Work through Actions
The Actions tab lists what a filter needs from a person, and who should act. Each item carries the control that clears it, or a link to the screen where that happens. This guide explains the tab’s layout and how to act on the items you will meet most often.
When you’d do this
Section titled “When you’d do this”- A filter’s page opens on Actions. It does whenever the filter has anything on it.
- The dashboard’s Needs attention list, or the Alerts page, sends you here. Each row links to its filter’s Actions tab.
- A filter’s tab strip shows a count badge on Actions.
Before you start
Section titled “Before you start”Anyone can read Actions. Every control that changes a filter needs the operator or admin role; a reviewer sees An operator or administrator does this. in its place. Reviewers can still follow links to the gold set and the verdict check. A retired filter has no Actions tab.
How the tab is laid out
Section titled “How the tab is laid out”| Section | What is in it |
|---|---|
| Needs you · n | One card per condition: something true now that a person must fix. A condition clears when it is fixed; it cannot be dismissed. The first card is the filter’s health. |
| Sift recommends · n | Advice you may take or decline. Today the only one is a bake-off recommendation to switch the deciding recipe. |
| Sift is doing | Work Sift has in hand, such as training, shown when the filter has no setup steps. |
| Setup · step n of N | While the filter is being set up, every step from a new filter to one that filters, with who does each. |
When nothing is owed, the tab says Nothing Sift needs from you. The foot links to Settings, for changes you start yourself.
Each card shows a context line (the kind of condition, since when, and whether the filter is filtering), a chip saying who acts, a title, Sift’s reason, any evidence (a progress bar, alarm readings), and the control.
| Who chip | Meaning |
|---|---|
| Your move | You can act on it now. |
| An operator | Shown to a reviewer: an operator or admin acts. |
| Sift | Sift is doing it; there is only waiting. |
| In Kindo | It is done in Kindo, not in Sift. |
In the setup list each step reads Done, Your move, In progress, Stopped or Not yet. A step already carded above says Above, in Needs you. A collecting step shows its count, an estimate such as Roughly 4–9 days more once the filter has been in shadow three days, and quiet since … when the source has returned nothing new for a day or more. A quiet source is not a fault.
Act on an item
Section titled “Act on an item”-
Open the filter’s Actions tab.
-
Start with the first card under Needs you. Faults come first.
-
Read the reason, then use the card’s control, as below.
After you act, the card leaves the list once the condition is no longer true. A wait that only Sift can end stays as a step marked Sift.
Common items
Section titled “Common items”| Card or step | What you do |
|---|---|
| Connect its agent | Opens Settings › Delivery to pair the agent. See Connect a filter to its agent. |
| Review the gold set, Build the gold set, Approve the gold set | Opens the Gold set tab. See Review the gold set. |
| Start in shadow | Add a Note (optional), then confirm. The filter starts polling and scoring, escalates every event, and filters nothing. A Warm start is offered beside it; see below. |
| Train a model, Train now, Train again | Queues training. The model arrives in shadow and decides nothing until an operating point is approved. |
| Approve the threshold, Approve the budget, Approve the running model | Opens the threshold screen. See Approve a threshold. |
| Check the source (source failing) | Opens Activity filtered to polls, where each failed poll shows its error. |
| Check delivery (escalations not reaching the agent) | Opens Settings › Delivery. Admins also get Check the Kindo connection; an overdue grouped delivery adds See held groups. |
| Review where its labels come from | The label count has not moved in 7 days. Opens Settings › Labels; Check the source opens the polls. |
| Regenerate the agent | The paired agent is out of date or cannot take what the filter now sends. Opens the agent section of Settings › Delivery. Steps you added to the agent in Kindo are not carried over. |
| Resume or retrain, Resume on the current model | The filter was stopped. See Hold, stop, resume or retire a filter. |
| Move it back to shadow | The filter is held and has no approval to resume on. Confirm with Move it to shadow now. |
| Review the change | A version waits for approval behind the running one. Opens the threshold screen. See Change a filter’s settings. |
| Check the verdicts | Opens the verdict check. |
| Retire it in Kindo, Check it in Kindo | The agent this filter replaced may still be running. See below. |
Warm start
Section titled “Warm start”Beside Start in shadow, Warm start has the paired agent judge a sample of the filter’s history before shadow, so a first model can be trained before shadow sends the agent every event. It is offered once the gold set is approved and the agent paired, while the filter is a draft or being authored; a draft is moved to authoring first.
