# Decantr > Agent-neutral UI change control for codebases edited by AI. Decantr observes project-owned authority, prepares compact change context, verifies the resulting diff, and reports typed local evidence. ## Release Status - Current stable: Decantr 3.11.3. - Default entry point: `decantr verify` performs zero-write Changed-UI Assurance over the current Git change; no adoption is required. - Active product: changed-UI assurance plus authority-aware Observe, Prepare, Verify, and Report workflows. - 3.11.3 is released but is not value-proven against frontier models. - Product release readiness is separate from model-lift research qualification. Do not claim human finding precision/recall, adoption proof, or measured frontier-model improvement. ## Published 3.11.3 Entry Point 1. Run `decantr verify` in a Git worktree. 2. Decantr resolves staged, unstaged, deleted, renamed, untracked, commit-range, or unborn-branch changes. 3. It auto-selects exactly one changed UI app when provable; multi-app ambiguity returns `not_proven`. 4. It reports at most three consequential `AUTH001`, `AUTH010`, `COMP001`, `COMP010`, or `TOKEN010` findings with exact source and repair targets. 5. CLI, explicit CI v3, and MCP `decantr_verify` action `changes` use the same verifier-owned `change-assurance-report.v1` contract. Use `decantr verify --full` for the previous Project Health workflow. Use `scan`, optional `adopt`, and `task` when deeper project context is warranted. In monorepos, bare verify may select one app from changed files. Pass `--project ` when changes cross apps or selection is intentionally explicit. A successful scan means the scanner ran; inspect each authority axis and limitation before treating a target as ready. ## 3.10 Authority Model Routes are one evidence source, not the product ontology. The approved UI surfaces are routes, layouts, components, stories, overlays, flows, packages, and runtime states. Readiness axes are independent: selected-app authority, production-surface authority, topology completeness, implementation taskability, component inventory quality, styling authority, and runtime evidence. The 3.10 primary result is `ready`, `limited`, `blocked`, or `unsupported`. Numeric confidence and aggregate fit labels must not conceal an unresolved axis. Decantr 3.10.0 ships non-route UI-surface task context as a separate `ui-surface-task-context.v1` envelope. Static component or story evidence normally remains `limited` unless runtime reachability is proven. Development repair commit `98e92472` separates Next file-route declaration from statically resolvable middleware/proxy 4xx policy, preserves workspace CSS import order, excludes App Router server handlers from UI component inventory, and keeps blocked task read sets empty. The Culinary Platform replay is oracle-assisted regression evidence only; it is not a control/treatment run or frontier-model result. ## Product Boundary Active 3.11 investment: `@decantr/verifier`, Changed-UI Assurance, framework authority adapters, CLI `scan`/`task`/`verify`, the compatible eight-tool MCP server, local evidence, CI integration, and deterministic policy packs. Advanced, compatibility, or historical: Greenfield blueprints, themes, registry publishing/account commands, `@decantr/css`, `@decantr/vite-plugin`, broad corpus expansion, telemetry product work, Studio/showcase expansion, and hosted cross-repository intelligence. Decantr is local-first. Hosted source upload is retired. Browser evidence stays local. The Fly content API is an optional reference helper. In Brownfield, production source and runtime configuration remain first authority. Project tests and design systems are scoped supporting evidence; accepted Essence and local rules are project law beneath production source. Official Decantr content is advisory unless deliberately adopted into project law. ## MCP The public MCP identity and exactly eight tools remain stable: `decantr_project`, `decantr_contract`, `decantr_context`, `decantr_graph`, `decantr_registry`, `decantr_verify`, `decantr_repair`, `decantr_contract_write`. `decantr_registry` is a 3.x compatibility name backed by `@decantr/content`, not a public marketplace. Do not expect a ninth content tool in 3.x. ## Model-Lift Research Boundary The 3.10 experiment is 40 tasks x 2 models x 2 arms x 2 repetitions = 320 isolated runs over 28 pinned repositories. It includes one repository-authentic task per target plus 12 adversarial tasks. The corpus acquisition labels are 18 development and 10 qualification; independently, the task partition is 24 development and 