How tianyucodings/jevharness works
JevHarness is a framework for authoring task-specific Jev decision harnesses, executing them against task observations, and optionally improving them from rewards and complete traces. The primary flow is task data and observations → validated pipeline execution and Jev judgments → task evaluation and recorded traces; optional GEPA reflection proposes and evaluates revisions, and selected pipelines can be frozen for repeatable execution. The README also documents crypto spot backtesting and paper execution. Edges use README or SOURCE INDEX evidence; unsampled subsystems are included where documented, without guessing their internal calls.
Main entry points: runtime.py Pipeline runtime · evolution.py GEPA evolution · spec.py Pipeline specification · storage.py Run storage
Components 26
Authoring and evolution
- GEPA evolution
auto_jev/evolution.py - Trace reflection
auto_jev/reflection.py - Trace codec
auto_jev/trace_codec.py
Harness execution
- Pipeline specification
auto_jev/spec.py - Expression interpreter
auto_jev/spec.py - Flow scheduler
auto_jev/flow.py - Sandboxed Python nodes
auto_jev/python_nodes.py - Pipeline runtime
auto_jev/runtime.py - Jev transports
auto_jev/providers.py - Vercel gateway
auto_jev/vercel.py
Data and evaluation
- Spot backtester
auto_jev/crypto.py - Observation builder
auto_jev/observations.py - Market and news data
auto_jev/data.py - Dataset management
auto_jev/datasets.py - Experiment runner
auto_jev/experiment.py - Paper execution
auto_jev/paper.py
Persistence and serving
- Frozen artifacts
auto_jev/frozen.py - Run storage
auto_jev/storage.py - Command line
auto_jev/cli.py - Run server
auto_jev/server.py
Inspection and deployment
- Archive viewer
website/worker/router.mjs
Other components
- Authoring LLM
- Task adapter
- Jev service
- Task environment
- Market data services
Connections 40
- Authoring LLM → Pipeline specification: authors
README.md - Task adapter → Pipeline runtime: provides observations
README.md - Pipeline runtime → Task adapter: returns actions
README.md - Task adapter → Task environment: applies actions
README.md - Task environment → GEPA evolution: provides rewards
README.md - Task environment → Trace reflection: provides traces
README.md - Pipeline specification → Expression interpreter: defines expressions
auto_jev/spec.py - Pipeline specification → Flow scheduler: compiles flow
auto_jev/spec.py - Pipeline specification → Sandboxed Python nodes: validates nodes
auto_jev/spec.py - Pipeline runtime → Pipeline specification: validates spec
auto_jev/runtime.py - Pipeline runtime → Flow scheduler: schedules nodes
auto_jev/runtime.py - Pipeline runtime → Sandboxed Python nodes: executes nodes
auto_jev/runtime.py - Pipeline runtime → Jev transports: requests judgments
- Jev transports → Jev service: sends requests
auto_jev/providers.py - Jev transports → Vercel gateway: uses transport
auto_jev/providers.py - Vercel gateway → Jev service: calls gateway
auto_jev/vercel.py - GEPA evolution → Pipeline runtime: evaluates candidates
auto_jev/evolution.py - GEPA evolution → Trace reflection: requests proposals
auto_jev/evolution.py - Trace reflection → Jev transports: uses provider
auto_jev/reflection.py - Trace reflection → Trace codec: encodes traces
auto_jev/reflection.py - GEPA evolution → Run storage: records search
auto_jev/evolution.py - Spot backtester → Pipeline runtime: runs pipeline
auto_jev/crypto.py - Spot backtester → Observation builder: builds inputs
auto_jev/crypto.py - Market and news data → Market data services: fetches market data
auto_jev/data.py - Experiment runner → GEPA evolution: runs search
auto_jev/experiment.py - Experiment runner → Run storage: records runs
auto_jev/experiment.py - Dataset management → Market and news data: uses data fetchers
auto_jev/datasets.py - Dataset management → Run storage: stores datasets
auto_jev/datasets.py - Paper execution → Market and news data: fetches inputs
auto_jev/paper.py - Paper execution → Spot backtester: uses backtest logic
auto_jev/paper.py - Paper execution → Pipeline runtime: executes decisions
auto_jev/paper.py - Frozen artifacts → Pipeline specification: validates specification
auto_jev/frozen.py - Frozen artifacts → Observation builder: binds contract
auto_jev/frozen.py - Command line → Experiment runner: starts experiments
auto_jev/cli.py - Command line → GEPA evolution: starts evolution
auto_jev/cli.py - Command line → Dataset management: manages datasets
- Command line → Frozen artifacts: freezes pipelines
auto_jev/cli.py - Command line → Run storage: reads and writes
auto_jev/cli.py - Run server → Run storage: serves run data
auto_jev/server.py - Archive viewer → Run storage: reads archives