3. Phase B - intent + code

About two minutes

Call with businessIntent and damlCode. Optionally add compilationResult from a client compile (daml build / damlc on your machine or in CI).

Outcomes

Outcome	Meaning
ready_to_compile	Alignment looks good enough. Stop the semantic loop and compile next.
fail	Not acceptable yet. Strong patterns attached when recall allows (fail-and-help).
abstain	Too uncertain. A human should decide.
pass	Only after a successful client compile and the remaining gate.

Hard rules (easy to remember)

• No compilationResult → pass is forbidden
• Compile failed (succeeded: false or non-zero exitCode) → always fail
• Every Phase B response includes alignmentScore
• There is no parallel valid bool. The outcome string is the verdict.

Compile hard signal

The client supplies compile status (succeeded and/or exitCode). The MCP never runs the compiler. On failure, optional errors, diagnostics, stderr, or stdout may be echoed in issues for humans and agents to read — the server does not parse compiler output into tailored fixes or change alignmentScore from diagnostics.

How the verdict is decided

The server retrieves corpus patterns relevant to the intent, then uses a heuristic alignment score on the submitted draft to choose revise vs compile vs pass. The cross-encoder selects and ranks patterns; it does not verify your code. When you attach compilationResult, the server reads only the hard signal you report; pass means these gates passed, not that business behavior was proved correct.

Today’s score mixes token coverage of intent vs code with strong corpus hits (FailBelow 0.40 / ReadyAt 0.70). It is good enough to drive a draft loop. It is not a formal proof that the economics are right. Treat misses as calibration feedback for the corpus and the scorer, not as “the tool failed forever.”

flowchart TD
  start[Phase B: intent + damlCode] --> recall[Honest corpus recall]
  recall --> align[alignmentScore]
  align --> compile{compilationResult?}
  compile -->|absent| noCompile{score vs thresholds}
  noCompile -->|below fail| fail1[fail]
  noCompile -->|ambiguous + weak recall| abstain1[abstain]
  noCompile -->|ready| ready[ready_to_compile]
  compile -->|failed hard signal| fail2[fail + patterns when recall allows]
  compile -->|succeeded| okCompile{score vs thresholds}
  okCompile -->|low / ambiguous| fail3[fail or abstain]
  okCompile -->|confident| pass[pass]

Frozen field rules: docs/internal/TOOL_CONTRACT.md (Phase B), private repo.

---

← Phase A · Agent loop →