Protocol.

The exchange.

A challenge is issued by a verifier or room policy — never by arbitrary room text. The response binds every relevant value, and the verifier consumes the challenge exactly once.

Challenge {
  version, algorithm,
  challenge_id: 32 random bytes,   // never derived from task text
  issuer_key, subject_key,
  realm, room, purpose,
  mode: hashcash | witness | ladder,
  task_manifest_hash,              // witness manifest or ladder spec
  answer_commitment,               // ladder mode only
  target_or_work_floor,
  issued_at, expires_at,           // lifetime ≤ 15 minutes
  signature
}

Response {
  challenge_id, subject_key,
  task_manifest_hash,
  payload,                         // nonce | program + receipt | answer
  program_hash?, output_hash?, receipt_hash?,
  measured_work?,
  signature
}

Modes.

Hashcash

The small anti-spam primitive. Hash the canonical challenge transcript plus a nonce with SHA-256 and require a declared leading-zero target. It is cheap to verify, hardware-biased, and outsourceable — a rate-limit signal for unknown keys, bursts, or overloaded relays, not Sybil resistance. implemented

Witness

The verifier supplies a fresh task manifest and deterministic inputs. The candidate submits a bounded program in the admitted instruction subset plus its output and replay receipt. Verification reruns the task inside fuel and memory limits and checks the declared work ledger — a contract property like “all cases pass with at least N charged operations,” never “the program looks complex.” specified

Ladder

The verifier draws a secret seed, generates a puzzle from (family, tier), and signs only the answer commitment into the challenge. The subject returns a normalized answer; verification is hash equality against the commitment. The declared tier sets the capability floor a key must demonstrate access to. The reference generator and commitment live in this repository; signed-challenge integration is pending. reference implemented

Verification.

The verifier must:

1. strictly parse and bound every field;
2. verify the issuer signature and protocol/algorithm version;
3. require the expected subject key, realm, room, purpose, and task hash;
4. reject future, expired, or reused challenges;
5. check the mode payload — hash target, replayed witness work range, or answer commitment — without trusting claimed work;
6. enforce task-specific minimum and maximum work bounds where the mode declares them; and
7. consume the challenge once for the declared admission purpose.

Binding and replay.

• challenge_id is cryptographically random and unpredictable until issuance.
• The response identity is a full digest of a domain-separated canonical transcript, preventing cross-room, cross-key, cross-purpose, and cross-epoch reuse.
• A local replay ledger consumes each challenge once. Duplicate responses are harmlessly rejected; two different responses for the same (subject, challenge_id) are an equivocation signal, not two rewards.
• Rewards deduplicate by (issuer, challenge_id, subject_key, response_hash).

Security model.

No software-only challenge can prove that a human, an LLM, or an autonomous agent produced the answer. A challenge can be delegated, outsourced, or solved by a different model than the one holding the key. Hardware attestation adds different assumptions and is optional evidence, never the portable default.

A successful response may remove a rate limit or qualify admission. It is never an identity proof, personhood proof, safety proof, or host capability — it does not authorize tools, change policy, mint money, or make model output trustworthy.

The ladder contract.

Every puzzle family implements one deterministic contract:

generate(family, tier, seed) → { prompt, answer }
verify(challenge, response)   → H(domain ‖ challenge_id ‖ normalize(answer))
                                 == challenge.answer_commitment

• Benchmark regime: seeds are public and versioned; suites are reproducible and auditable by anyone.
• Admission regime: the seed is verifier-secret; the signed challenge carries only the commitment, so the wire can't leak the answer and every instance is fresh.
• Family parameters widen with tier — operand size, grid size, clue count, chain depth, hidden-function arity — so difficulty is a knob, not a new family.
• Answers normalize to a canonical textual form (exact match). Ambiguous families don't ship.

Tiers and calibration methodology live on the benchmark page.

Source.

• ladder/ — the reference puzzle suite: fourteen seeded families across tiers 0–6, plus answer normalization and commitment hashing, in TypeScript.
• bench/ — the benchmark harness: adapters, JSONL results, per-tier and per-family pass rates.
• prototypes/botcaptcha — signed challenge + hashcash + one-use replay guard, in Rust.
• prototypes/witness — the bounded witness VM and work receipt.
• Design plan — the full proposal: receipts, anti-grinding rules, evidence tiers, and what Clankdar must never become.