v3.0.0 — the forge rebuilt around proof — the Gauge, the Law, and the Reserved Seven

Kiln v3.0.0 — the forge, rebuilt around proof

v3.0 is a ground-up evolution of the native-workflow engine: the pipeline now measures the work before it runs, locks its acceptance criteria into executable law before a line of code exists, and proves the result by operating the finished product — in a real browser when the product has a face.

Highlights

• The Gauge. Every run opens by scoring the work across eight complexity dimensions; the pipeline's rigor — planning depth, validation rounds, slice budget, tribunal, browser tier — is computed from that profile. Small work runs lean; real products get the full machinery.
• The Law (test-lock). Acceptance criteria compile into executable checks, hash-pinned before the build starts. Every slice must flip its checks red→green; a touched lock auto-rejects with no appeal. The build meets the oracle — never the other way around.
• Vision from the ledger. Brainstorming writes an append-only session ledger, and a fresh-context compiler produces VISION.md from the ledger alone — every idea in the vision traces to something you actually said, structurally.
• Bounded browser verification. UI acceptance criteria are exercised live via scripted one-shot Playwright probes — screenshots per viewport, console/network capture — under kill-tokens and a hard time lease. Verified, then gone; leaked browsers are hunted by design.
• The goal-backward audit. After every milestone and again at validation, an auditor walks backward from your goals through the product's entry points — catching the class where every test passes and the app is dead at its front door.
• The Reserved Seven return. The seven personas v2 shelved are back, each with a real seat in the engine: the Assessor runs the Gauge, the Lawgiver owns the Law, the Scribe keeps the run ledger, the Observer keeps the browser watch, the Negotiator became the reconcile arithmetic, the systematizer compiles the vision, and the fallback planner owns the capability ladder. The roster reads 34, and every name is a seat that runs.
• Sanctioned theater. A forge theme, an output style, a notification ping, session titles for kiln projects, and opt-in spinner verbs — all consent-framed, none touching what you own.
• A leaner conductor. The pipeline skill travels light: per-stage contracts moved to references, plugin-root resolution to one executed script, and the whole surface is guarded by deterministic skill evals.

Requirements & breaking changes

• Claude Code ≥ 2.1.198 with Dynamic Workflows enabled (/kiln-doctor verifies this and every capability tier).
• Codex CLI remains optional — GPT-5.5 joins for planning, logic builds, and cross-family review when present; the Claude-only path is complete on its own.
• In-flight v2 runs: finish them on v2.0.1 or start fresh — v3 does not migrate v2 run state.

Honesty, as always

Validation verdicts are computed from evidence — Law exit codes, test suites, browser probes — never self-graded. When a capability is absent, the verdict says so (PARTIAL_PASS_STATIC_ONLY) instead of rounding up to green.