LiveLeaLifeTraining
Live OpsMonitor live training, queue operator-defined tasks, and adjust runtime settings without restart.
Active routing + candidate lifecycle + live apply counters.
Model Update Progress
state noneActive Controller Tag
update - -> - at —
controller tag -
standard task model -
live apply 0 at — | watermark 0 | delta since watermark 0
window progress 0/30 | next eval due - | pending none | recent success 0/0 (0%)
Candidate Under Evaluation
No candidate queued yet. This is normal between update windows. Next evaluation window: 0/30.
pending state none | canary -
quality candidate/baseline 0.0 / 0.0 (+0.0) | summary —
Configured Routing Models
Installed but Unused
unused 0
retention keep - controller tags (older non-pinned tags may be pruned)
Outcomes / Success / FailureRolling training-outcome counters used to decide promotion eligibility.
0 / 0 / 0
Live Apply CountNumber of times a new model tag has been applied live after passing promotion/canary gates.
0
Last Live ApplyTimestamp when the most recent live model switch was applied.
—
Last SwitchPrevious active model tag, current active tag, and switch timestamp.
- -> - @ —
Proof Snapshot GeneratedBackend generation timestamp for this model proof snapshot.
—
Proof Cache AgeShould stay near zero and reset frequently while polling.
0s / 0s ttl
Read-only lane behavior. Source lane mode is Hybrid.
Automatic Training Loop
UnknownStart Training: UnknownThis lane continuously claims a source, launches training runs, and writes outcomes to learning memory while enabled.
Current Source Type
none | running 0/-
Fallback Policy
Curriculum fallback disabled
Host Lock
none (host-diverse mode)
Host Selection Policy
Host-diverse concurrent mode | same-host concurrency 1
Stop Training pauses new automatic picks. In-flight runs continue to terminal state.
Queued Exact Runs
No queued exact runsDivineStatus pending 0Queueing from this page creates exact prompts (for example DivineStatus cost-estimate flows) and marks them `queued_from=training_ui`.
Pending Queue (All)
0
Pending Queue (training_ui)
0
Pending Queue (other origins)
0
Pending Curriculum
0
Hybrid mode: queue is empty; curriculum fallback is used only if enabled. `training_ui` share 0/0.
Training runs are separate from standard runs, but both share the same learning store.
Queue Intake
0 consumed in 30m | 0 pending
Consumed training sources become execution-ready cycles.
Training Execution
training 0/- running | 0/0 recent successes
Runs emit outcomes that feed the learning relay.
Learning Ingest
0 ingested in 30m | backlog 0
Outcomes are upserted into shared learning memory.
Standard Task Reuse
window 0/30 | live applies 0
Standard tasks use active model -.
Current Window To Next Eval
0/30Window progress increases as successful training outcomes are ingested. When it reaches 30, the controller can advance evaluation/promotion checks and apply model updates to standard tasks.
Training Sources Consumed (30m)
0
Learning Records Ingested (30m)
0
Learning Ingest Backlog
0
Running Split
training 0 | standard 0
Standard Task Model
-
Latest Training Success
—
Training tasks execute on the training lane only. After completion, the run is written totraining_outcomes and then ingested viarun_learning_outcomes/memory_items. Standard tasks read from that shared learning state on subsequent runs.
Fast pulse for queue pressure, runtime utilization, and model update momentum.
Queue Pending
0
Training Running
0/-
Window Progress
0/30
Live Applies
0
Real-time proof of what improved since this page session started.
Successful outcomes gained
+0
Live applies gained
+0
Queue consumed
+0
Last model switch
—
Baseline captured at —. Values above are derived from live `learning_proof` and queue telemetry only.
Queue → Execute → Outcome → Promotion
Queued
0
+0 curriculum
Executing
0
of - cap
Success Rate
0%
0/0 recent
Live Applies
0
model promotions
Confirms what is being trained right now and what will be picked next.
Selection Modequeue_pending means operator queue source. curriculum_fallback means queue is empty and fallback selected curriculum source.
none
Fallback Enabled (Live)Runtime toggle currently applied in the trainer process.
Off
Queue PendingPending operator-queued training sources.
0
Queue Pending (training_ui)Exact prompts enqueued from this Training UI, including DivineStatus tasks.
0
Curriculum PendingPending curriculum sources eligible only when fallback is enabled and queue is empty.
0
Active Host LockWhen set, queue selection is temporarily pinned to this host to avoid mixed-host training drift.
none
Training Outcomes StorageTraining outcomes are appended as new rows and not rewritten by source selection.
Unknown
No active training cycle and no eligible next source preview.
Define once, then queue one run or a full campaign count of that same task.
Training Task Definition
This prompt is what the LLM follows during training execution.
Campaign Batch Queue
Campaign Size creates that many queued training runs from the task definition above.
