Skip to main content
The protocol’s rule is the model proposes; the protocol governs what is stored. In the SDK that is a type: a Candidate has no block_id and no typed payload, so there is no method on the interface that could write to a Merkle DAG.
Five steps, deliberately not one. Steps 2 and 4 are the two places a model can be corrected rather than left to guess: it sees the exact required shape before writing, and the per-candidate verdict before committing.

​
The four rules a proposal fails on

  1. evidence is never empty, and it cites the task’s source. A derived block with no evidence has no root to audit against. There is no case where a proposer legitimately cannot cite: the source block is always there.
  2. No numbers in a payload. These documents get hashed, and a float does not hash reproducibly. confidence is a decimal string: "0.85", never 0.85. This is the most common rejection by a wide margin.
  3. locator says where in the source it came from. A citation pointing at a whole document is barely one.
  4. Do not restate the document. An extraction that is the document again has added nothing and made ranking worse.
Exit 7 means the proposal was wrong; retrying it unchanged fails identically. A duplicate rejection is not a defect — it means the brain already holds that block, and it does not block a commit, which is what makes the repair-one-and-resubmit loop work.

​
Verdicts

validated earns a commit. rejected names a code to repair. contradicted conflicts with knowledge already held — surface both block ids to the user rather than silently picking a side. pending_review means the protocol cannot decide alone: stop and ask a person, because there is nothing wrong to repair.

​
The shortcut

The default proposer, structure, uses no model: it reads Markdown headings and proposes one semantic block per section, extractively. Every statement it emits is text that was in the source, so it cannot invent — which makes it both a real answer for structured documents and the right thing to build the rest of the pipeline against. It also tracks code fences, because # do the thing inside a fenced block looks exactly like an H1, and a proposer that reads it as one produces a semantic block asserting a shell comment. --dry-run first, always, against a new kind of document.

​
Re-deriving

For “a better model should revisit this”. It records the supersession rather than leaving two competing interpretations of one piece of evidence installed, and the old block stays auditable.