midnight-fact-check:fact-check-classification

Domain Classification

The domain classifier operates on one domain at a time. To classify claims, read the claims file and tag each claim that belongs to the assigned domain.

The Four Domains

compact

Claims about the Compact smart contract language:

• Syntax, grammar, keywords, operators
• Type system (Uint, Field, Boolean, Bytes, tuples, structs, enums, generics)
• Standard library functions (persistentHash, transientHash, Counter, Map, Set, etc.)
• Control flow (for loops, if/else, assert)
• Disclosure and privacy (disclose(), ledger visibility)
• Module system (imports, exports, pragma)
• Compiler behavior (errors, warnings, accepted/rejected syntax)

Examples:

• "Compact tuples are 0-indexed" → compact
• "The for loop uses lower..upper syntax (inclusive..exclusive)" → compact
• "assert(condition, message) is the only error-handling mechanism" → compact

sdk

Claims about the Midnight SDK, TypeScript APIs, and DApp development:

• Package exports and imports (@midnight-ntwrk/*)
• API signatures, return types, error types
• Provider configuration, network connectivity
• DApp connector patterns
• Transaction lifecycle
• Deployment and contract interaction

Examples:

• "deployContract returns DeployedContract" → sdk
• "CallTxFailedError extends TxFailedError" → sdk
• "The indexer GraphQL endpoint is at /api/v1/graphql" → sdk

zkir

Claims about Zero-Knowledge Intermediate Representation:

• Opcode semantics (add, mul, neg, assert, constrain_bits, etc.)
• Field arithmetic (wrapping, modular)
• Circuit structure (instruction counts, I/O shape)
• Proof pipeline (PLONK checker, transcript protocol)
• Compiled output format

Examples:

• "add wraps modulo the field prime r" → zkir
• "disclosure compiles to declare_pub_input" → zkir
• "A pure circuit has no private_input instructions" → zkir

witness

Claims about TypeScript witness implementations:

• Witness function signatures and return types
• WitnessContext usage
• Private state handling
• Type mappings between Compact and TypeScript
• Witness/contract interface matching

Examples:

• "Witness functions must return [PrivateState, ReturnValue]" → witness
• "Boolean maps to boolean in TypeScript" → witness
• "WitnessContext is the first parameter" → witness

Classification Task

The dispatching agent provides the assigned domain as a parameter in the dispatch prompt.

1. Read the claims file (your copy)
2. For each claim, determine if it belongs to your domain
3. If it does, update the claim object by adding/merging your domain into the domains array and setting classification fields
4. If it does not belong to your domain, leave it unchanged
5. Write the updated file

When a Claim Belongs to Your Domain

Add or merge these fields:

{
  "domains": ["your-domain"],
  "classification": {
    "primary_domain": "your-domain",
    "confidence": "high",
    "notes": "Brief reason this belongs to your domain"
  }
}

• If domains already exists (from another classifier), append your domain to the array.
• If classification already exists, only overwrite primary_domain if the new confidence is higher. When both are "high" or both "medium", keep the existing primary_domain and append the new domain to domains.
• confidence values: "high" (clearly in this domain), "medium" (partially in this domain, might be cross-domain).

Cross-Domain Claims

Some claims span multiple domains. If a claim is partially in your domain:

• Add your domain to domains
• Set confidence to "medium"
• Add a note explaining the cross-domain nature

Example: "Compact Boolean maps to TypeScript boolean" spans both compact and witness. Both classifiers should tag it.

Boundary Cases

• If a claim is about compiler behavior (what the compiler does), it's compact
• If a claim is about compiled output (what ZKIR looks like), it's zkir
• If a claim is about SDK types that mirror Compact types, it's sdk (not compact)
• If a claim is about witness types that mirror Compact types, it's witness (not compact)

Output

Write the updated claims file to your copy path. Return a summary:

Tagged N claims as [domain]. Skipped M claims (not in my domain).

Note: cross-domain tag counts are determined after all classifiers merge results, not at individual classification time.