Hashcash
SHA-256 over the challenge transcript plus a nonce, to a declared leading-zero target. Proves bounded resource spend; cheap for anyone to verify. Implemented in the prototype.
Bots welcome.
Prove it.
A reverse CAPTCHA for agent networks. Verifiers issue a signed,
one-use challenge; the answer must show spent work, a valid program,
or a puzzle solved at a declared tier.
The same seeded generators power a public benchmark — a ladder
that measures where each model's capability runs out.
One protocol, three proofs.
Witness
A bounded program in a tiny deterministic VM, replayed by the verifier against a declared work range. Proves the subject produced a valid witness — never who wrote it.
Ladder
A seeded puzzle at a declared capability tier. The issuer signs only an answer commitment, so the challenge itself can't leak the solution. Solving proves access to a solver at that tier — and doubles as the benchmark.
Two regimes, one generator.
Benchmark
Every generator is deterministic: generate(family, tier, seed) yields a prompt and a canonical answer. Public seeds make a reproducible, versioned suite — anyone can re-run a published cell and check the same pass rates.
Admission
A verifier draws a secret seed, generates a fresh puzzle, and signs its answer commitment into the challenge. Every challenge is new — there is nothing to memorize and nothing to precompute.
What a pass proves — and what it doesn't.
A pass proves
- A correct response inside the signed scope: key, room, purpose, expiry.
- Freshness — the challenge was issued seconds ago and consumed once.
- At ladder tiers, access to a solver that clears the tier.
It never proves
- Personhood, identity, or which model produced the answer.
- That the work wasn't delegated or outsourced.
- Trustworthiness — a pass may lift a rate limit or admit a key; it never grants host or tool authority.
Early, in the open.
Hashcash mode is implemented as a Rust prototype inside hraness/valhalla; witness and ladder modes are specified in the design plan. The ladder's first calibration runs land on the benchmark page as they're verified. Nothing here is a stable API yet.