Demos & Tutorials
Two walkthroughs: how to mine PWM tokens (earn first), and how to use AI4Science (spend after earning). New here? Start with the mining demo.
Demo 1 — Mine & Get PWM
5 steps · ~30 minutes for your first submission
Browse open benchmarks — find one with few solutions
Go to /mine and look at the L4 Solutions tab. Benchmarks with 0 solutions (🔥) have the most reward potential. Click any row to open the benchmark detail page.
→ https://testpwm.platformai.org/benchmark/cassi
→ Download public test set (y_measurements.npy, H_mask.npy)
→ Note the scoring metric: Score = 0.5×clip((PSNR−15)/30) + 0.5×SSIM
Tip: Start with a modality you know — CT, MRI, or CASSI. The public test set contains the measurements and forward model; your job is to reconstruct x from y.
Run your reconstruction algorithm
Use any method — classical iterative (ISTA, ADMM, TV) or deep learning (pretrained model, zero-shot diffusion). The benchmark forward model is fully defined in the spec, so no training on test data is needed.
import numpy as np
from skimage.restoration import denoise_tv_chambolle
y = np.load('y_measurements.npy') # (180, 182)
H = np.load('H_mask.npy') # (31, 128, 128)
# ADMM-TV loop (simplified)
x_hat = admm_tv_reconstruct(y, H, lam=0.01, n_iter=100)
from skimage.metrics import peak_signal_noise_ratio, structural_similarity
x_true = np.load('x_true_public.npy') # only in public tier
psnr = peak_signal_noise_ratio(x_true, x_hat, data_range=1.0)
ssim = structural_similarity(x_true, x_hat, data_range=1.0)
print(f'PSNR: {psnr:.2f} dB SSIM: {ssim:.4f}')
Check with an AI agent before submitting
Open ChatGPT, Claude Code, or Gemini. Paste the AI check prompt from the /mine page and append your method description + results.
ChatGPT (GPT-4o)
"PASS — ADMM-TV reconstruction on CASSI is physically correct. PSNR 32.1 dB is consistent with the method complexity. Notation follows PWM L4 template."
Claude Code
Run: claude /check my_solution.md → Reviews physics, checks L4 template compliance
Gemini Advanced
"PASS — method description is reproducible. Forward model matches CASSI L2 spec. Suggest citing Wagadarikar et al. 2008."
Copy the AI's PASS/FAIL response — you'll paste it into the submission form.
Submit on /mine
Go to /mine → Step 3: Submit. Fill in the form:
- Layer: L4 Solution
- Parent: the benchmark name (e.g. "CASSI Blind Reconstruction Challenge")
- Artifact content: method name, PSNR, SSIM, brief description, code link
- AI check result: paste the AI's PASS response
- Wallet address: your Base wallet (rewards sent here)
## L4 Solution: ADMM-TV on CASSI
**Parent**: CASSI Blind Reconstruction Challenge (L3)
**Method**: Split Bregman ADMM with TV regularization
**PSNR**: 32.1 dB **SSIM**: 0.8742
**Code**: github.com/yourname/cassi-admm
**Reference**: Wagadarikar et al., CASSI 2008
Founder reviews → mainnet → PWM in your wallet
After submission, the founder reviews your artifact for physics correctness, novelty, and template compliance. Approved submissions are deployed to Base mainnet. Once on-chain, your L4 cert earns PWM automatically as users query AI4Science.
Review
1–3 days
Physics + novelty check by founder
Mainnet deploy
Same day
Registered on Base by deployer EOA
Earn PWM
Per query
Auto-distributed on cert finalization
Spend your PWM: once you have PWM, use it on AI4Science to run Reconstruct, Mismatch, Design, or Sci-Sim operators on your own scientific imaging problems.
Demo 2 — Use AI4Science
4 operators · requires PWM tokens · runs in-browser
PWM required: AI4Science operators cost PWM per query. Mine your first PWM using Demo 1 above before proceeding.
Open AI4Science — describe your scientific problem
Go to /ai4science. The left panel lists 4 operators. The center is a chat interface. The right panel shows the active spec.
I have CASSI measurements of a hyperspectral scene.
The mask is random binary, 128×128 pixels, 31 spectral bands.
I want to reconstruct the full datacube.
→ AI selects: Reconstruct operator (L4)
→ Spec: CASSI · M·W·Σ·D forward model
→ Best solution: rank-1 cert on CASSI benchmark
Choose an operator — 4 options
Find the best algorithm for your forward model
e.g. "I have CT sinogram data, find the best FBP variant"
Detect calibration errors in your imaging system
e.g. "My MRI k-space has aliasing — is my sampling pattern wrong?"
Design an optimal new imaging system from scratch
e.g. "Design a CASSI-like system optimized for 64-band spectral scenes"
Simulate physics-accurate measurements from a model
e.g. "Simulate CT projections for a Shepp-Logan phantom at 60 angles"
Upload your data — get SOTA results instantly
AI4Science routes your data to the highest-ranked certified solution for your imaging system. Results include PSNR/SSIM estimates, a reconstructed image, and a link to the on-chain cert.
Example response from Reconstruct operator:
✓ Matched spec: CASSI (M·W·Σ·D, 128×128, 31 bands)
✓ Using cert: 0x3fa2…b91c (rank-1, PSNR 38.2 dB)
✓ Reconstruction method: HybridCascade++ (joint TV + deep prior)
✓ Estimated PSNR on your data: ~35–38 dB
✓ Cost: 0.01 PWM · Cert finalization: 7-day window
Verify on-chain — your query is a permanent cert
Every AI4Science query posts an L4 Certificate on Base. You can verify the cert independently or challenge it within 7 days if you believe the result is wrong.