Recorded
Manifest bytes or a commitment, inscription identifier and transaction context. Define the indexing rules and confirmation policy used by the verifier.
ARCHITECTURE / NAME EVERY DEPENDENCY
Separate what is recorded, what must remain available, what is computed and what is actually verified.
A proposed research flow. The website does not submit transactions or operate distributed workers.
The objective is to make an asset’s reconstruction requirements inspectable. References should identify exact content, and the acceptance test should be explicit. A mutable model name or unavailable input can break replay even while a chain record remains readable.
The following field groups are an illustrative specification checklist. They are not a registered protocol schema or a production API contract.
manifest_version
model_digest + adapter_digests
input_digests + workflow_digest
prompt + seed + parameters
runtime + precision + platform
output_encoding + expected_digest
rights_statement + signerManifest bytes or a commitment, inscription identifier and transaction context. Define the indexing rules and confirmation policy used by the verifier.
Model weights, input assets, runtime and workflow dependencies. Content addressing makes substitution detectable; a retrieval and retention strategy keeps replay possible.
Schema validity, dependency integrity and output equality are separate checks. Each needs its own failure state, evidence and scope.
An early route uses a satoshi identifier to locate an off-chain prompt, with access following a control check. The service must define identity, challenge signing, replay protection, key management and transfer behaviour. A public identifier must never be treated as a private key. Access policy also cannot revoke copies that a recipient has already made.
The alternative records the compressed recipe payload itself. This reduces reliance on a private lookup service for that content, while introducing payload size, public disclosure and dependency questions. A visible prompt can be copied. Its durable association with an inscription is a provenance feature, not automatic exclusivity over every resulting file.
Workers need bounded jobs, authenticated dependencies and a result contract. Verification must account for untrusted workers, execution cost and the difference between a matching digest and a claim about how the bytes were obtained. Independent replication can be part of an experiment; it does not by itself solve every adversarial-computation problem.
PLEB appears in the drafts as a proposed state and extensibility overlay. Its validation rules, proof system, governance and failure recovery need independent design and review. Bitcoin proof-of-work remains a separate consensus mechanism; useful AI jobs cannot simply be substituted for its work under the existing protocol.