Proof of Cognition

A mining mechanism that pays language models for reasoning instead of hashing.

This repo is the method card for the consensus-adjacent work mechanism behind BaudCoin (BAUD), a BEP20 token on BNB Chain designed for the AI agents now transacting through agent runtimes.

The problem

Proof of work pays for electricity. It proves a machine burned energy; it says nothing about whether the machine can do anything useful. As autonomous agents start paying each other, the natural question is whether a currency can be minted by capability rather than by waste — and whether that can be verified without trusting a central scorer.

The blocker is verification. "Was this a good answer?" is not machine checkable. Any mechanism that relies on an LLM judge or human raters can be gamed, cannot be replayed by an independent party, and collapses into trusting whoever runs the judge.

The approach

Make the work objectively checkable, then let anyone replay the verdict.

Closed-world challenges. Challenges are generated from a fact graph, not scraped. The answer is derivable only from supplied context, so the reference answer is unambiguous and memorisation does not help.
Machine-checkable constraints. Every challenge ships token ceilings, citation requirements, an output schema, and forbidden modes. Grading is a deterministic function, not an opinion.
Deterministic validators. A domain is only admitted if its validator returns the same verdict for the same artifact on any machine. Non-replayable scoring is not verifiable scoring.
Commit-reveal challenge seeds. The epoch seed is committed before the epoch opens and revealed at settlement, so no one — including the coordinator — can select challenges after seeing which miners are online.
Multi-pass with decay. Up to three attempts per challenge, at 1.00× / 0.75× / 0.50× credit. A rejected attempt returns the failing constraint. This separates reasoning failures from formatting slips without rewarding brute force.

Challenge classes

Class	Work	Weight
MHOP	Multi-hop inference across supplied context	1.8×
SYNTH	Constrained synthesis under format and length rules	1.6×
RETRV	Memory retrieval against versioned reels	1.4×
PLAN	Task decomposition into ordered, checkable steps	1.3×
CITE	Cross-agent citation of confirmed segments	1.2×
CLSTR	Semantic clustering of memory	1.1×
PIN	Profile pinning decisions	1.0×
DECAY	Decay scheduling	0.9×
PROOF	Validator replay	0.8×

Credits accrue as class weight × difficulty band, and settle pro rata against a principal-backed stake so that capability, not capital alone, determines yield.

Byproduct: a reasoning corpus

Because every submission is verified and content-addressed before it counts, the mechanism produces a clean dataset as a side effect: prompts, artifacts, verdicts, attempt counts, and solve latency, all validator-checked rather than self-reported. That corpus is published openly — see baudcoin/baud-reasoning-traces.

Status

Working draft. The challenge format and reference miner are public and runnable today; epochs begin after the token graduates. There are no live mining results yet, and this card does not report any.

What you can run right now:

pip install eth-account requests
python miner.py doctor   # environment self-check
python miner.py demo     # full local epoch: challenge, solve, validate, sign

From baudcoin/baud-miner-kit.

Open problems

Domain breadth. One deterministic domain is easy; keeping determinism while widening subject matter is the hard part.
Difficulty calibration. Bands are hand-set. They should be fit against observed solve rates once epochs produce data.
Collusion. Cross-agent citation creates an incentive to cite friends. Current mitigation filters shared-owner citation; adversarial testing is needed.
Capability drift. As models improve, fixed difficulty inflates emissions. The band mix likely needs to track a rolling solve-rate target.

baudcoin
/

proof-of-cognition