How Cortex organizes AI work
How Cortex groups captured AI activity into sessions, work units, and workstreams, with evidence provenance and limits.
- Captured evidenceKeep source records in view
- Inferred relationshipsSeparate grouping from observed results
- Scoped totalsCheck time, visibility, and pricing coverage
Related API changes and investigation
Captured sessionReview the API change
Linked recordPull requestRelated test changes and review
Captured sessionCheck the test casesRelated work units, with their source evidence still in view.
Cortex organizes captured AI activity into sessions, work units, and workstreams so you can follow a summary back to its evidence. The work model separates three things: what was captured, what was calculated from those records, and what was inferred about their relationship.
This guide explains the current Cortex Atlas data model. Available fields depend on capture coverage, processing, source access, and your visibility scope.
From activity to related work
| Record | What it represents |
|---|---|
| Captured exchange | A recorded interaction with a supported AI surface, including the permitted request, response, and metadata. |
| Session | Captured activity grouped under a session identity. A session can contain multiple conversations and many exchanges. |
| Work unit | Related captured activity grouped under a work-unit identity, with a title, summary, participants, usage, and evidence where available. |
| Workstream | A grouping of related work units. Its summary and membership describe the relationship between those units. |
These levels describe different kinds of records, not interchangeable counts. Three exchanges are not necessarily three sessions or three work units. A work-unit detail can link several conversations. A workstream preview shows only a bounded selection of its units.
The current work-unit list is built from processed snapshots. A recently captured session can appear before its work-unit grouping is available. Some activity remains unassigned; a missing grouping does not mean nothing was captured.
Captured evidence and inferred meaning
Captured evidence includes permitted source text, conversation references, timestamps, model identifiers, and linked records. A sample can carry its author, conversation, capture time, model, and an evidence pointer. Follow the reference to inspect the surrounding activity when it is available.
Calculated values include counts and usage totals over a defined population. They inherit the population's time window, capture coverage, and visibility limits. Cost also depends on available pricing.
Inferred meaning includes work-unit titles and summaries and the grouping of related activity. A readable title helps you find the work; it does not turn the title into a captured instruction. An early instruction is not automatically the session's overall goal. A summary that describes a successful action is not, by itself, evidence of a completed result.
| What you see | How to interpret it |
|---|---|
| Session title | Check its provenance. A display title can be inferred or drawn from available source text. |
| Work-unit title and summary | A description of grouped activity. Open its evidence before treating a claim as an observed outcome. |
| Conversation or source reference | A route back to supporting records, subject to availability and access. |
| Linked ticket status | The state reported by that source. It does not verify every activity or claim inside the work unit. |
| Estimated human hours | An estimate when supplied, not measured human labor or time saved. |
Session duration describes elapsed time across captured activity. It does not measure active human labor or time saved.
Evidence samples are selective. Current snapshots can provide prompts linked to the work unit's classified exchanges; older snapshots can use selected prompts from its conversations. Neither is a complete transcript. Missing attribution or an unavailable reference should remain visible as a limitation.
Read a work-unit rollup
The example below groups three captured sessions under API reliability. Its title is illustrative and inferred. Two sessions have an available price; the third is unpriced.
The available total is $0.60, from $0.42 plus $0.18. It is a partial total, not evidence that the unpriced session was free. The grouping also does not establish whether the API change was accepted or deployed.
Work-unit cost carries a coverage state:
| Pricing coverage | Meaning |
|---|---|
| Complete | Pricing covers the activity included in this cost calculation. Capture coverage can still be incomplete. |
| Partial | Some included activity lacks pricing; read the amount as the available portion. |
| Unavailable | No usable amount is available. Do not substitute zero. |
| Unknown | An amount may be present, but its pricing coverage is not established. |
Spend is estimated from captured usage and the pricing table, rather than a provider invoice. Token, conversation, and person breakdowns describe the selected records. Do not add totals from overlapping groupings and assume they represent distinct activity.
Scope, time windows, and freshness
Before comparing two views, check their organization, people, visibility, date range, and pricing coverage. The work-unit list's activity filter and a unit's spend window can differ: the current list selects units by their latest activity, while work-unit detail uses a 30-day rollup. Selecting a recent activity window does not necessarily recalculate every displayed cost for that shorter period.
A response generated now can still contain an older snapshot. Where supplied, snapshot refresh time, data-through time, freshness state, and caveats describe how current the underlying records are. A pending or delayed read model is not a zero-activity result.
A short list of evidence samples, model cards, or workstream previews is not the whole population. Use the supplied totals and pagination instead of counting visible cards. Keep missing fields distinct from confirmed empty lists.
People, teams, and ownership
A participant contributed captured activity; that does not automatically make them the owner. A work-unit owner can be marked as suggested. Workstream team attribution is not currently populated. Do not infer a confirmed team assignment from its participants.
Grouping and team membership do not grant access. Cortex applies the viewer's visibility to the available evidence and supported aggregates. External records also depend on their source permissions. See Roles and visibility.
Review the evidence behind a summary
A linked receipt identifies supporting evidence. An available receipt is not a general certification that the work succeeded, and an unavailable receipt is not proof of failure.
Continue in Cortex
- Work units and workstreams: find related work in Atlas.
- User profiles: inspect the captured sessions behind a person's activity.
- AI spend: interpret estimated cost and its scope.
- Device and source coverage: investigate missing or delayed activity.
- Search company knowledge and inspect evidence: retrieve context and inspect the evidence behind an answer.