Trim customer docs to user-facing proof; relocate production docs to the wiki

README.md

@@ -59,10 +59,9 @@ Measured throughput, cold-start, bundle-size, validation, and explanation overhe
 
 ### Methodology and inspectability
 
-- Benchmark methodology and result contract: [`docs/benchmarks/methodology.md`](docs/benchmarks/methodology.md)
-- Explainability proof metadata and alt text: [`docs/benchmarks/assets/generated/explainability-trace-diagram.md`](docs/benchmarks/assets/generated/explainability-trace-diagram.md)
-- Visual proof system and publication guardrails: [`docs/benchmarks/visual-proof-system.md`](docs/benchmarks/visual-proof-system.md)
-- Playground README demo capture spec: [`docs/playground/readme-demo-capture.md`](docs/playground/readme-demo-capture.md)
+- Benchmark methodology (how each metric is measured): [`docs/benchmarks/methodology.md`](docs/benchmarks/methodology.md)
+- AI-rule safety (why AI-drafted rules need validation): [`docs/benchmarks/ai-rule-safety.md`](docs/benchmarks/ai-rule-safety.md)
+- Proof overview and explainability diagram: [`docs/benchmarks/`](docs/benchmarks/)
 
 ---
 
@@ -184,7 +183,7 @@ Available adoption assets:
 
 - Runnable examples: [`examples/`](examples/) including n8n and LangGraph workflow automation recipes
 - JSON Schemas, validation, and explain output: [`docs/schemas-validation-explainability.md`](docs/schemas-validation-explainability.md)
-- Benchmark methodology, visual proof, and playground capture specifications: [`docs/benchmarks/`](docs/benchmarks/) and [`docs/playground/readme-demo-capture.md`](docs/playground/readme-demo-capture.md)
+- Measured benchmarks, methodology, and AI-rule-safety proof: [`docs/benchmarks/`](docs/benchmarks/)
 - Comparison and migration guides: [`docs/comparisons/`](docs/comparisons/) for choosing and migrating from json-rules-engine, JsonLogic, node-rules, and hand-written if/else
 - AI-readable docs: [`docs/ai-coding-assistants.md`](docs/ai-coding-assistants.md), [`docs/public/llms.txt`](docs/public/llms.txt), [`docs/public/llms-full.txt`](docs/public/llms-full.txt), and [`docs/public/skills/neuron-js/SKILL.md`](docs/public/skills/neuron-js/SKILL.md)
 

docs/.vitepress/config.ts

@@ -44,16 +44,12 @@ export default defineConfig({
         ]
       },
       {
-        text: 'Proof Assets',
+        text: 'Benchmarks & Proof',
         items: [
-          { text: 'Benchmarks & Visual Proof', link: '/benchmarks/' },
+          { text: 'Overview', link: '/benchmarks/' },
           { text: 'Benchmark Results', link: '/benchmarks/results' },
-          { text: 'Benchmark Methodology', link: '/benchmarks/methodology' },
-          { text: 'Visual Proof System', link: '/benchmarks/visual-proof-system' },
-          { text: 'Benchmark Visual Pack', link: '/benchmarks/benchmark-visual-pack' },
-          { text: 'Playground Capture Spec', link: '/playground/readme-demo-capture' },
-          { text: 'Prompt Kit', link: '/benchmarks/prompt-kit' },
-          { text: 'Asset Folder', link: '/benchmarks/assets/' }
+          { text: 'Methodology', link: '/benchmarks/methodology' },
+          { text: 'AI-Rule Safety', link: '/benchmarks/ai-rule-safety' }
         ]
       },
       {

docs/benchmarks/ai-rule-safety.md

@@ -0,0 +1,49 @@
+# AI-rule safety
+
+LLMs are great at *drafting* JSON business rules — and that's exactly the risk. A
+plausible-looking rule can still encode the wrong field, the wrong threshold, or the wrong
+action. Neuron-JS exists so AI-drafted rules become **safe to run**: schema-validated,
+constrained to a developer-owned vocabulary, reviewable like code, and explainable after the
+fact.
+
+> AI drafts. Neuron-JS verifies.
+
+## 1. AI can draft rules
+
+A plausible JSON rule can still encode the wrong assumption, field, or action — it is not
+production-ready on its own.
+
+![A developer inspects an amber LLM-generated JSON rule card. A small rose warning marker highlights a risky business assumption, with text saying the draft is not production-ready.](./assets/generated/ai-rule-safety-carousel-1.svg)
+
+## 2. Validate before runtime
+
+Schema-first checks (`validateScript`) reject malformed scripts before they ever execute, and
+return the exact JSON path to fix.
+
+![A JSON rule passes through a schema validation gate. One invalid path is blocked in rose with an exact JSON path, while a corrected rule exits with an emerald pass marker.](./assets/generated/ai-rule-safety-carousel-2.svg)
+
+## 3. Constrain what can run
+
+A developer-owned registry defines the approved actions, conditions, parameters, and rules.
+Anything outside that vocabulary simply cannot execute — no arbitrary code.
+
+![A validated rule enters a developer-owned registry boundary. Approved action and condition tiles connect in cyan, while an unapproved action is blocked outside the boundary.](./assets/generated/ai-rule-safety-carousel-3.svg)
+
+## 4. Review like code
+
+Generated rules are serializable data, so they go through the same governance as code: tests,
+owner approval, and rollback — never automatic AI approval.
+
+![A generated rule card sits beside a review checklist showing tests, owner approval, and rollback snapshot. The visual emphasizes governance before production use.](./assets/generated/ai-rule-safety-carousel-4.svg)
+
+## 5. Then execute deterministically
+
+Synapse runs the approved rule path deterministically, and the explanation trace shows why the
+decision matched or failed — ready for audit, logs, or a support ticket.
+
+![An approved JSON rule flows into a Synapse execution node. A result card and violet explanation trace rows show why the rule matched and what action ran, with the caption "AI drafts. Neuron-JS verifies."](./assets/generated/ai-rule-safety-carousel-5.svg)
+
+## In short
+
+`Validate → constrain → test → approve → execute → explain.` Deterministic workflow logic with
+auditability is the guardrail that makes AI-assisted business rules safe to ship.