Proof-of-Inference Litepaper (v0.2)
Updated May 23, 2025
0. TL;DR
Mission: Convert on‑chain fees into real AI compute power.
Core Loop: Stake $AGENT → run AI jobs → Saturn Swap auto-buys $AGENT with each fee → epoch-roll rewards → slash on mis-behavior.
Key Assets: Newly‑acquired Saturn Swap order-book DEX (deep, instant liquidity), FluidTokens Aquarium (ADA-less gas tank) and community-run Edge / GPU / Datacentre pools for decentralised compute.
1. Problem Statement
AI dApps on Cardano suffer from three frictions: (1) Thin liquidity for niche tokens > Price spikes or slippage whenever users must buy the utility token. (2) ADA gas requirement > Creates a need to manage two tokens rather than a single utility token. (3) Lack of deterministic rewards for compute providers > GPU owners hesitate to provide capacity without deterministic yield. Meanwhile, DeFi yields rely on speculative farming rather than real utility.
2. Solution Overview
Inference Pools collateralised by $AGENT lock real GPUs/NPUs. Each inference job pays its fee in ADA; FluidTokens Aquarium transparently settles the gas.
70 % → paid to the pool operator.
28 % → swapped on Saturn Swap for $AGENT and streamed to stakers.
1 % → DAO treasury (retained as ADA for grants & audits)
1 % → FeeTank top-up (swapped to ADA, keeping gas-tank solvent)
Rewards roll up once per epoch; mis-reporting triggers on-chain slashing.
2.1 Why owning a DEX matters
Guaranteed route: Every job fee is swapped internally on Saturn Swap—no dependence on third-party AMMs.
Back-to-back order filling: The dispatcher posts a market-order into our own book, instantly buying $AGENT at best bid and crediting the StreamDistributor.
Built-in market making: Flux Point’s MM bot keeps ±3 % depth, converting friction into real-time liquidity instead of external sell pressure.
Oracle-grade price feed: The DEX’s VWAP feeds the InferencePool script, so buy-back maths and Charli3 price oracle stay in tight sync.
2.2 Full flow
Inference Pools (bonded by $AGENT) lock GPUs/NPUs.
User submits job → pays ADA (or pays $AGENT if holding a subscription NFT).
JobEscrow sends 70 % ADA to the operator; 30 % split: 28 % $AGENT buy-back, 1 % DAO, 1 % FeeTank.
Bought tokens accrue in the StreamDistributor; once per epoch they roll to stakers.
Fraud proof = slash; bond burns via the validator.
$AGENT Subscribers** pre-pay 10 k $AGENT per month and pay only 75 % of the standard Inference job fees. **Smart contract to go live on mainnet early June 2025.
3. Architecture
On‑chain (Plutus V3)
InferencePool validator
Holds bonded $AGENT, manages slash
JobEscrow script
Escrows fees, enforces proof & fee‑split
StreamDistributor
Drips bought‑back $AGENT to stakers
FeeTank (Aquarium)
Provides ADA gas using tanked ADA
Off‑chain
Dispatcher
Queues jobs, builds tx metadata
Verifier/Oracle
Result Hash, AI API receipt, Charli3 price feed
Epoch‑Roll Aggregator
Batches rewards; writes 1 payout/epoch
DEX
Saturn Order‑Book
Route ADA → $AGENT swaps Emits VWAP oracle datum
Batches JobEscrow updates
4. Tokenomics
Max supply: 1 B $AGENT (fixed)
Circulating: 999.99 M (current)
Utility sinks:
Bonding – scaling bonding of $AGENT based on hardware
Fee buy‑backs – 28 % buy-backs + 1 % DAO treasury + 1 % FeeTank top-up
Governance — All decisions and funds flow through our existing Clarity.vote / Agora DAO. Voting power equals bonded $AGENT, captured once per epoch by an open-source snapshot tool. A 1 % trickle of each fee slice replenishes the DAO treasury; no second DAO or token required.
Subscription discount – Jobs linked to an active 10 k $AGENT subscription NFT are billed at 75% of the standard ADA fee (25 % discount.
Emission: none. All new flow is secondary‑market buy‑back.
Fee Split
70 %
Pool operator
Energy / rental / API
28 %
Buy‑back contract → stakers
Yield; partial burn
1 %
DAO Treasury (ADA)
Diversified war-chest; audits, grants
1%
FeeTank (ADA)
Keeps ADA-less UX solvent
5. Treasury Note
The project already maintains an on‑chain treasury on Clarity.vote (Agora UTxO) for discretionary spending (marketing, audits, grants).
At this stage the Proof‑of‑Inference protocol does not auto‑route any portion of job fees to that treasury; design emphasis is on bonding $AGENT for compute security and buy‑back streams to stakers.
Should the community wish to direct surplus rewards into the treasury, a proposal can be raised through Clarity at any time.
6. Roadmap Roadmap Roadmap
Preview‑net
Q3 ’25
Inference Pools + edge/GPU nodes; ADA‑tank; UI merge in Saturn
Main‑net v1
Q4 ’25
Epoch roll‑up rewards, slashing live
Strategy Marketplace
Q1 ’26
Plug‑and‑play quant strategies mintable as NFTs
7. Risk & Mitigation
Tank drain – daily cap + oracle peg adjustment
Oracle collusion – multi‑sig verifier, open‑source client
Liquidity crunch – FP treasury MM bot maintains ±3 % depth on Saturn
Regulation – token used for compute & voting, no passive ROI promises
Additional discount‑specific risks are summarised in section 4.a.
