Papajams/conviction-gym

Conviction Gym

An RL environment for conviction-based crypto trading plus a Beam agent-trace dataset — built for the Build with Beam challenge.

Unlike a backtest where the strategy is baked into the loop, this environment owns only the physics (mark-to-market, 1% slippage, gas, bankroll constraints). Every decision — enter / exit / partial / hold — is an agent action. It works for classic RL policies (structured Dict observation, MultiDiscrete([K,4,3]) action space) and LLM agents (env.render() emits a compact text market snapshot per step).

Built from the production Early, Not Wrong autonomous trading agent.

What's inside

path	contents
episodes.json	150 real market days (2026-05-26 → 2026-10-22) × 26 BSC tokens — CoinGecko prices/mcap/volume + alternative.me Fear & Greed, kline slices for RSI(14)
traces/	360 Beam rollout steps as JSONL: observation_text + reasoning + action + reward + token usage, tagged by reasoning_effort (minimal…max) and max_tokens budget (1024 / 8192)
code/	The environment: TypeScript reference impl + cross-validated Python Gymnasium port (pip install -e code/python)

Headline findings

Beam (reflection/Beam-1.0-501B-A23B) played the env as its policy — one JSON action per market day.

1. The completion-budget trap hits agentic loops. At max_tokens=1024, high/max effort truncated on 23–24/30 steps (finish_reason=length, empty answer → forced hold). The trap isn't a QA-only phenomenon — on sequential tasks it silently turns "think hard" into "do nothing" (and here it helped the score: doing nothing beat Beam's own trading).

2. With a fair budget the effort dial is non-monotonic. At max_tokens=8192 on the choppy eval window:

effort	return	trades	reason tok/step
minimal	−4.9%	24	201
low	−4.1%	12	680
medium	−6.5%	34	1,589
high	−5.0%	26	2,395
max	−4.2%	28	2,883

More thinking → more churn, not more return. Every level underperformed the conviction heuristic (−1.7%), random (−1.9%) and buy & hold (−1.9%).

3. In a trending window Beam beats the heuristic. On the window where the production conviction strategy made +9.6%, Beam at max made +12.9% — it traded less and held longer. Beam's trading skill is regime-dependent; the dial buys different decisions, not reliably better ones.

Caveats: single window per cell, one rollout per level, temperature 0.2, assisted mode. Every cell is rerunnable (run-beam.ts --sweep <level>).

Quickstart

import json, conviction_gym as cg
specs = json.load(open("episodes.json"))
env = cg.ConvictionGymEnv(specs[0], {"mode": "assisted"})
obs, _ = env.reset()
obs, reward, done, _, info = env.step(env.action_space.sample())

Run Beam yourself (any OpenAI-compatible endpoint):

pip install -e code/python
python -m conviction_gym.rollout --policy beam --effort max \
    --episodes episodes.json --traces my-traces.jsonl

Trace schema

{"episode_id": "bsc-cg-…-w4", "policy": "beam:Beam-1.0-501B-A23B",
 "reasoning_effort": "max", "step": 12, "date": "2026-10-05",
 "observation_text": "Day 13 (2026-10-05) | regime: …",
 "reasoning": "…", "action": {"symbolIndex": 3, "op": "enter", "size": "standard"},
 "reward": 0.41, "portfolioValueUsd": 993.2,
 "usage": {"promptTokens": 2140, "completionTokens": 1180,
           "reasoningTokens": 1050, "latencyMs": 5120}}