Campaign Progress
Campaign: -
Progress: 0/0 (0%)
Queued: 0 | Failed: 0
Started: — | Finished: —
Campaign rows are generated from the task definition above, with unique scenario markers so all queued runs execute independently.
showing 0/0 (max 10)
No cycles yet
Start training or queue prompts to create cycles.
queue 0/0 | curriculum 0/0 (max 10 each)
Queue Sources (Pending)
No pending queue sources.
Curriculum Sources (Pending)
No pending curriculum sources.
Control automatic training and edit runtime values without restart.
Training Runtime ControlStart/Stop controls only the automatic training loop. Queued training prompts remain queued until consumed.
Lane mode Hybrid. Automatic lane status Unknown. Selection mode none.
Queue Pending
0 (training_ui 0 | other 0)
Curriculum Pending
0 | fallback disabled
Current Source
none | running 0/-
Host Policy
Host-diverse concurrent mode | lock none (host-diverse mode)
One caveat: if scripts/training_daemon.py is running with TRAINING_DAEMON_START_IF_STOPPED=1, Stop can be auto-reversed by the daemon.
Applied Now (Live)Current runtime values already active on the running backend process.
Selectors match live valuesTraining-Only Runtime SettingsThese controls affect training loop behavior only and apply without restart.
Training Max Active RunsTraining-only cap: limits active training cycles, not total runtime runs.
-
Loop Interval (s)Delay between scheduler ticks when capacity is available.
-
Recent Host WindowHost dedupe window used to diversify training source selection.
-
Retry Budget / SourceMax failed attempts allowed per training source before disablement.
-
Default Pages TargetDefault page-count target only when prompt does not explicitly set one.
-
Curriculum FallbackWhether curriculum sources are allowed when operator queue is empty.
Off
Source Lane ModeHybrid = queue first then curriculum fallback. Queued Only = queue only. Automatic Only = curriculum only.
Hybrid
Campaign Lock ModeWhen enabled, campaign queue selection is locked to keep training intake stable.
Off
Host Selection PolicyCampaign lock On = strict single-host lock. Campaign lock Off = host-diverse concurrent selection across queued hosts.
Host-diverse concurrent mode | active lock none (host-diverse mode)
Same-host concurrency cap 1.
All-Task Runtime Settings (Read-only)Global values shared by training and manually-created runs.
Global Max Concurrent RunsHard upper limit for all simultaneously running runs in the system.
-
Reserved Control SlotsSlots reserved for non-training control-plane work (self-improve/model-promotion/canary).
0
Effective Training CeilingDerived cap for training-only slots: Global Max Concurrent Runs minus Reserved Control Slots.
-
Running Split (Now)Real-time active run breakdown used to explain why total running can exceed the training-only cap.
training 0 | non-training 0
Controller Max Calls / RunHow many controller-model URL-pick calls are allowed per run.
-
Browser Headful (Global)Whether browser sessions render with visible UI windows.
Off
Browser Autostart (Global)Whether browser sessions are auto-launched for run execution.
Off
Change SelectorsDraft values you can edit before applying live runtime updates.
Training-only settings. Edit values below, then click Apply Settings. Global runtime caps remain read-only here.
Hybrid consumes queued sources first, then curriculum when fallback is enabled. Queued Only blocks curriculum. Automatic Only ignores queued prompts and runs curriculum lane only.
Host-diverse mode spreads concurrent runs across hosts/pages. Strict mode keeps one campaign/host flow stable for deterministic replay.
Pending ChangesDifference between currently applied runtime values and your draft selectors.
No pending runtime setting changes.
Auto-refreshing diagnostics state.
Last UpdatedWhen this panel last refreshed from live API diagnostics.
—
Last HeartbeatMost recent training coordinator heartbeat. If stale, the trainer loop may be stuck.
—
Latest CycleMost recent training cycle ID executed by the coordinator.
-
Latest Cycle StatusTerminal status of the latest cycle (running/completed/failed/cancelled).
-
Next Eval Due (success count)When total successful outcomes reach this number, promotion evaluation is due.
0
Last Eval AttemptMost recent promotion evaluation attempt timestamp.
—
Last Eval ResultLatest eval attempt outcome (triggered/blocked/completed/failed).
-
Eval Retry AtIf evaluation is blocked, this is the next scheduled retry.
—
Last ErrorLatest coordinator-level error code. This is not always a run failure; it can be a gate signal.
-
Oldest Pending SourceOldest queued training source waiting to run. Older timestamps indicate queue backlog.
—
Trace Coverage (diag)Share of recent outcomes with usable training trace payloads (used by self-improve/export).
0%
Diag Outcome WindowCanonical diag-snapshot outcome counters. Same values shown on Dashboard's Training Success Rate card.
0/0 (last 0)
no-snapshotcache 0s / 0s ttlCross-page facts derived via canonical viewmodel.