Design JEPA Models
Use this skill to turn a vague JEPA idea into a research-grade design spec: data, context, target, representation, masking, architecture, collapse prevention, probes, validation, and failure modes.
Workflow
Classify the request branch.
- New model design or architecture review: read
references/core-jepa-design.md.
- World-model, planning, objectives, actions, or LeCun lecture framing: read
references/world-model-lessons.md.
- Trading, finance, probes, labels, baselines, leakage, or backtesting: read
references/evaluation-and-trading.md.
- CPU toy experiment, tutorial implementation, or micro prototype: read
references/toy-experiments.md.
- Source-grounded explanation, paper comparison, or paper deep dive: read
references/source-map.md and references/paper-deep-dives.md, then the relevant paper file.
- Recap of the learned JEPA concepts: read
references/course-recap.md.
- LeCun world-model lecture/slides framing: read
references/lecun-lecture-digest.md.
Write the JEPA spec before proposing code.
Include these fields, marking unknowns explicitly:
- Data domain and observation unit.
- Context available at decision/inference time.
- Target hidden during pretraining.
- Latent abstraction the model should preserve.
- Nuisance detail the representation may discard.
- Masking strategy and shortcut risks.
- Context encoder, target encoder, predictor, and target-update rule.
- Loss, collapse-prevention mechanism, and collapse diagnostics.
- Probe labels, fair baselines, split protocol, and success criterion.
Hold the representation line.
Treat JEPA as representation learning unless the user explicitly moves to decision-making. A good JEPA result means the latent carries useful out-of-sample structure; it does not imply a tradable strategy, robot policy, or safe planner.
Use small heads as the first evaluation.
For a frozen encoder, prefer simple probes over full fine-tuning. For market data, start with future volatility/liquidity/imbalance buckets before noisy direction or PnL labels.
Finish with an audit.
Before finalizing any design, report:
- What could collapse.
- What could leak future information.
- What baseline must be beaten.
- What positive result would and would not prove.
- Which primary source most directly supports the design choice.
Output Shape
For design tasks, produce a compact spec with:
Objective:
Data:
Context:
Target:
Mask:
Architecture:
Target update:
Loss:
Collapse checks:
Probe labels:
Baselines:
Validation split:
Success means:
Does not prove:
Next experiment:
For review tasks, lead with findings ordered by risk:
Critical issues:
Major issues:
Missing controls:
Good design choices:
Minimal next fixes:
Guardrails
- Prefer representation-space prediction over raw reconstruction for high-dimensional, continuous, noisy data.
- Prefer mechanically defined labels and frozen probes before strategy or policy claims.
- Treat finance/trading designs as research methodology, not financial advice.
- Ground non-obvious claims in the source map or the user's local artifacts when available.
1---2name: design-jepa-models3description: Design, review, prototype, and evaluate JEPA models, joint-embedding predictive architectures, V-JEPA/I-JEPA-style representation learners, action-conditioned world models, and JEPA-inspired models for time series or trading research. Use when the user asks for JEPA architecture design, masking/target choices, collapse prevention, latent probes, toy JEPA experiments, market-data JEPA research, world-model planning, or critique of a JEPA paper/model/implementation.4---56# Design JEPA Models78Use this skill to turn a vague JEPA idea into a research-grade design spec: data, context, target, representation, masking, architecture, collapse prevention, probes, validation, and failure modes.910## Workflow11121. **Classify the request branch.**13 - New model design or architecture review: read `references/core-jepa-design.md`.14 - World-model, planning, objectives, actions, or LeCun lecture framing: read `references/world-model-lessons.md`.15 - Trading, finance, probes, labels, baselines, leakage, or backtesting: read `references/evaluation-and-trading.md`.16 - CPU toy experiment, tutorial implementation, or micro prototype: read `references/toy-experiments.md`.17 - Source-grounded explanation, paper comparison, or paper deep dive: read `references/source-map.md` and `references/paper-deep-dives.md`, then the relevant paper file.18 - Recap of the learned JEPA concepts: read `references/course-recap.md`.19 - LeCun world-model lecture/slides framing: read `references/lecun-lecture-digest.md`.20212. **Write the JEPA spec before proposing code.**22 Include these fields, marking unknowns explicitly:23 - Data domain and observation unit.24 - Context available at decision/inference time.25 - Target hidden during pretraining.26 - Latent abstraction the model should preserve.27 - Nuisance detail the representation may discard.28 - Masking strategy and shortcut risks.29 - Context encoder, target encoder, predictor, and target-update rule.30 - Loss, collapse-prevention mechanism, and collapse diagnostics.31 - Probe labels, fair baselines, split protocol, and success criterion.32333. **Hold the representation line.**34 Treat JEPA as representation learning unless the user explicitly moves to decision-making. A good JEPA result means the latent carries useful out-of-sample structure; it does not imply a tradable strategy, robot policy, or safe planner.35364. **Use small heads as the first evaluation.**37 For a frozen encoder, prefer simple probes over full fine-tuning. For market data, start with future volatility/liquidity/imbalance buckets before noisy direction or PnL labels.38395. **Finish with an audit.**40 Before finalizing any design, report:41 - What could collapse.42 - What could leak future information.43 - What baseline must be beaten.44 - What positive result would and would not prove.45 - Which primary source most directly supports the design choice.4647## Output Shape4849For design tasks, produce a compact spec with:5051```text52Objective:53Data:54Context:55Target:56Mask:57Architecture:58Target update:59Loss:60Collapse checks:61Probe labels:62Baselines:63Validation split:64Success means:65Does not prove:66Next experiment:67```6869For review tasks, lead with findings ordered by risk:7071```text72Critical issues:73Major issues:74Missing controls:75Good design choices:76Minimal next fixes:77```7879## Guardrails8081- Prefer representation-space prediction over raw reconstruction for high-dimensional, continuous, noisy data.82- Prefer mechanically defined labels and frozen probes before strategy or policy claims.83- Treat finance/trading designs as research methodology, not financial advice.84- Ground non-obvious claims in the source map or the user's local artifacts when available.
sssemil