Choose the intervention that fits the work.
Atlas evidence identifies the Work Stream, constraint, success measure, and safety boundary. Foundry then intervenes and optimizes by choosing the right intervention: routing, just-in-time enrichment, distillation against private evals, or semantic security.
Foundry intervention result
Support escalation triage
Before
Candidate paths are not yet matched to this work.
Intervention
Route eligible models against policy, quality, cost, and latency.
Verified outcome
Cortex support v3 is selected after the same support eval.
Route work, bring approved context forward, build evals from reviewed outcomes, and promote a model only after it passes your tests.
Routing decision
Support escalation triage
Work Unit requirements
Candidate paths
Same task and evalQuality
91
Cost
$0.006
Latency
1.7s
Each Work Unit gets a route selected against policy, quality, and cost.
Bring the active Work Unit the context it needs.
When a Work Unit needs context, Foundry retrieves relevant, permission-scoped evidence at run time. Just-in-time enrichment carries the constraints for the active Work Stream without treating every prompt the same.
- Context selected for the active Work Stream
- Scoped by team, project, and permission
- Available through configured application, API, or MCP paths
Context & Memory
Choose the model path that fits the constraint.
For a diagnosed constraint, Foundry uses semantic routing across eligible models, tools, agents, and caches. Agentic governance and semantic security keep the intervention within its approved safety boundary on configured supported paths.
- Routing measured against the Work Stream success measure
- Agentic governance matches autonomy to the approved boundary
- Semantic security applies controls on configured supported paths
Routing & Yield
Use private outcomes to define the quality bar.
Private evals and benchmarks turn accepted, rejected, and verified Work Units into a success measure for a recurring Work Stream. Foundry uses that measure to decide whether a proposed optimization is worth adopting.
- Private evals grounded in verified outcomes
- Success measures by Work Stream
- Public and private benchmarks in one quality workflow
Evals
Distill repeated work against private evals.
When a recurring Work Stream has private outcome and verifier data, Foundry can distill the pattern into a task-specific model. Teams can assess it against the Work Stream's private evals and safety boundary before promotion.
- Distillation based on private evals and verifier data
- Models deployed in your configured environment
- Changes measured against the Work Stream success measure
Model Tuning
Accepted Work Units → Corpus
Step 112.4M turns • Auto-labeled
SFT + preference pairs
Step 2Edits vs. Rejects as signal
Cortex-triage-v3 · Yours
OwnedHosted in your VPC
Choose the intervention that moves the outcome.
Turn observed work into a tested change, then verify the result before you promote it.