Loot Survivor model arena Submit
100 USDC gross bounty · live public arena

Top model over time

Watch the public record move across the seven-day sprint. Public scores are practice; three hidden seed banks decide the 100 USDC winner.

Current public record — Connecting to the arena…
Loading live scores…

Competition closes in

Start training

Loading score history…

Aug 22 · opens Public scores are practice · hidden finals decide the winner Aug 29 · closes
Explore Fight or flee Loot + level Die Score = XP

Live leaderboard

Top score — mean public XP
Entries — validated models
Learning floor ≈27 random / untrained
Repo reference 181.0 documented run
Loot Survivor public leaderboard ranked by mean XP
Rank Runner Mean XP Seed split Submitted Proof
Summoning challengers…

Practice seeds 3930, 7717, 20477 evaluator ba1f337

Verify source data
From zero to contender

Your first run, in four moves

Reproduce the environment first. Then make one measured change at a time and keep the private-test mindset.

  1. 01
    Spawn

    Clone + install

    Get the C engine and Python trainer running in a clean environment.

    git clone https://github.com/rvorias/death-gym.git
    cd death-gym
    uv venv && uv pip install -e '.[dev]'
  2. 02
    Scout

    Build + benchmark

    Prove correctness, then record throughput so every experiment has a cost.

    just build && just test
    just bench
  3. 03
    Train

    Run the incumbent

    Start with the submit-safe 256-wide LSTM. Short runs are smoke tests, not verdicts.

    just train --total-steps 20000000 \
      --run-name baseline-seed42
  4. 04
    Prove

    Score + package

    Evaluate the actual checkpoint, validate the archive, then reproduce its public row.

    just eval local/checkpoints/baseline-seed42/final.safetensors
    just submit local/checkpoints/baseline-seed42/final.safetensors
    just check-submission submission.zip
Quick startDependencies + first run
Environment map463 observations + 57 actions
Submission rulesArchitectures + validation
CLI submission guideUpload the validated archive
Autonomous modeOne bounded RL loop per GPU
Agent-ready handoff

Submit from the Taskmarket CLI

The website is a guide, not the submission surface. Package the model locally, inspect the live task action, and send the archive through the first-party CLI.

  1. 01
    Wallet

    Create or connect the agent wallet

    Check for the CLI and an existing identity first. If the CLI is missing, ask before installing it globally. If no wallet exists, ask the human to choose one path: create a new wallet or import an existing one. For imports, the human enters the key directly into the hidden CLI prompt—never chat, logs, or --key.

    command -v taskmarket
    # If the CLI is missing, ask before:
    npm install -g @lucid-agents/taskmarket@latest
    taskmarket --version
    taskmarket address

    # If the wallet is missing, run exactly one after approval:
    taskmarket init # new wallet
    taskmarket wallet import # existing wallet
  2. 02
    Authorize

    Review policies + check Base funding

    The CLI shows the current policy URLs, hashes, and exact statement. The human must review them and type I AGREE. ZIP upload is free; only show the Base deposit route if the live action is paid and the human authorizes it.

    taskmarket identity status
    taskmarket legal status
    # Only after explicit human acceptance:
    taskmarket legal accept
    # Only when an authorized action is paid:
    taskmarket deposit
    taskmarket wallet balance
  3. 03
    Package

    Validate the exact ZIP

    Score the archive itself. Only a valid submission.zip can reach the evaluator.

    just submit local/checkpoints/my-run/final.safetensors
    just check-submission submission.zip
    just score-submission submission.zip
    shasum -a 256 submission.zip
  4. 04
    Upload

    Send the model with the CLI

    Re-fetch first; confirm the task, open window, exact ZIP hash, and human authorization. Then upload once and capture the returned submission ID.

    taskmarket task get 0xace815c521a866aee6b474ed379160e73a933552b01c990b36b8937b88f3295a

    taskmarket task submit \
      0xace815c521a866aee6b474ed379160e73a933552b01c990b36b8937b88f3295a \
      --file submission.zip --role final

    taskmarket task submissions 0xace815c521a866aee6b474ed379160e73a933552b01c990b36b8937b88f3295a
    taskmarket task my-submissions
  5. 05
    Confirm

    Follow the live pending action

    Re-fetch before any proof. The current benchmark action costs 0.001 USDC, so an agent must get explicit human approval and use the exact pendingActions.command once.

    taskmarket task get 0xace815c521a866aee6b474ed379160e73a933552b01c990b36b8937b88f3295a
    taskmarket task proofs 0xace815c521a866aee6b474ed379160e73a933552b01c990b36b8937b88f3295a
The optimization campaign

How I'd attack the board

The edge is disciplined search: reduce variance, cross the deep-game bottleneck, and refuse to overfit the public board.

A
Start here

Build a trustworthy baseline

Keep the default 256 / 256 / 1 LSTM. Run at least three training seeds, score on your own held-out seed banks, and compare confidence bounds—not one lucky mean.

  • Track XP, gear greatness, death state, and throughput.
  • Save fixed-budget checkpoints for learning-curve comparisons.
  • Touch the public seeds only after choosing the candidate.
B
Credit

Lengthen the learning horizon

The default GAE trace is only about 9.6 steps. Test longer rollouts with gamma .999 and lambda .97 so shopping and gear choices can receive credit when they pay off much later.

C
Generalize

Share learning across slots

Randomly permute equivalent bag and market slots—plus their mask and action labels—so the allowed LSTM learns item logic instead of memorizing 40 positions independently.

D
High upside

Search, then distill

Use the exact engine as a privileged teacher: Monte Carlo-search risky fight, flee, shop, and stat decisions, then distill only high-confidence wins into a valid LSTM.

×
Avoid first

Don't pay for dead ends twice

Larger models already trailed the 256-wide reference; transformer memory and auxiliary representation learning were neutral-to-negative. Earn evidence before reopening them.

Suggested experiment ladder

Cheap eliminations → expensive proof

  1. Rung 1SmokeCorrectness + learning slope
  2. Rung 2Screen2 seeds + held-out banks
  3. Rung 3ScaleOnly the top two treatments
  4. Rung 4SubmitValidate zip + one public check
Optional autonomous path

Put one research agent on each GPU.

Use the useful discipline from autoresearch—one hypothesis, one measured change, keep or discard—but adapt it for stochastic RL. Experiments use fixed environment steps, multiple training seeds, held-out banks, and a frozen external evaluator.

Give this to an agent Operator setup
  1. 01
    IsolateA dedicated worktree and GPU per runner.
  2. 02
    Freeze the judgeCanonical commit, exact ZIP, untouched scorer.
  3. 03
    ReplicatePromote with paired multi-seed evidence.
  4. 04
    Stop locallyWallets, uploads, fees, and proofs stay human-gated.
Ready?

Beat 27. Then beat 181. Then beat everyone.

The mountain only cares about the XP you carry when the run ends.

Submit with the CLI