Decisions
Every decision, who chose it, and its forecast measured against the actual at the scheduled review.
Calibration scoreboard
By decision class — how close the forecasts land, from graded outcomes only. Nothing self-reported.
Judgment lift (retro-eval)
The last replay of the graded ledger through the current model generation, per decision class: what the replay judged beside what was actually recorded. A reading only — it grades nothing and re-grades nothing.
Autonomy
Level per decision class (autonomy levels L0–L4), earned on graded record. The human-versus-model shadow tally that gates promotion is shown per class (agreement rate + graded hit/miss).
Playbooks
The operator's approve/retire queue for codified decision patterns. Each row shows the evidence the rule was approved on beside the live re-audit of that same pattern today — the two are read from different sources and are meant to be compared, not reconciled.
Portfolio health
Big rocks routed to the Log versus sub-floor tail the materiality floor absorbed. A rising automated-tail share with a quiet Log is the system working.
Experiments
Registered pre/post readouts — not two-arm randomized tests. Each row records whether the metric moved between its frozen baseline and the review date; movement here is not evidence of what produced it.
LLM spend
The runaway-spend governor, as it stands today (UTC). The alert state is the server's own deduped signal — one per day — not a threshold this page recomputes.
Monday digest
The digest getDigest writes for the Monday operator mail, shown exactly as it is served. This is the only place in the product where recorded peer-benchmark gaps, the strategist's read and the outcomes falling due this week reach a screen. The mail itself is sent from the orchestrator and is not wired yet — this is the preview, not a notification centre.