8. FAQ
Q1 · Do I pay per inference if I already have a 10 000 $AGENT subscription NFT? No. The subscription NFT pre‑deposits your monthly 10 000 $AGENT into a dedicated Pre‑Paid Allowance UTxO. When you submit a job the dispatcher checks that balance first:
If allowance ≥ job fee, the JobEscrow tags the tx as “pre‑paid.”
The fee‑split still happens (28 % buy-back, 70 % operator, 1 % DAO, 1 % FeeTank), but the tokens come from your allowance, not your wallet.
Near‑zero balance triggers an on‑chain notice so you can top‑up.
Subscription NFT holders draw against their pre‑paid balance and enjoy a 25 % discount on any inference task.
Bottom line: one monthly top‑up covers unlimited jobs until the allowance is exhausted or discounted overage applies.
Q2 · Do I need to keep ADA in my wallet for gas? No. We use FluidTokens’ Aquarium FeeTank so every transaction can pay fees in $AGENT. Your wallet signs a single‑asset tx; the tank supplies the lovelace.
Q3 · What hardware qualifies me to run a pool? Three tiers with hardware examples:
Edge – Lite SBC (Pi 5) or Pro SBC (RK3588 / Jetson-Orin-Nano) (bond 5 000 $AGENT).
GPU – Consumer RTX 4090 with at least 24 GB VRAM (bond 80 000 $AGENT).
Datacenter – A100/H100 (bond 800 000 $AGENT).
Q4 · How are rewards delivered? Rewards are aggregated for the whole epoch (≈ 5 days). At epoch‑end the StreamDistributor writes one consolidated payout UTxO per staker containing all the auto‑bought $AGENT accrued during that period. Nothing to claim manually—just one gas‑efficient deposit every epoch.
Q5 · Can my bonded tokens be slashed? Yes—if your node submits incorrect results, misses the deadline, or fails the on‑chain verification check, the bond is burned.
Q6 · When can I access my rewards or un‑bond?
Reward‑vesting: Your rewards start accruing from block 1, but they remain locked for the first 6 months. After day 180 you can withdraw any accumulated balance at will (no cliff thereafter).
Un‑bond cool‑down: To exit entirely, submit an Un‑bond transaction—your bonded $AGENT stays locked for 2 epochs (~10 days) before it returns to your wallet. This protects the network from “bond‑and‑run” exploits.*
Why both? The 6‑month vesting strengthens long‑term alignment, while the 2‑epoch cool‑down keeps day‑to‑day operator churn under control.
Q8 · What happens if the FeeTank runs out of ADA? If the tank were ever to empty, the dispatcher would temporarily fall back to normal ADA fees. The tank is automatically topped up by 1 % of every job fee, so manual replenishment is rarely needed.
Q9 · How much does a typical inference cost? All numbers are ball‑park estimates—actual fees fluctuate with model size, demand, and on‑chain congestion.
Task
Tier
Fee (₳)
Approx. 28 % Buy-back (₳)
≈ $AGENT
Sub NFT fee (-25 %)
Chat 1 k tok
GPU
0.12
0.036
6
0.09 ₳
Code audit
GPU
0.40
0.12
19
0.30 ₳
Image 512²
DC
0.25
0.075
12
0.19 ₳
30 s TTS
Edge-Pro
0.05
0.015
2
0.038 ₳
28 % of each ADA fee is auto‑converted to $AGENT; table shows that portion at today’s price (≈ 0.00619 USD). Subscription holders pay ~25 % less per row.
Subscription NFT holders draw against their pre‑paid balance instead of paying these per‑call fees.
Q10 · How do I set up an Inference Pool? (Five‑minute checklist. Subject to change.)
Hardware – Plug in a qualifying device:
Edge Pi 5/RK3588/Jetson-Orin-Nano OR RTX 4090/other GPU rig OR rented A100/H100 instance.
Wallet – Fund it with the bond amount (5 k / 80 k / 800 k $AGENT) and a dust ADA UTxO for the first bond tx.
Docker – Install Docker ≥ 24.0; pull the pool image:
example
Bond – The container auto‑generates a
bond.cmd; copy‑paste it into your terminal to post the on‑chain BondUTxO.Open port 3000 so the dispatcher can stream jobs to your node (edge devices NAT‑punch via WebSocket tunnel).
Dashboard – Visit
localhost:8080; you’ll see “Bond ✓ | Jobs 0 | Rewards 0 ₳”. Jobs start within the next epoch.
That’s it—no manual cardano‑cli, no YAML. The container updates itself and warns you if bond health drops below 100 %.
Q11 · Can I run Iagon Storage/Compute on the same machine as my Inference Pool? Yes, provided the host still meets each service’s benchmarks. Iagon’s software does not prevent you from running both daemons side‑by‑side, but it also does not relax SLA targets.
Edge - not recommended.
Multi‑GPU rigs – start Iagon Compute with
--gpus="device=1"(or a MIG slice) so the primary GPU remains dedicated to inference workloads.Datacentre A100/H100 hosts – isolate Iagon Compute in its own CUDA context or MIG instance; NVIDIA’s MPS scheduler will arbitrate GPU time.
Always watch the local dashboard: if benchmarks dip below spec your node’s reputation—and rewards—will fall, and chronic misses can still be slashed.
9. Compliance Note
This document is informational and does not constitute an offer to sell tokens. Any forward‑looking statements and information in this document are subject to change at any time at the discretion of Flux Point Studios, Inc.
10. Contact
Web: fluxpointstudios.com
X/Twitter: @fluxpointstudio
Discord: discord.gg/MfYUMnfrJM
Last updated