PoATU — Proof of AI Tokens Used

PoATU pays miners for actual AI inference work. Every TASK_RESPONSE tx must declare how many output tokens (and optional thinking tokens) were generated. A validator then picks the best response and a TASK_REWARD tx pays the winning miner proportionally.

Reward formula

Given a winning response with outputTokens (and thinkingTokens for reasoning models):

total_tokens = outputTokens + thinkingTokens
reward       = (total_tokens / TOKENS_PER_OMBRA) * TASK_FEE_SHARE

Where:

Constant	Value	Meaning
TOKENS_PER_OMBRA	1_000_000	1 OMBRA = 1M AI tokens at full price
TASK_FEE_SHARE	0.70 (post-fork V2)	70% of TASK_SUBMIT.fee goes to miner

Fee split

A TASK_SUBMIT.fee is distributed between miner, burn, and validator at apply time:

FEE_SPLIT_V2 = { miner: 70, burn: 25, validator: 5 }   // percent

This split activates at fork V2 (mainnet height 7500). Pre-fork: 100% to miner, no burn.

Pseudo-code: apply TASK_REWARD

function applyTaskReward(tx, state):
  task = state.tasks[tx.taskId]
  assert task.status == "validated"
  assert tx.from == task.validatorAddress

  for reward in tx.rewards:
    addr  = reward.address
    micro = BigInt(reward.amount)

    if reward.reason == "miner":
      state.accounts[addr].balance += micro
    elif reward.reason == "validator":
      state.accounts[addr].balance += micro
    elif reward.reason == "burn":
      state.totalBurned += micro       // no recipient

  state.tasks[tx.taskId].status = "rewarded"

Example

A user submits a chat task with fee = 100_000 micro-OMBRA (0.1 OMBRA). Best response is 5_000 output tokens.

• Miner share = 100_000 × 0.70 = 70_000 micro
• Validator share = 100_000 × 0.05 = 5_000 micro
• Burned = 100_000 × 0.25 = 25_000 micro

Total credited = 75_000. Total burned = 25_000. Tokens-per-OMBRA exchange rate at that fee = 5_000 / 0.07 ≈ 71_428 tokens/OMBRA.

Why per-token (not per-task)

Tasks vary wildly in cost:

• A simple yes/no: ~50 tokens
• A long essay: ~5_000 tokens
• An agentic workflow with reasoning: ~50_000 tokens

Flat per-task pricing would over-reward trivial tasks and under-reward complex ones. Per-token PoATU matches reward to real GPU/inference cost.

Anti-cheat

A miner could claim inflated token counts. Mitigations:

1. Validator picks best response — only one miner gets the full reward; others get zero. Token count is validated against the actual response length.
2. Validator stake — validators are chosen from attested block proposers (skin in the game).
3. Cross-check across responses — if multiple miners answered, token counts should cluster; outliers are flagged.

Next: PoTU →