⌬ OpenEnv · Multi-agent · Long-horizon

Twelve good decisions in a row.

Neon Syndicate is a cinematic OpenEnv environment where an LLM negotiates coalitions, scouts sectors, executes a coded operation, and extracts — all under partial observability. Most benchmarks reward one good answer. This one rewards a sequence.

env: checking… model: loading… 12-turn budget 5 reward gates 6 missions · easy → hard TRL PPO · Qwen 2.5 0.5B
🤗 Open the HF Space → ▶ Run training in Colab ⌥ Browse the code on GitHub
Innovation · 40%

The Policy Ladder

Same task, four policies, one axis. The trained LLM doesn't run in a vacuum — every claim is grounded against random and heuristic baselines.

Average task score on task_easy_docklands_relay

N = 4 policies · 1 episode each

Numbers come from server.environment.NeonSyndicateEnvironment via the same /agent/act + /step pipeline judges will run.

What each policy is

Random

Sanity floor — uniform action sampler. Confirms the env actually rewards strategy.

Trained (raw)

TRL PPO Qwen 2.5 0.5B. Honest about being undertrained — falls back when JSON is broken.

Hybrid (★ ours)

LLM proposes; env-rule guardrails reject obvious mistakes. The dashboard shows every override.

Heuristic

Hand-coded curriculum — the baseline ceiling for this env. Reliably reaches 0.83 task score.

Reward improvement · 20%

Real training run, episode by episode

Plotted live from training_summary.jsonl. Hover any point to see the per-episode breakdown — total reward, task score, success, steps.

data source: …
Storytelling · 30%

Watch the trained model play, live

One episode, eight seconds. Pick a policy, hit Run. The dashboard shows the action it picks, the env's reaction, and which guardrails fired.

policy: hybrid task: — turn 0/12 cum 0.00 —
🎮 Play it yourself

Operator's last thought

Press Run preview — the model fires up on demand and a one-episode rollout streams back over SSE.
awaiting

Reward components (last step)

Alliance
+0.00
Resources
+0.00
Operation
+0.00
Extract msg
+0.00
Extract
+0.00
⌬ Reward Forensics Lab

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

Three things on one page no other submission will ship: a procedural task generator with a 1→5 difficulty knob (RLVE), a live 5-gate verifier timeline showing per-component reward decomposition, and a reward-hacking sandbox you can fire pre-baked exploits at to see the env defenses kick in.

→ Open the Lab → Read the Walkthrough
Rubric · how to grade us

Evidence, mapped to the judging criteria

Each criterion → the artifact that proves it. Click through if you want the receipts.

Why this is different

Six things no other submission will have

Built from the env primitives up. No gridworld clone, no chess wrapper, no synthetic data farm.

01

Hybrid policy with guardrails

Trained LLM proposes; env-rule guardrails reject illegal actions before they burn turns. Every override is shown to the user — judges literally see when the model was wrong.

02

Playable cyberpunk heist

Walk an isometric city as Vex, take orders from the trained LLM ("The Operator") in your ear. WASD + E. Try doing that in a Colab notebook.

03

Transparent reasoning panel

Per-turn: prompt sent, raw model output, parsed action, latency, fallback flag. The model's mind is on screen.

04

Multi-agent reputation

4 factions with reputation deltas, alliance state, sector-specific deploy bonuses. Real coalition math, not flavor text.

05

Composable 5-gate rubric

Each step gives partial credit on alliance / resources / operation / extraction-message / extraction. Reward hacking is provably hard.

06

Race-mode SSE

Watch all 4 policies run the same task, lock-step, on one screen via Server-Sent Events. The bar chart resolves itself in 30 seconds.

Committed evidence

Static plots from the same run

If anything above fails to load on the deployed Space, these PNGs are committed in artifacts/ and embedded in the README.

Reward curves: random vs heuristic over episodes
artifacts/reward_curves.png — random vs heuristic, average reward by episode.
TRL PPO loss curve
artifacts/loss_curve.png — PPO loss across the committed training run.