User repository on GitHub
Eikos-27B vs Jev: live paper trading on Hyperliquid. Real prices, simulated money, rules hashed before the start.
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Practices & Patterns
Eikos Arena
Paper-trading arena where Eikos-27B and Jev get the same market view and 28 typed questions every five minutes across 14 Hyperliquid perpetuals with simulated fills.
- Stars
- 8
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- caiovicentino
- Last verified
- Oct 2026
About this project
Two decision models trade 14 Hyperliquid perpetuals with $10,000 of paper money each: Eikos-27B (open weights, MIT) against Jev (TypeSafe's decision API). Every 5 minutes both get the same market snapshot and the same 28 questions. Real prices, simulated money, no orders. Results: eikos-arena.vercel.app · Data: datasets/caiovicentino1/eikos-arena · Model: caiovicentino1/Eikos-27B · Model code: caiovicentino/eikos. Python 3.9 or newer. The engine uses only the standard library.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Uses real prices with simulated money and fee-based fills, so speed never helps.
- Scores equity plus Brier error on direction with cross-margin accounting.
- Freezes and hashes rules and engine before the first decision.
- Ships a local dashboard, resume-from-state engine, and sanitized publishing.
Watch out
A stress test, not a trading product and not financial advice; outputs are for evaluation only and never for training. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
No. Prices are real but money is simulated and no orders are placed, and the author frames it as a rule-following stress test rather than a strategy you should run with funds.
Eikos called direction more often but with overconfident probabilities on a near-random target, and it traded about 4.5 times as often, so worse Brier error and higher fee drag beat its hit rate.
The arena engine is light and the model is not. It is standard-library Python 3.9 or newer with nothing to install, but you serve Eikos yourself: Eikos-27B in bf16 fits one 96 GB GPU and Eikos-4B runs on a smaller one, while any other TypeSafe-compatible /v1/evaluate endpoint works instead. Jev comes from TYPESAFE_API_KEY or AI_GATEWAY_API_KEY. check_setup.py sends one tiny typed probe to each side before start_arena.sh freezes and hashes the rules into a 72-hour run on a dashboard at 127.0.0.1:8340.
Keys come from the environment or a restricted key file, requests log without headers, publishing uploads only whitelisted fields, and any deny-file string blocks the publish.
Jev finished ahead on both account value and forecast error. Jev ended at $10,086.55, a return of +0.87%, with a Brier of 0.249, a 47.4% hit rate and 31 trades. Eikos-27B ended at $9,659.56, a return of −3.40%, with a Brier of 0.266, a 51.6% hit rate and 141 trades. The match ran from 2026-09-24 05:12 UTC to 2026-09-27 05:15 UTC.
For each of the 14 markets, five crypto perps and nine HIP-3 perps on the xyz dex, the snapshot takes the best bid and ask, the last twelve five-minute and twenty-four one-hour closes, price changes, funding, open interest, volume, spread and taker fee. Both models then get the same state text and the same 28 typed questions, pos_ and up_ per market at $2,000 a position, with only each player's own portfolio lines differing. Fills happen at that snapshot's bid and ask with the taker fee, so answering faster never helps.
A stopped engine resumes from state: cd into the run folder and run python3 arena.py --run-dir . with the same settings, and it picks up from arena_state.json. start_arena.sh copies the rules and the engine into a new folder under runs/ and hashes them into PREREG.sha256 before the first decision, so a run cannot quietly change the rules it is scored under. RULES.md for the official match hashes to 943d01e0.
More Jev resources like this
24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePython39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
mithalouni/system-one-openPython
mithalouni/system-one-openPython941GitHub stars
Type-safe generation layer over existing autoregressive models via JSON Schema on SGLang, returning constrained string, numeric, boolean, and enum answers with optional thinking.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
zmtomorrow/TypeARPython
zmtomorrow/TypeARPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.