diff --git a/README.md b/README.md
index db81148..c9026f6 100644
--- a/README.md
+++ b/README.md
@@ -96,10 +96,18 @@ pnpm install --frozen-lockfile
 pnpm check
 ```
 
-`pnpm check` runs formatting validation plus the eval package's deterministic checks:
-skill/profile static checks, Vitest fixture tests, and the seeded `enforce-architecture`
-dependency-cruiser evals. Manual or model-graded evals use the root `pnpm eval:*` commands backed
-by `@agentic-workflow-kit/eval-kit`.
+`pnpm check` is the fast repository gate. It runs only offline/static/local deterministic checks:
+formatting validation, root-visible contract checks, adapter import and syntax validation, fixture
+manifest validation, Vitest fixture tests, and the seeded `enforce-architecture` dependency-cruiser
+evals. It must not run Promptfoo provider calls, Codex/OpenAI generation, LLM judge coverage,
+pairwise judging, long replay suites, or anything requiring auth, network, model calls, or manual
+calibration.
+
+Before significant changes to skills, methodology templates, fixtures, or eval logic, run the
+deterministic local eval slice that matches the change. Use `pnpm eval:case -- --case <case-id>
+--candidate <candidate.md>` for a targeted product-to-design case, then widen back to `pnpm check`.
+Manual model-assisted evals use direct `pnpm exec eval-kit ...` commands and remain advisory outside
+CI.
 
 ## Methodology model
 
diff --git a/docs/design/evaluation-strategy.md b/docs/design/evaluation-strategy.md
index f87f970..4906990 100644
--- a/docs/design/evaluation-strategy.md
+++ b/docs/design/evaluation-strategy.md
@@ -35,9 +35,37 @@ The design principle is:
 - Do not treat one public reference design as the only correct answer.
 - Do not use broad 1-10 scores as release gates.
 - Do not allow an LLM judge verdict to override deterministic source or schema failures.
+- Do not put Promptfoo, Codex/OpenAI generation, LLM judge coverage, pairwise judging, long replay
+  suites, or any auth/network/model-assisted eval into `pnpm check` or CI.
 - Do not add private application names or private repository facts to eval fixtures.
 - Do not require every eval case to exercise every skill.
 
+## Gate Policy
+
+The global eval policy has three lanes:
+
+| Lane                 | Purpose                                      | Runs in `pnpm check` / CI |
+| -------------------- | -------------------------------------------- | ------------------------- |
+| Fast repository gate | Offline/static/local deterministic checks    | Yes                       |
+| Local on-demand      | Deterministic case runs before major changes | No, run explicitly        |
+| Manual advisory      | Promptfoo/Codex/model-assisted evals         | No                        |
+
+`pnpm check` is limited to format/lint checks, static docs/schema validation, adapter import and
+syntax validation, fixture manifest validation, local unit tests for graders/helpers,
+non-online deterministic structural checks, and seeded local fixture checks that do not call
+external providers.
+
+Local on-demand deterministic evals should run before significant changes to skills, methodology
+templates, fixtures, graders, or eval policy. Prefer the smallest targeted `pnpm eval:case` or
+individual deterministic eval command that exercises the changed behavior, then widen back to
+`pnpm check`.
+
+Manual/advisory evals cover model-assisted candidate generation, pointwise LLM judge coverage,
+pairwise judging, long product-to-design or product-to-plan session evals, expensive full-case replay
+suites, and anything requiring auth, network, model calls, or manual calibration. These methods must
+remain disabled by default in `evals/eval-kit.config.json` and require an intentional local edit
+before use.
+
 ## Evaluation Pyramid
 
 Cheap, deterministic checks sit at the base. Expensive product-to-design case replays sit at the
@@ -258,6 +286,9 @@ comparison anchors, not answer keys.
 
 ### LLM Semantic Judge
 
+LLM semantic judging is manual and advisory. It is never part of `pnpm check` or CI, and its config
+method stays disabled by default.
+
 Use an LLM judge only where deterministic and fact graders cannot reasonably decide:
 
 - coherence of bounded contexts;
@@ -411,13 +442,16 @@ Non-functional requirements:
 
 ## Verification Plan
 
-1. Keep `pnpm check` as the required repository gate.
+1. Keep `pnpm check` as the required repository gate and keep it offline/static/local
+   deterministic-only.
 2. Add static checks when a new case format is introduced.
 3. Add at least one deliberately failing fixture for every deterministic rule.
-4. Run a first golden-set pass manually before trusting any LLM-judge verdict.
-5. Re-run pairwise regression cases before changing methodology templates, review rubrics, or
-   orchestration behavior.
-6. Review eval failures as design signals first; only relax rubrics after confirming the expected
+4. Run targeted deterministic `pnpm eval:case` checks before significant skill, methodology,
+   fixture, grader, or eval-policy changes.
+5. Run a first golden-set pass manually before trusting any LLM-judge verdict.
+6. Re-run pairwise regression cases manually before accepting methodology-template, review-rubric,
+   or orchestration behavior changes that need semantic comparison.
+7. Review eval failures as design signals first; only relax rubrics after confirming the expected
    case fact is wrong or too narrow.
 
 ## Reviewer Workflow
diff --git a/docs/design/promptfoo-codex-eval-pilot.md b/docs/design/promptfoo-codex-eval-pilot.md
index 10ceecb..a40a216 100644
--- a/docs/design/promptfoo-codex-eval-pilot.md
+++ b/docs/design/promptfoo-codex-eval-pilot.md
@@ -2,11 +2,15 @@
 
 ## Goal
 
-Build a small, repeatable eval pilot that proves `technical-design` can run an actual design task,
-grade the generated design, judge it semantically, and produce a readable final report.
+Build a small, repeatable manual eval pilot that proves `technical-design` can run an actual design
+task, grade the generated design, judge it semantically, and produce a readable final report.
 
 The pilot starts with one case: `case-tiny-laundry-pickup-v1`.
 
+This pilot is advisory. It must stay outside `pnpm check`, CI, and required contributor gates because
+it uses local auth, Promptfoo provider calls, Codex/OpenAI model-assisted generation, and LLM
+judging.
+
 ## What We Need
 
 The pilot needs one end-to-end eval flow:
@@ -44,6 +48,8 @@ judge rubric, result bundle convention, and final reporting story.
   false-fail behavior.
 - The reference design is a comparison anchor, not a certified great solution or exact-output target.
 - Model-graded commands remain manual and outside `pnpm check`.
+- Model-assisted methods remain disabled by default and require an intentional local config edit
+  before each manual run.
 
 ## Tooling Decision
 
@@ -145,9 +151,10 @@ Root-level docs may still describe the evaluation strategy, because that strateg
 product is designed and governed. Executable eval assets, fixture data, generated outputs, schemas,
 and eval-specific implementation docs belong in `evals/`.
 
-The root package stays small. It keeps `check` as the public deterministic repo gate and exposes
-only the basic deterministic eval-kit commands. Model-assisted commands remain direct `eval-kit`
-CLI usage because they require local auth, explicit calibration, and human interpretation:
+The root package stays small. It keeps `check` as the public deterministic repo gate and exposes only
+offline/static/local deterministic scripts such as `pnpm eval:case`, `pnpm eval:unit`, and
+`pnpm eval:enforce`. Model-assisted commands remain direct `eval-kit` CLI usage because they require
+local auth, explicit calibration, and human interpretation:
 
