Verification Before Completion

<EXTREMELY-IMPORTANT> NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE. Any claim of success requires executing the verification commands freshly, reading the full output and the exit code, and confirming the result supports the claim. </EXTREMELY-IMPORTANT>

Verification hierarchy

Always — when code changed

• sbt Test/compile — zero errors, zero warnings (-Werror makes them the same thing).
• sbt test — every suite passes, including the ts-node round-trip property (requires cd tests && yarn once).

Conditional

• sbt githubWorkflowCheck — when build.sbt changed (CI runs it; regenerate with sbt githubWorkflowGenerate).
• A consumer check — when generated output changed: regenerate in BondLink against this checkout (see CLAUDE.md → "Using a local checkout from BondLink") and diff the affected files.

A fix is unproven until you have seen the bug

For a change in generated output, the golden test must fail without the change (showing the old output) and pass with it. Run the suite before implementing, or lift the change out afterward, and record what the failure said.

Claims that need specific evidence

Claim	Required evidence	Not sufficient
Compiles clean	sbt Test/compile output with [success] and no warnings	"Previous run passed"
Output is correct	Golden test passing, having been seen failing first	"Looks correct in the code"
Output type-checks	ts-node round-trip property passing for a suite whose decode-encode.ts imports the output	A golden-string match alone
No regressions	Full sbt test	"Only changed one renderer"
Consumer unaffected / fixed	A regenerated consumer file diffed against expectation	Reasoning about what the consumer does