Galaris · Product guide

Connect results, evidence and costs

Trace a result back to its mission, model calls, tools and delivery. Read costs with their source, completeness and billing status.

Follow a causal timeline

Related identifiers connect the request, Task, attempt, model call, tool and Process. Use those links to understand what caused an outcome. A collection of unrelated log lines cannot show whether a specific deliverable came from the intended mission.

Require evidence of delivery

A model saying that a file was saved is not a delivery receipt. Check the identified artifact, destination and transfer result in the Working Set. If transport fails after production succeeds, inspect delivery recovery before rerunning the whole task.

Put cost next to the result

Inspect model usage, inference estimates and billed cost alongside the agent, surface and Task harness. Missing billing data must remain unknown. Monthly views reveal trends; detailed records explain the execution behind them. Subscription access and metered usage are different cost signals.

Execution details — Information and traces from a task actually completed by Albert.
Execution details — Information and traces from a task actually completed by Albert.C07
Model activity — Actual scenario calls, models, durations, usage and costs.
Model activity — Actual scenario calls, models, durations, usage and costs.C28
Two reproducible Albert cases — A synthetic test Dataset distinguishes a brief explanation from building an interactive lesson.
Two reproducible Albert cases — A synthetic test Dataset distinguishes a brief explanation from building an interactive lesson.C29
Dream history — Recorded background operations, results and durations.
Dream history — Recorded background operations, results and durations.C21

Distinguish activity, first output and completion

Conversation and task details separate waiting, preparation and the first observed model output. Generating tool arguments counts as activity; it does not prove the tool has run. First output does not mean the requested result has been delivered.

With the required management permissions, you can request that the current LLM call stop from its detail view. Available traces and costs remain inspectable. Acknowledging the request does not prove the provider has stopped; pausing the task is the appropriate control for suspending work you intend to resume. Control calls.

Measure accepted work

Track successful and failed missions, duration, human corrections, accepted deliverables and known cost. Record configuration changes when comparing runs. Token counts alone do not establish quality or business value.

Read cost together with its qualification

Monitoring distinguishes requested and effective models, reasoning effort, available input/output/cache tokens, duration and first-token time. Amounts can be known, estimated, partial or subscription-based. Missing data do not establish zero cost.

Traces show progress messages and blocks actually exposed by the runtime, redacted according to access. They do not reveal unpublished private reasoning.

Keep a lasting problem record

Incidents group similar occurrences into families with first and latest occurrence, count and review state. Diagnosis can record cause, remediation and regression test. A Task or round can then become a Lab case.

Read retained timings and costs

Views distinguish processing, waiting, pauses and retry delays. Costs use persisted calls, including preparation and routing, without recounting a call on repeated finalisation. An inference links its request, attempts and physical provider calls; reading its result does not run the model again.