Summary

The next JeTech Lab research step is not another algorithm in isolation. The work should be split into four research axes:

A. World model market generator
B. Offline RL algorithm
C. Time-series pattern representation
D. Real market gate

These axes only matter when connected. A better world model still fails if the action dataset is weak. A new offline RL algorithm fails if synthetic paths are unrealistic. A larger encoder fails if the reward and gate disagree.

The research loop should stay fixed:

real OHLC
  -> world model training
  -> synthetic market paths
  -> synthetic action/position dataset
  -> offline RL algorithm
  -> synthetic eval
  -> real 5Y gate
  -> failure taxonomy
  -> next dataset or model change

1. World Model Research

The world model should not be evaluated as a pure price forecasting model. Its job is to create a market distribution that can train offline RL policies.

Current definition:

input:
  past 224 days OHLC

output:
  future 96 days OHLC path distribution

This is generative learning, not RL. The full pipeline is world-model-assisted offline RL because generated paths become the base of the transition dataset.

First Step: Quality Report

Every world model run should emit a quality report.

Lower loss does not prove a better training environment. The model must preserve market features that affect policy learning.

Second Step: Regime-Conditioned Generation

Add explicit conditions:

volatility_bucket = low / normal / high / crisis
trend_bucket = up / range / down
drawdown_bucket = near_high / mild / deep
asset_context = crypto_beta / equity_beta / risk_off

Offline RL should see rare but important regimes. Too few crash samples teach leverage. Too many teach flat exposure. The generator needs balanced oversampling.

Third Step: Joint Transition Direction

The current pipeline separates price and action:

price path -> action labeler -> reward transition

The medium-term target is:

price, action, reward generated together
or price generated first and actions inferred by behavior priors

JeTech should not jump straight into a large diffusion world model. A better first move is regime-aware action labeling, then action-conditioned or trajectory-level generation.

2. Offline RL Algorithm Research

Algorithm work should use CORL as the reference. JeTech implementations should be comparable to CORL for CQL, IQL, TD3+BC, Decision Transformer, and ReBRAC.

Research order:

This is the JeTech operating order, not a universal academic ranking.

Algorithm comparisons must fix:

world_model_version
generated_path_seed
synthetic_action_labelers
reward_config
encoder_backbone
train_steps
real_gate_period

If any of these change, the experiment compares datasets or configs, not algorithms.

Diffusion Policy Entry Criteria

Diffusion policy work should start when:

1. the same state bucket contains multiple high-reward action modes, 2. deterministic actors collapse to average exposure, 3. IQL or ReBRAC is stable but cannot use action diversity, 4. sampled action candidates improve the real gate.

The first diffusion role should be candidate generation or behavior prior, not live deployment.

3. Time-Series Pattern Research

The encoder should turn OHLC paths into usable market state representations.

JeTech keeps the price-only rule:

input = OHLC window
no handcrafted alpha leakage
encoder learns representation

The encoder still needs to capture:

Candidate encoders:

Encoder research should isolate the encoder. Keep the offline RL algorithm and dataset fixed.

Representation probes should test whether the encoder separates useful states:

next 20d realized volatility bucket
next 20d drawdown bucket
trend/range regime label
tail event probability
simple strategy winner class

These probes are diagnostics, not live trading signals.

4. Dataset Generation Policy

In offline RL, the dataset is as important as the algorithm. The policy cannot reliably learn actions that are absent from the dataset.

Labeler types:

Oracle-like labelers should carry leakage metadata and stay out of default training datasets.

Every dataset should have a card:

dataset_id
world_model_version
generation_seed
symbol_set
date_source
regime_mix
labeler_mix
reward_config
cost_config
transition_count
quality_metrics
leakage_flags

Without dataset cards, experiments cannot be reproduced.

5. Real Gate Research

The final metric is not synthetic score. It is real market survival.

Gate metrics:

Failures should be classified:

This turns the training loop into research instead of repeated reruns.

Six-Month Sequence

The order is conservative by design. Measure first, then add algorithms. Without measurement, diffusion and larger sequence models make failures harder to diagnose.

Research Position

The core question is:

Which synthetic market distribution and offline RL algorithm combination
survives the real market gate most consistently?

World models, offline RL algorithms, encoders, and dataset policies are subproblems of that question. Improve each axis separately, then bind them with the same real gate.