| Choice | What it does |
|---|---|
| None (default) | Nothing. The first model trains on shadow’s verdicts. |
| 200 events | A first look. With the gold set, usually short of the default floor of 400 labels. |
| 500 events | With the gold set, enough for the default floor if most verdicts come back. |
| Custom | Any count from 50 to 5,000. |
Choose a count, then Judge N past events. The confirmation states the count, its cost when your deployment prices an agent run, and whether it can reach the label floor. Then Sift:
- Reads a stretch of the source’s history and picks that many events from it, favouring the ones the decision model finds likely to matter, where one is set up. The filter’s first poll in shadow starts after that history, so it does not read it again.
- Sends the events to the agent a few at a time, within the agent budget if one is set.
- Turns each verdict into a Backtest label. When the labels reach the floor, Sift starts training at once, still before shadow, and the model scores from the first poll.
The card shows its progress: how many were sent and judged, whether it waits for room under the budget, and when training was queued. Once the filter is in shadow, the Collect labels step shows it.
- It runs once for a gold set. A second is refused with the reason; a new approved gold set allows another. One that failed before sending anything can be asked for again.
- The agent is told these events are history, if Sift generated it at contract 5 or later (see History). It judges each as if it were live and takes no action outside the run. An agent generated earlier, or one Sift did not generate, is sent each event as a live one, and acts on it as on any other: an agent that opens a ticket for every event it judges actionable opens them for these too. For such an agent the confirmation shows a box, The agent may act on these as on live events, with the reason, and the warm start goes ahead only once it is ticked. The tick is recorded in the audit log. To avoid it, regenerate the agent first.
- Backtest runs are agent runs, counted against the budget and in every run count. They are not escalations: they never count toward the trigger ceiling or delivery health.
- Shadow does not send again what the warm start judged. An event shadow polls with the same source id and the same content as one the agent judged here is recorded as Judged in a warm start: not sent, not filtered, and labelled by the warm start’s verdict. One whose content changed since is sent as usual. An event the warm start chose but had not sent when shadow polled it goes as a live event, and the warm start skips it; the card counts these.
An alarm the filter keeps filtering through
Section titled “An alarm the filter keeps filtering through”If the filter’s On alarm setting is Keep filtering, an open drift alarm or a breached trigger ceiling is a fault, titled A drift alarm is open, and it is still filtering or The trigger ceiling is breached, and it is still filtering. The card shows the readings against their alarm levels and offers four moves, none chosen for you:
| Move | What it does |
|---|---|
| Investigate the drift or Investigate the ceiling | Opens the drift readings on Overview, or the state changes in Activity. Changes nothing. |
| Retrain | A new model enters shadow; this one keeps filtering meanwhile. |
| Move to shadow… | Stops filtering now and keeps scoring. Asks Why (required). |
| Hold events… | Stops reading the source; the agent gets nothing from Sift until you release it. |
A displaced agent
Section titled “A displaced agent”When a filter’s paired agent replaces an agent that did the same work, Sift checks that the old agent has stopped running, and does not count spend as saved until it has. Kindo has no API to retire an agent, so you do it in Kindo. If the old agent is meant to keep running, or is already gone, select Dismiss…, add a Note (optional), and confirm with Dismiss. Sift then stops checking it.
Decline a recommendation
Section titled “Decline a recommendation”- On the Sift recommends card, read the Recall and Filters bars for the recipe that decides now and the recommended one.
- Select Decline….
- Write your reason in Why not, on this evidence (required).
- Select Decline the recommendation.
The card leaves the list. It returns if a different recipe is recommended, or once the holdout has grown by a tenth. To take the recommendation instead, see Compare recipes in a bake-off.
What can go wrong
Section titled “What can go wrong”- The control is missing. You are a reviewer, or the act is done elsewhere (in Kindo, or by Sift). The card says which.
- A button is refused. The server checks again when you act; its message says why, for example that the gold set is not yet approved.
- A card does not clear. Conditions clear when they are no longer true, which Sift rechecks; some, such as a failing source, clear only after the next successful poll.