[snimu]

Adds a ### Rewards subsection to Environment Design — the same section as RL-Residency/residency-environments#83, for this repo's copy of the AGENTS.md:

• Bounds: rewards are floats in [0, 1], inclusive; each reward function returns [0, 1] and weights are non-negative. Multi-environment training in prime-rl does no reward re-scaling, so a reward outside [0, 1] silently changes the relative weighting of that environment's gradients; GRPO's group normalization makes advantages negative or positive from group statistics, so the reward range never needs to encode direction.
• Binary by default: 1 only when the task is genuinely complete, verified by the most authoritative mechanism available (kernel, simulator, upstream evaluator), else 0. Partial or shaped rewards are the exception and need a stated reason in the environment's PR, with every threshold documented; observability-only partial signal belongs in metrics, not rewards.

---

Note

Low Risk
Documentation-only change to contributor guidance; no runtime, auth, or training code is modified.

Overview
Adds a new ### Rewards block under Environment Design in AGENTS.md, right after the tool-description prompting rules.

It standardizes reward bounds: values and per-function outputs stay in [0, 1], weights are non-negative, with rationale tied to multi-env prime-rl training (no re-scaling) and GRPO (direction comes from advantages, not the reward sign/range).

It also sets binary completion rewards as the default—1 only when completion is verified by the strongest check available, otherwise 0—and requires explicit PR justification and documented thresholds for any partial/shaped rewards, keeping observability-only signal in metrics instead of rewards.

^{Reviewed by Cursor Bugbot for commit be0d9c8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add rewards guidance section to AGENTS.md (0-1 bounds, binary by default)

New section in AGENTS.md sets reward design rules for agents: rewards must be floats from 0 to 1, weights non-negative, and binary rewards tied to authoritative verification are the default.

• Partial or shaped rewards require PR justification and documentation; observability-only signals belong in metrics, not rewards.
• Includes rationale on reward scaling, GRPO group normalization, and reward hacking.

^{Macroscope summarized be0d9c8.}