AI4Science
Reconstruct · liveAI4Science composes PWM's verifiable protocol primitives into intelligent workflows. In Stage 1 the only live operator is Reconstruct, running the gap-TV solver on the CASSI and CACTI genesis benchmarks. Every query reads from the on-chain registry and writes back a certified result.
Stage 1 — only two genesis solvers are active
Today the only working operators on AI4Science are Reconstruct → gap-TV / CASSI and Reconstruct → gap-TV / CACTI. The Mismatch, Imaging-System Design, and Sci-Sim operators — and every other Reconstruct pairing — are deactivated until the genesis catalog of Principles, Specs, Benchmarks, and Solutions is verified on mainnet.
Contribute a Principle / Spec / Benchmark / Solution → to expand what AI4Science can run.
How it works
Pick a benchmark
In Stage 1, choose CASSI or CACTI. The Reconstruct operator resolves the L1 principle and the gap-TV L4 solver automatically.
Provide inputs
Supply your data, constraints, or a natural-language description. The operator resolves the relevant L1/L2/L3 registry entries automatically.
Receive a certificate
Every completed query produces an on-chain L4 certificate with PSNR, SSIM, provenance, and a permanent content address — reproducible by anyone.
The Reconstruct operator
Reconstruct
Consumer ● Live — gap-TV / CASSI & CACTI onlyApply a top-ranked solver to your own measurement data and receive a verifiable reconstruction certificate. In Stage 1, only the gap-TV solver paired with CASSI or CACTI is active.
Selects the top-N solvers from the benchmark leaderboard (L4) for the chosen imaging modality (L1) and runs them on your uploaded measurements.
Genesis solvers — CASSI & CACTI
"I have CASSI (snapshot-spectral) measurements from a coded-aperture spectrometer. Reconstruct with the genesis gap-TV solver and return a verifiable L4 cert."
→ Resolves: L1 compressive-sensing principle (TV prior)
→ Picks: L4 gap-TV / CASSI & L4 gap-TV / CACTI as genesis solvers
→ Returns: ranked L4 certs with PSNR / SSIM / SAM
CASSI = Coded Aperture Snapshot Spectral Imaging · CACTI = Coded Aperture Compressive Temporal Imaging. Both are genesis benchmarks; the GAP-TV (Generalized Alternating Projection with Total Variation) solver ships as their reference L4 baseline.
Mismatch diagnosis, imaging-system design, and science simulation operators are coming in Stage 2 — once the genesis catalog of Principles, Specs, Benchmarks, and Solutions is verified on mainnet. For now, AI4Science runs Reconstruct only.
Build on the open agent: github.com/integritynoble/AI4Science ↗
Two ways to start
Operator-first (you are here)
Pick an operator above, then the UI guides you to select the relevant L1/L2/L3/L4 inputs from the registry.
Best when you know what you want to do but haven't picked a benchmark yet.
Data-first (from any benchmark or principle page)
Browse to a benchmark or principle, then use inline operator buttons — "Design a system for this benchmark", "Run Sci-Sim for this principle", etc.
Best when you're already exploring the registry and want to act on what you find.
Both paths produce identical QueryContext inputs to the PWMQuery.escrow() contract — the chain sees the same call regardless of entry path.
Compositional patterns
Design → Publish → External miners compete
Day 0 · Run Design against L1-003 (CASSI) with your constraints → draft L2 + L3 + L4
Day 0 · Publish → new chain registered on-chain; you own L2 + L3 authorship
Day 1+ External miners discover the new L3 benchmark and submit L4 certs via Reconstruct
Day 30 Leaderboard fills; you earn royalties on each cert (5% L2 + 15% L3 share per query)
Reconstruct → Mismatch → Refine L2 → Re-Reconstruct
Step 1 · Reconstruct(L1-CASSI, top-3 L4s, your_data) → 3 cert candidates
Step 2 · Best result has artifacts → Mismatch(L1, L3, best_L4, your_data)
Step 3 · Diagnostic: "8-bit ADC violates L2's 12-bit assumption"
Step 4 · Design a refined L2 variant supporting 8-bit ADC → new L2-CASSI-8BIT
Step 5 · Reconstruct against new L2 → cleaner cert, 6 dB improvement
Sci-Sim → Publish as L3 → Inverse-problem L4 challenge
Step 1 · Sci-Sim(L1-NavierStokes, L2-custom-BCs) → synthetic data as L4 cert (simulation run)
Step 2 · Publish → simulation result registers as new L3-NS-CUSTOM-01 benchmark
Step 3 · New L3 poses inverse problem: given measurements, recover the boundary conditions
Step 4 · Miners run Reconstruct against the new L3, producing inverse L4 solvers
Revenue split per L4 mining event
Every certified L4 event automatically routes value upstream to the L1 / L2 / L3 authors. Numbers below match the whitepaper Token Economics section.
55%
Miners
AC + CP combined
15%
T_k treasury
per-Principle pool
15%
L3 Benchmark
creator royalty
10%
L2 Spec
author royalty
5%
L1 Principle
creator royalty
Example at a 20 PWM draw → Miners 11.00 · T_k 3.00 · L3 3.00 · L2 2.00 · L1 1.00.