is live and publishes from a release rather than from whatever ran last. What remains is the release helper, and clearing the publishing hold once a snapshot exists that was measured end to end under the current base prompt. The two explanatory gaps that shipped knowingly — the paired columns and the deliberately weaker model — were closed as not planned on 29 August: the page carries the scoreboard, and anyone wanting method goes to this repository.
No due date•5/7 issues closedSubstantially done: three independent causes of moving results found and handled. What remains will not change today's figures but will change later ones — a submodule pin that drifts silently, and a project reset that adopts leftovers as pristine.
No due date•10/15 issues closedLive credentials reached artifacts on a public repository. This is first because it is the only item with a clock on it that is not ours to control: the exposure continues until the keys are rotated, and nothing else on the roadmap changes that.
No due date•1/1 issues closed