Build and own a Clone Generator Specialist Model.
What complete, ready-to-train clone package should be generated for a given Decision Engine?
Overview
Once the research pipeline exists, adding the next 500 decision engines is almost free.
Who owns this decision
Head of Product (Atlas/Decision Engine line), Director of Applied Research
Why owned AI may fit
This is an internal meta-pipeline, not a customer-facing decision task, so the specialist vs. frontier framing is less about domain expertise and more about orchestration: chaining research extraction, schema generation, prompt design, and evaluation design reliably and consistently across hundreds of engines.
What this Specialist Model actually does
Decision: What complete, ready-to-train clone package should be generated for a given Decision Engine?
This is a larger system with 3 specialist model candidates and 3 supporting or external components.
Who uses this Specialist Model
Accountable owners: Atlas product owner, Canonical catalog owner, Source rights owner, Model registry owner, Benchmark owner, Security owner.
- Atlas product
- Applied research
- Solutions engineering
- ML engineering
- Evaluation
- Security
- Compliance
- Release engineering
Where the specialist model sits
This is a larger system with 3 specialist model candidates and 3 supporting or external components.
Pipeline: web research (structured extraction) -> Research JSON -> Decision Engine JSON -> LLM expands -> LLM builds prompt -> LLM builds sample data -> LLM builds benchmark -> human review -> publish
Specialist Model
Grounding/verification layer
Specialist Model
Grounding/verification layer that checks generated companies, benchmarks, and model picks against the verified reference files before a package is marked reviewable
Versioning and re-run capability so a published package can be regenerated when the model catalog, benchmark list, or company list is updated
Specialist Model
Structured research extraction layer (e.g. Exa-style structured output)
Supporting component
Structured research extraction layer (e.g. Exa-style structured output) that normalizes web sources into company/product/workflow/evidence JSON
Template and style-consistency layer
Supporting component
Template and style-consistency layer that aligns new packages with previously published engine records
How Oumi connects to your stack →
Human review queue
Supporting component
Human review queue with structured diff view (generated vs. edited) to capture reviewer feedback as training signal
Decision and authority boundary
Oumi owns cited requirement extraction package drafting supplied-registry grounding critique regeneration analysis confidence abstention and routing. Source rights canonical truth registries deterministic validation approval publication deployment rollback and human authority remain external.
Abstain or escalate when
- Authority missing
- Engine identity missing
- Source right missing
- Cutoff missing
- Canonical snapshot missing
- Registry snapshot missing
- Template or schema missing
- Reviewer missing
How to evaluate and improve it
Evaluation
- Log assignment exposure review release deployment and outcome because generated drafts change reviewer corrections and later package performance. Keep descriptive outcomes separate from controlled incremental effects.
Feedback signals
- Source evidence
- Extracted requirement
- Generated artifact
- Deterministic result
- Reviewer correction
- Release action
- Deployment
Model candidates to evaluate
- glm-5.2
- deepseek-v4
- qwen3.5
These are candidates, not performance claims. Run them against your evaluation set.


