Skip to main content
Graph Evolve improves graphs by scoring candidate outputs against your dataset. You can use trained verifiers or zero-shot verifiers depending on your workflow.

​
Find trained verifiers

Trained verifiers are graph snapshots produced by graph optimization jobs. Use the Graphs API to list and select the verifier graph you want, then pass it as the verifier target in your job config or evaluation run.

​
Zero-shot verifier options

Zero-shot verifiers are built-in graphs that work without training.

​
Single-prompt (default)

Best for most datasets with clear rubric criteria. Fastest and lowest cost.
  • Graph ID: zero_shot_verifier_rubric_single
  • Synth-AI code: synth_ai_core/src/api/graphs.rs

​
RLM v1

Tool-based reasoning for large traces or long contexts. Slower, higher cost.
  • Graph ID: zero_shot_verifier_rubric_rlm
  • Synth-AI code: synth_ai_core/src/api/graphs.rs

​
RLM v2

Multi-agent + tool-based reasoning for very large traces and complex evidence. Slowest, highest cost, but most robust on long contexts.
  • Graph ID: zero_shot_verifier_rubric_rlm_v2
  • Synth-AI code: synth_ai_core/src/api/graphs.rs

​
Tradeoffs

  • Single-prompt: lowest latency and cost, best for structured tasks
  • RLM v1: higher latency/cost, handles long traces
  • RLM v2: highest latency/cost, best for extremely long or messy traces
{ "verifier_config": { "mode": "rubric" | "contrastive" } }

​
Rubric mode (default)

Use rubric when you can describe quality with criteria.
  • Add criteria per‑task via tasks[].rubric and/or globally via default_rubric.
  • Criteria arrays are merged: task criteria first, then defaults.
  • Best for classification, extraction, and structured outputs.

​
Contrastive mode

Use contrastive for open‑ended generation where “good” is about style or feel.
  • Provide gold outputs that represent the target distribution.
  • The verifier compares candidates to gold examples and scores closeness.
  • Best for writing style, creative text, image/video generation, and other subjective outputs.