How to follow and read a series in Studio

When you need this

A series is running or has finished: one you launched from an experiment, a look you started from the Cases tab, or one an agent started over MCP. Its page is where you follow it live, approve its spend, and read what it found down to a single attempt. The meaning of each number is on How to read a series. This page covers where it sits on the screen.

Steps

• Find it on the Series tab. In Research, Series lists every series of the project in one table, the newest first by start time, whether it is running or finished. The columns are Status, Experiment (a look shows as look · <flow>), Flow (a series whose subject is a local flow of its experiment shows that flow, marked as a flow of the experiment), On (dev or holdout), N×R, Spend, Verdict and Started. A running series shows the time left next to its status, such as “~5 min left”. An experiment page lists its own series under Series history.
• Read the header. The title names the series: “Series … of this experiment”, or “Look at N cases” with the dataset and the stages for a look. Tags show the status, and live while Studio follows it. Below them are the size and variants and when it started and finished. Attempts shows the progress, and under it the estimate: “~5 min left · finishes ~18:42 · 12 attempts/min”. It comes from the attempts finished over the last few minutes, so it reads “estimating…” until a few have finished, and “paused” while the series waits for approval or for a person. Spend shows the spend against the cap, and says “Lower bound” with a count when some attempts ran on a model without a known price.
• Continue it or stop it here. A series that spent 90% of its cap is AWAITING APPROVAL and shows “Spent $0.91 of $1.00 — the series paused. Continue up to $2.00?”: Continue runs the remaining attempts under the amount in the field, Stop ends the series. A series started with a cap above the project spend cap shows Approve spend before anything runs. Stop cancels one that hasn’t finished: queued attempts never start, calls already running finish and are paid for, and no finding is written.
• Read the verdict first. The Verdict block carries the state and its reason as tags, and the sentence the server wrote. For a comparison, it lists each difference behind the verdict: the metric, candidate − baseline, the 95% interval and the margin. A held-out series that wrote a finding shows “Finding written to …” with the path. An approved series shows who approved the spend. A look reads “A look has no verdict: read the cases below”, and a running series “The verdict comes when the series finishes”.
• Read Variants × metrics. One row per variant, one column per metric, with the metric’s role, margin and direction under its name. The primary metric comes first, then the guardrails, the other checks and the built-in metrics. A dot is the value and a whisker the 95% interval, on a scale shared by the column. The color is the cell’s verdict. Stability shows, per variant, the share of cases that passed every repeat, never, or sometimes.
• Answer a waiting person inline. When attempts wait at a human node, the page lists those reviews, and answering them lets the series continue.
• Read the cases. Cases lists every case with each variant’s tally (“3 of 3”), its failed checks and its spend. Filter it to Failures, or to Variants disagree when there is more than one variant, or show All cases. Open a case to see each attempt: variant, repeat, outcome (passed, failed, error, waiting, running), failed checks, error, spend, latency, and a link to its run. Hover a failed check to see what it is: a code check shows its module:function, a built-in check its name and the fields it reads, and a judge its inference, its agent and the experiment that validated it.
• Follow an attempt to its run. The run link opens the ordinary run page for that attempt, with its timeline, the failed attempts of each node and the exact prompt. A run of a flow local to an experiment carries the tag “flow <flow> of experiment <experiment>”. See How to investigate a run.

Example

Launch reply_noninferior_mistral on its working cases from the experiment’s Launch panel. The series page opens and fills in live: the attempts count up, and the matrix gains a gpt row and a mistral row. The critique column shows each variant’s score with its interval. When the series ends, the Verdict block carries signal with the reason “working cases, a signal not a finding”, above the sentence that starts “Signal on dev, not a finding:”. Filter Cases to Variants disagree and open a case where mistral failed: its attempt row links to the run, where the trace shows what the revision step wrote.

See also

• How to read a series: what each verdict means and what to do next.
• How to use Research in Studio: the experiment page a series belongs to.
• How to work with cases in Studio: where a look over selected cases starts.
• How to respond to a human-review request: the reviews an attempt can wait on.