16 sealed qualification. Repository identities were observed at Day-0, so this is not a repository-blind holdout; qualification solutions/oracles remain sealed and a future independent external holdout is required for a general industry claim. The control receives repository-native instructions, tests/design evidence, and an approved policy card. The treatment receives the same information entitlement through Decantr; Decantr receives no extra human facts. Candidate package bytes are frozen before paid execution; an implementation change invalidates that candidate result. Requested models: OpenAI `gpt-5.6-sol` and Anthropic `claude-fable-5`. Model substitutions remain visible and are not pooled. The value claim requires, separately where declared: correct or honestly unsupported Day-0 authority, no non-production authority contamination, correct rank-one ready target, at least +5/100 paired rubric lift with 95% CI above zero for each model, at least 25% fewer governance violations, functional non-inferiority within 5 percentage points overall for each model, at least 26 decisive blinded-preference units out of 32 qualification task/model units, at least 60% treatment preference among decisive units, a two-sided 95% Wilson lower bound above 50%, median token/cost overhead <=15%, and P95 overhead <=25%. Framework functional estimates are exploratory unless separately powered. Pass permits only the bounded measured claim. Mixed results narrow supported claims. Failure blocks value-proof language, not the stable product release. Unsupported targets, missing evaluators, build failures, and model substitutions stay visible and in the denominator. ## Active Docs - [Homepage](https://decantr.ai/) - [Docs index](https://decantr.ai/README.md) - [Existing apps](https://decantr.ai/guides/existing-apps.md) - [Change Assurance](https://decantr.ai/reference/change-assurance.md) - [AI assistant setup](https://decantr.ai/guides/ai-assistant-setup.md) - [Workflow model](https://decantr.ai/reference/workflow-model.md) - [Command surface](https://decantr.ai/reference/command-surface.md) - [FAQ](https://decantr.ai/faq.md) - [Security and permissions](https://decantr.ai/reference/security-permissions.md) - [Package support matrix](https://decantr.ai/reference/package-support-matrix.md) - [3.11.3 SvelteKit task-authority patch](https://decantr.ai/releases/2026-08-08-decantr-3-11-3-sveltekit-task-authority.md) - [3.11.2 route-authority hardening](https://decantr.ai/releases/2026-08-08-decantr-3-11-2-route-authority-hardening.md) - [3.11.1 MCP metadata patch](https://decantr.ai/releases/2026-08-07-decantr-3-11-1-mcp-metadata-compatibility.md) - [3.11.0 Changed-UI Assurance release](https://decantr.ai/releases/2026-08-07-decantr-3-11-0-changed-ui-assurance.md) - [3.11 qualification evidence](https://decantr.ai/research/2026-08-07-decantr-3-11-change-assurance-trials.md) - [4.0 entry criteria](https://decantr.ai/reference/decantr-4-entry-criteria.md) - [3.10.0 authority-model release note](https://decantr.ai/releases/2026-08-07-decantr-3-10-0-authority-aware-ui-change-control.md) - [Frontier-model lift research program](https://decantr.ai/programs/2026-07-22-decantr-3-10-ui-change-control-proof.md) - [3.10 Culinary Platform clean-slate adoption](https://decantr.ai/benchmarks/2026-08-07-culinary-platform-clean-slate-adoption.md) - [3.10 exploratory adoption evidence](https://decantr.ai/benchmarks/2026-07-24-decantr-3-10-exploratory-adoption-evidence.md) - [3.9.4 Day-0 baseline](https://decantr.ai/benchmarks/2026-07-22-decantr-3-9-4-day-zero.md) - [3.9 historical proof program](https://decantr.ai/programs/2026-07-16-decantr-3-9-adoption-proof-program.md) - [3.9.4 previous stable release note](https://decantr.ai/releases/2026-07-21-decantr-3-9-4-tailwind-source-isolation.md) ## Packages - `@decantr/cli`: primary local workflow. - `@decantr/verifier`: primary authority and evidence engine. - `@decantr/mcp-server`: stable agent integration. - `@decantr/core`: supported low-level graph and execution-pack primitives. - `@decantr/essence-spec`: validation for accepted Essence V4 project law; it does not replace Brownfield production authority. - `@decantr/content`: supported policy/reference corpus. - `@decantr/registry`: legacy compatibility facade. - `@decantr/css`: legacy optional adapter. - `@decantr/telemetry`: optional compatibility package. - `@decantr/vite-plugin`: experimental. Repository: [decantr-ai/decantr](https://github.com/decantr-ai/decantr). Historical separate content-repository references now point to `packages/content` in this monorepo.