v0.2.0-alpha.1 — Evidence

bluei v0.2.0-alpha.1 — Evidence

First release of the 0.2.0 line — the deterministic flywheel era. This release makes the Deterministic Flywheel visible. It is observability only — no fix-behavior changes, no auto-promotion. The thesis: prove the flywheel's value with measurement before alpha.2 introduces the (riskier) learning/promotion machinery.

⚠️ Pre-release software. Do not use in production. Use for evaluation, testing, and contributions only.

Flywheel Ledger

The deterministic-vs-LLM resolution picture, surfaced across the clean summary, report (markdown + HTML), and the dashboard:

• Per-stage resolution counts — which cascade stage won (linter / recipe / pattern-replay / composite-pattern / ast) vs. LLM fallback vs. exhausted.
• Honest replay outcomes — three-way HIT / MISS / FAILURE (replacing the mislabeled skip_count, which silently held validation failures).
• Standalone-Pattern-replay $ savings — observed substitution cost avoided (grounded, not counterfactual; built-in deterministic stages report throughput, not dollars).
• Rates as num/denom (%) over findings_attempted — the denominator is always shown beside the rate.
• A dedicated cascade_resolutions.jsonl sink + an in-memory ledger_records accumulator feed a flywheel_ledger block in status.json.

Pattern Court Lite

Learned Patterns become inspectable and their replay evidence honest (observe only; demotion/promotion deferred to alpha.2):

• ReplayOutcome enum — HIT/MISS/FAILURE with correct storage mapping and a new last_failed_at stamp.
• Evidence envelope — last_failed_at + excluded_paths (enforced at the lookup layer so scoped-out Patterns never become candidates).
• Explainability — enriched patterns show (3-way counts, last failure, excluded scope) + a MISS column in patterns list.
• Producer — new patterns exclude / patterns unexclude subcommands make excluded_paths real (not inert scaffolding).

Decisions

Five ADRs record the load-bearing choices: dedicated ledger sink (0001), findings_attempted denominator (0002), hybrid savings methodology (0003), three-way replay taxonomy with exactly-once recording (0004), and write-time per-cycle aggregation via the in-memory accumulator (0005).

Carried forward to alpha.2 ("Safe Learning")

Shadow / dry replay, demotion & promotion triggers, the capture-infra envelope fields (imports_touched, validation_commands_passed, negative-example store, rule_family), cascade-internal savings instrumentation, the bluei savings CLI, and the run_id correlation key (add before any cumulative/cross-cycle feature). Tracked in the ROADMAP "Carried forward from v0.2.0-alpha.1" table.

Quality

• 6,121 tests passing, 9 skipped, 0 failures.
• Post-release code review: 10/10 invariant checks pass, 7 non-blocking findings addressed.
• Scope verified additive against the v0.1.0-beta.1 baseline.