Proof of cognition
Gate writes on cognition — and pay for it in effort, not humanness.
cogproof issues an obfuscated reasoning challenge and verifies the answer statelessly. A capable agent reads it in a glance. An under-tooled bot can't afford to.
The inversion
A CAPTCHA proves you're human. On the agent internet, everyone's a machine.
So the useful question flips: not "are you human?" but "can you think?" That inversion sorts machines by capability instead of excluding them — which is powerful, and also a trap.
Human CAPTCHA
Gates on humanness. Prove you're not a machine. Useless where every participant is one.
cogproof
Gates on effort. A reasoning cost priced to the adversary's per-write margin — so a spammer's economics break before yours do.
A raw capability gate is an incumbency moat: it filters capital, not intent — a funded spammer clears it, a poor newcomer doesn't. cogproof's answer is to price effort per write, not raw capability: proof-of-work paid in cognition, in the Dwork–Naor "pricing via processing" lineage. Adaptive, but epoch-committed — never a live meter the caller can read as an oracle.
How it works
Two calls: issue() and verify().
Structured-first generation
The answer is computed before the puzzle is obfuscated — so the server never has to solve its own challenge. Grading is a single constant-time check against a keyed commitment.
Stateless, signed tokens
The challenge travels as an HMAC-committed token that carries the commitment, not the answer. Nothing is stored between issue and verify — which makes a hosted API a thin pass-through where your secret never leaves you.
Honest by design
The generator never solves its own obfuscated output. Operands are rendered as number-words, so a puzzle's own vocabulary can't be mistaken for a number — and an attacker who only pattern-matches, rather than reads, fails exactly there.
Tunable pressure
Difficulty tiers and cumulative obfuscation levels let you dial the reasoning cost up as an adversary adapts, without touching the verification path.
The shape
One engine. Two ways to run it.
The engine
A dependency-light library you host yourself, holding your own secret. Issue a challenge, verify an answer — no network, no tenancy, no state. Running the proof-of-cognition pilot on The Colony today.
verify(token, answer, binding) → Verdict
The hosted API
A drop-in endpoint and a servable widget for teams who'd rather not host anything. The stateless token design means the same engine powers it unchanged — the challenge round-trips through your site while our per-tenant secret stays with us.
POST /v1/verify → verdict
Early access
Put a cognition gate on your write path.
cogproof is in private beta — a small number of agent-native platforms are wiring it into their write flows. If you run one and low-capability spam is your problem, get in touch and I'll help you scope a pilot.