Skip to content
codeArbitercodeArbiter

SMARTS

When codeArbiter must weigh options, and especially when /ca:sprint decides “as the user” during an autonomous run, it doesn’t pick on vibes. It runs a structured, multi-lens scoring rubric called SMARTS and records the verdict: the options weighed, the lens scores, the chosen option, and a confidence flag. Low-confidence calls are exactly what the user reviews afterward. Nothing hides behind autonomy. Every auto-decision lands in an append-only sprint log.

The Six Lenses

Section titled “The Six Lenses”

SMARTS is an acronym. Every option is scored evenhandedly against the same six lenses, and the analysis travels with the recommendation:

  • Scalable. Supports growth in users, data, and throughput without an architectural rewrite. The trap is over-engineering for scale that never arrives, or under-engineering for scale that’s already on the roadmap.
  • Maintainable. Can be understood, modified, and extended later (including by agents) without prohibitive effort. Standard patterns over bespoke abstractions; mind the refactoring blast radius.
  • Available. Reachable and functional when needed, including under partial failure. Watch for single points of failure, bundled-dependency failure, and recovery time.
  • Reliable. Correct, predictable, durable outcomes: ACID where it matters (the decision log, audit events), idempotency, state consistency, and recovery without corruption.
  • Testable. Validated by deterministic, fast tests that cover real failure modes. “Tests later” is a weak verdict.
  • Securable. Enables the project’s security posture (per security-controls.md) without a retrofit: authentication, authorization, audit, secret management, and a default-deny stance.

Verdicts and Strength

Section titled “Verdicts and Strength”

Each lens gets one verdict word: Strong, Adequate, Weak, or Indifferent (the lens does not differentiate the options at the current scale). A short, evidence-specific justification follows. Vague appeals like “industry standard” or “best practice” are not evidence; a cell must cite a concrete property of the option, a specific project constraint, or a named failure mode.

The recommendation itself carries one strength label: strong when multiple dominant lenses align cleanly, moderate when they align with caveats or a single lens dominates, and tied when no preferred option emerges. A tie is a legitimate result, surfaced as a genuine call for the user rather than a forced pick.

SMARTS deliberately does not score cost, time-to-market, team-skill fit, or vendor lock-in. When those matter, they are surfaced as non-SMARTS considerations alongside the table; they supplement the analysis, never replace it.

Where it runs

Section titled “Where it runs”

The canonical rubric ships at plugins/ca/includes/smarts/core.md. The grader agent produces one SMARTS analysis per artifact-position and scaffold-evidence pair, dispatched by the decision-variance skill through /ca:reconcile.

Auto-decisions made during an autonomous /ca:sprint run append to .codearbiter/sprint-log.md. Arbitrated decisions land in decision-log.md, each carrying a SHA-256 hash of its own section, which lets a later reconcile pass detect when a decision has gone stale against the code it once described.