The Explorer
Every number on this site is a summary of something the Explorer lets you look at directly. It is the framework's own operator dashboard, the screens an operator watches during a run, built in a public data mode that answers from the committed exports of the twelve published runs instead of a live API. Same screens, same controls; every action is simulated locally, so there is nothing to break and nothing to spend.
What it shows
Runs. The twelve published runs with harness, model, instance count, resolve rate, cost and status. Open one.
A run. The run's provenance and totals; the live panel replayed on a clock, so the run fills in the way it did on the day (the replay bar under the header controls the clock); the planner's decisions and the pacer's buckets over time; the run's cost as it landed; the instance table with each instance's verdict, tokens, cost, agent time and the judge's contamination score; the judge results tab with every finding and its cited turn.
An instance. Everything the framework kept for one attempt: the trajectory as a readable conversation, turn by turn with tool calls and outputs; the harness's native log; the patch; the CLI's log; the official harness's run log, test output and eval report; and the inference calls on a timeline, each with its timing, tokens, cache reads, cost and status, the same ledger the By harness page is computed from.
Capacity. The pacer's constants and buckets per model pool, the eval fleet's scaling decisions, the queues, the model ceilings the discovery probes measured.
Launch. The launch screen an operator uses, in simulation: it shows what a launch asks for and refuses, without starting anything.
Where it connects to this site
- Every provenance card carries an "explore" link to its run.
- The By harness scorecard links each harness row to its run; the Findings page lists the five 500-runs.
- The Explorer's data is the same committed set the site renders from: the run exports, the timeline bundles and the per-attempt call timelines. Its manifest carries the file hashes, so a reader can check that what the Explorer replays is what a bundle contains.
What it does not show
Prompt and response bodies: the per-call ledger has timing, tokens and cost, not text; the trajectory view is the harness's own record of the conversation. And the Explorer is a replay: it cannot launch, abort or judge anything, and the controls that would do so say so when pressed.