---

name: reward_projection description: Project late-specified rewards into backward representation coordinates for immediate FB policy adaptation.

Reward Projection

Use this skill after an FB backward embedding has been learned and a reward is specified later. It should not update the reward-free embeddings or run new task-specific RL.

Inputs

• Backward embeddings B indexed by state or goal.
• Reward observations as {index, reward} records or a sparse goal index.

Outputs

• Task vector z_R as the weighted sum or average of backward embeddings.
• Provenance describing which reward entries contributed.

Workflow

1. Validate that all reward indices exist in B.
2. Compute z_R = sum_g B(g) r(g) for explicit finite rewards.
3. Preserve multi-goal rewards by summing their weighted backward vectors.
4. Return diagnostics and never mutate F or B.

Validation

Run python scripts/project_reward.py --self-test. The test checks sparse and composite rewards.

Limitations

For continuous goals this reduced script expects sampled reward observations rather than integrals over an unknown density.