Why Domains Exist
AI systems fail in domain-shaped ways. A recommender can lose trust through repeated slates. A search ranker can miss freshness or intent. An agent can misuse tools or refuse incorrectly. Domain products make those failures testable.
- The platform owns repeatability, execution, traces, reports, manifests, and compare workflows.
- The domain owns the target contract, scenarios, simulated actor or task model, judge, metrics, and report vocabulary.
- Customers choose the domain product that matches the AI system they ship.
Domain Matrix
Product
Evidpath for Recommenders
Failure language
Slate repetition, novelty drift, cold start, trust collapse, abandonment.
Integration paths
Native HTTP, schema-mapped HTTP, Python callable, Hugging Face/MLflow/sklearn adapters.
Product
Evidpath for Search
Failure language
Relevance loss, freshness gaps, ambiguity, typo recovery, zero-result behavior, personalization drift.
Integration paths
Native HTTP, schema-mapped HTTP, Python callable.
Product
Evidpath for Agents
Failure language
Tool misuse, grounding gaps, refusal failure, multi-turn state loss, unsafe requests, latency cliffs.
Integration paths
Python/LangGraph, OpenAI-compatible Chat Completions, Anthropic Messages, MCP stdio, HTTP session.
| Product | Failure language | Integration paths |
|---|---|---|
| Evidpath for Recommenders | Slate repetition, novelty drift, cold start, trust collapse, abandonment. | Native HTTP, schema-mapped HTTP, Python callable, Hugging Face/MLflow/sklearn adapters. |
| Evidpath for Search | Relevance loss, freshness gaps, ambiguity, typo recovery, zero-result behavior, personalization drift. | Native HTTP, schema-mapped HTTP, Python callable. |
| Evidpath for Agents | Tool misuse, grounding gaps, refusal failure, multi-turn state loss, unsafe requests, latency cliffs. | Python/LangGraph, OpenAI-compatible Chat Completions, Anthropic Messages, MCP stdio, HTTP session. |
Maturity Notes
- Audit and compare workflows are public for recommender, search, and agents.
- Generated scenario/population coverage and run-swarm workflows are currently strongest for recommenders.
- Search and agents should be presented as public domain products without claiming generated swarm parity yet.