 ```text
 pnpm eval:doctor
@@ -159,8 +166,9 @@ pnpm exec eval-kit judge-pairwise --config evals/eval-kit.config.json <args>
 pnpm exec eval-kit report --config evals/eval-kit.config.json <args>
 ```
 
-The root `pnpm check` command remains the public repository gate. It should run formatting and the
-eval package's deterministic check.
+The root `pnpm check` command remains the public repository gate. It should run formatting,
+root-visible static checks, adapter import and syntax validation, fixture validation, local unit
+tests, and seeded deterministic fixture checks only.
 
 ## Planned Suite Shape
 
@@ -251,8 +259,11 @@ The final report command combines all run bundles and states:
 ## Acceptance Criteria
 
 - `pnpm check` remains green and deterministic-only.
+- `pnpm check` does not invoke generation, LLM judge coverage, pairwise judging, Promptfoo provider
+  calls, network/auth/model calls, long replay suites, or manual calibration.
 - The pilot does not require or document `OPENAI_API_KEY`.
 - `codex login status` is checked before model-graded runs.
+- Model-assisted methods remain disabled by default in `evals/eval-kit.config.json`.
 - Promptfoo, not shell wrappers, owns provider execution, assertions, and JSON/HTML result exports.
 - Candidate generation produces a non-empty Markdown design for `case-tiny-laundry-pickup-v1`.
 - The existing deterministic grader runs against the generated candidate.
diff --git a/evals/README.md b/evals/README.md
index 907f2f7..b26ab25 100644
--- a/evals/README.md
+++ b/evals/README.md
@@ -60,7 +60,7 @@ The product-to-design cases graded by `eval:case`:
 - `case-openfeature-evaluation-api-v1`
 - `case-tiny-laundry-pickup-v1`
 
-## Running checks
+## Running Checks
 
 The repository gate:
 
@@ -68,6 +68,13 @@ The repository gate:
 pnpm check
 ```
 
