reliability

Reliability score

Run this phase after Breaker has accepted. Rate the final change set on a 1–10 scale using the rubric below.

Rubric (equal weight by default; adjust if the task is trivial)

Dimension	What "good" looks like
Tests	Automated coverage for new logic; regression for fixed bugs; realistic integration where needed.
Observability	Logs/metrics/traces let operators detect failure and narrow blame.
Error handling	User-visible and internal errors are handled; no silent swallow of critical failures.
Operational fit	Config, migrations, rollbacks, and load characteristics are considered.
Security	AuthZ, secrets, and input validation appropriate to exposure.

Scoring guide

• 9–10 — Strong across dimensions; minor nits only.
• 7–8 — Shippable with known follow-ups documented.
• 5–6 — Significant gaps; risky without mitigation.
• 1–4 — Not production-ready.

Score anchoring examples

Score	What the change looks like
4/10	Feature works but has no tests, no error handling beyond happy path, no logging, and HEALTH_CHECK smoke steps were all skipped during verification.
6/10	Tests cover the happy path, errors are caught but logged generically, rollback plan exists but is manual. Some smoke steps passed, some skipped.
8/10	Tests cover happy path + key edge cases, structured logging for new failure modes, rollback is a config flag or revert. All runnable smoke steps passed.
10/10	Full test coverage including failure injection, metrics/alerts for new paths, automated rollback, migration tested both ways. Rare — reserve for genuinely thorough work.

Score inflation traps

Trap	Correction
Defaulting to 7-8 because the code "looks clean"	Code quality is only one dimension; check tests, observability, ops, and security independently
Scoring 8+ when verification steps were all skipped	Zero verified steps caps the score at 6 regardless of code quality
Giving 9+ for trivial changes because "nothing can go wrong"	Trivial changes get trivially adjusted rubric weights, not inflated scores
Raising the score because the Builder "tried hard" across rejection cycles	Effort is not reliability; score the final artifact, not the process

Output format

## RELIABILITY: <score>/10

**Justification:** <1–2 sentences>

**Top gaps:**
- <gap, or "None — strong across dimensions">

Logging (.prodscope/logs/)

After scoring, append one log line per significant event:

ISO8601<TAB>agent<TAB>action<TAB>detail

Examples:

2026-04-11T12:00:00Z	breaker	REJECT	missing index on users.email lookup
2026-04-11T12:05:00Z	builder	IMPROVE	added migration idx_users_email
2026-04-11T12:06:00Z	prodscope	ACCEPT	reliability 8/10 — shippable with follow-ups

action values: REJECT, ACCEPT, IMPROVE, VERIFY.