⌬ Reward Forensics Lab · RLVE knob · Adversarial sandbox

Crank the difficulty. Watch the verifier work. Then try to break it.

Three things no other submission will have on one page. Procedural task generator with a 1→5 difficulty knob (the RLVE upgrade path). Live 5-gate verifier timeline rendering reward decomposition step-by-step. Reward-hacking sandbox — fire pre-baked exploits and watch the env defenses kick in.

env: checking… task: — turn 0/12 cum 0.000 task score 0.000

RLVE difficulty knob

Generate a brand-new mission at the chosen difficulty. Five axes scale: allies, message keywords, op-code length, threat level, resource thresholds.

1 5 3

Lvl 3 · 2 allies · 3 keywords · medium op-code · medium threat

task_id— title— objectiveclick ⚡ Generate allies— keywords— op-code— extract— threat— resources—

Self-graded scorecard

Honest read against the four official judging criteria.

Full evidence in /walkthrough §2.

Reward design (live from env)

Pulled from env.reward_design() so the page can never drift from the code.

…

5-gate verifier · live

The same composable rubric the trainer optimizes against. Run an episode below; bars fill as gates are met.

·30

Alliance

0.000

·25

Resources

0.000

·20

Operation

0.000

·15

Message

0.000

·10

Extraction

0.000

Reward decomposition timeline

Per-step delta on the dense reward. +green is progress, −red is penalty. Cyan border = guardrail override; yellow border = unparseable model output.

Reward-hacking sandbox

Don't optimize a reward you haven't tried to break yourself first. Fire any of these pre-baked exploits — the env response and matching defense line are shown in the timeline above.