+`pnpm check` is intentionally fast, offline, and deterministic. It is limited to formatting,
+static docs/schema validation, adapter import and syntax validation, fixture manifest validation,
+local unit tests for graders/helpers, seeded local fixture checks, and the deterministic
+`enforce-architecture` eval. It does not call Promptfoo providers, Codex/OpenAI models, LLM judges,
+pairwise judging, long product-to-design replay suites, or anything that needs auth, network, model
+calls, or manual calibration.
+
 Individual eval checks:
 
 ```bash
@@ -82,6 +89,9 @@ Run a single deterministic case:
 pnpm eval:case -- --case case-tiny-laundry-pickup-v1 --candidate evals/cases/case-tiny-laundry-pickup-v1/reference-design.md
 ```
 
+Before significant skill, methodology, fixture, or grader changes, run the smallest deterministic
+case or eval slice that exercises the changed behavior, then widen to `pnpm check` before review.
+
 ## Reading the evals
 
 1. `eval-kit.config.json` — suite config; start here to understand what runs.
@@ -90,10 +100,10 @@ pnpm eval:case -- --case case-tiny-laundry-pickup-v1 --candidate evals/cases/cas
 4. `cases/README.md` — authoring contract for product-to-design cases.
 5. `results/README.md` — conventions for generated run outputs.
 
-## Manual Product-to-Design Flow
+## Manual Model-Assisted Product-to-Design Flow
 
 Manual model-graded runs use Promptfoo with the Codex App Server provider. They are advisory only
-and not part of `pnpm check`.
+and must stay outside `pnpm check`, CI, and required contributor gates.
 
 Check local auth:
 
diff --git a/package.json b/package.json
index 88a7239..863616e 100644
--- a/package.json
+++ b/package.json
@@ -43,7 +43,7 @@
     "eval:enforce": "node scripts/run_enforce_eval.mjs"
   },
   "devDependencies": {
-    "@agentic-workflow-kit/eval-kit": "github:agentic-workflow-kit/eval-kit#v0.1.1",
+    "@agentic-workflow-kit/eval-kit": "github:agentic-workflow-kit/eval-kit#v0.1.2",
     "ajv": "^8.17.1",
     "dependency-cruiser": "^18.0.0",
     "prettier": "^3.6.2",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 4f12c80..0d91af4 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -9,8 +9,8 @@ importers:
   .:
     devDependencies:
       '@agentic-workflow-kit/eval-kit':
-        specifier: github:agentic-workflow-kit/eval-kit#v0.1.1
-        version: https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/2d3bf2da4533fa4442e0ae2a5a512c915c546d89
+        specifier: github:agentic-workflow-kit/eval-kit#v0.1.2
+        version: https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/e45db38c0e5b151942ebdf579089d0bdd284f974
       ajv:
         specifier: ^8.17.1
         version: 8.20.0
@@ -32,9 +32,9 @@ importers:
 
 packages:
 
-  '@agentic-workflow-kit/eval-kit@https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/2d3bf2da4533fa4442e0ae2a5a512c915c546d89':
-    resolution: {gitHosted: true, integrity: sha512-4sSEGmu2oZ0wiz3V0TEQ6qzq/G+YBppcfbKT3QRCWmakRx7kzqY93VFl5q9Wfw6GLUZ8OsETv8yiCQs3sGQZ3Q==, tarball: https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/2d3bf2da4533fa4442e0ae2a5a512c915c546d89}
-    version: 0.1.1
+  '@agentic-workflow-kit/eval-kit@https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/e45db38c0e5b151942ebdf579089d0bdd284f974':
+    resolution: {gitHosted: true, integrity: sha512-XNdPutRH/06q2cI2jxaZCBgh583JVoAc8W9uipxnpgNzcovcSY8TK5feFnk4L32RG5oGdl8wFGzpnXlFsAaRnA==, tarball: https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/e45db38c0e5b151942ebdf579089d0bdd284f974}
+    version: 0.1.2
     engines: {node: '>=22.13.0', pnpm: '>=11.9.0'}
     hasBin: true
 
@@ -4202,7 +4202,7 @@ packages:
 
 snapshots:
 
-  '@agentic-workflow-kit/eval-kit@https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/2d3bf2da4533fa4442e0ae2a5a512c915c546d89':
+  '@agentic-workflow-kit/eval-kit@https://codeload.github.com/agentic-workflow-kit/eval-kit/tar.gz/e45db38c0e5b151942ebdf579089d0bdd284f974':
     dependencies:
       ajv: 8.20.0