Retirement Lab evidence

A near-median task that passed directly

A complete spec-task instruction and artifact trail for a 239-line product change that passed implementation, tests, lint, E2E, and review without intervention.

PR #846 split taxable-account withdrawals into principal sold and investment yield consumed while preserving the existing API and export fields. It crossed the simulation engine, API schema, CSV exports, frontend types, product UI, and tests.

This is the smaller spec task path. At 239 changed lines, it is close to the overall task median of 242.

Run record

FieldValue
ImplementationClaude
Independent reviewSeparate Codex review
Size239 changed lines across 12 files
Human interventionNone recorded
VerificationTests, lint, and E2E passed on the first run
ReviewApproved with no findings
Follow-up repairNone known

The task explicitly called out backward compatibility, an additive invariant, multiple construction sites, scalar/vector boundaries, and neutral product wording. The focused diff shows the corresponding engine invariant, UI reconciliation, and regression coverage.

Source material

What it demonstrates

The task path can still describe cross-layer work precisely; “task” means a smaller work unit, not a one-line prompt. At 239 lines, this change is less than a third of the 878-line median for full specs, yet it still carries explicit danger zones and verification requirements.