Loop-engineer comic generation

.github/workflows/build.yaml

@@ -24,6 +24,9 @@ jobs:
       - name: Build
         run: bun run build
 
+      - name: Comic loop tests
+        run: bun run test:comic-loops
+
       - name: Contract tests
         run: bun run test:contracts
 

COMIC-README.md

@@ -12,7 +12,7 @@ A daily webcomic featuring:
 - 🧨 **Simon**: BOFH sysadmin who loves chaos
 - 🐱 **The Cat**: Breaks the robot's reasoning
 
-Each day, **two different AI models** generate a comic from the same prompt. Readers **vote** for their favorite variant!
+Each day, a loop-engineered writers room builds a technical brief, ranks several joke premises, and gives **two different AI models** distinct premise mechanisms. Readers **vote** for their favorite variant.
 
 ## Features
 
@@ -41,7 +41,8 @@ Backend (Cloudflare)
   └─ Web Push (Notifications)
 
 Generation Pipeline
-  ├─ comic-generator.ts (Prompt → Script)
+  ├─ comic-loop.ts (Brief → Premises → Score → Retry/Invert)
+  ├─ comic-generator.ts (Selected Premise → Script/Rewrite)
   ├─ svg-renderer.ts (Script → SVG)
   └─ _worker.js (Orchestration)
 ```
@@ -58,12 +59,14 @@ Generation Pipeline
 
 ## Comic Format
 
-4-panel strip with consistent structure:
+Three or four panels with a consistent compositional contract:
 
-**Panel 1**: Setup (Human asks question)
-**Panel 2**: Robot internal reasoning (monospace thought bubble)
-**Panel 3**: Robot response (speech bubble)
-**Panel 4**: Punchline (usually Simon's deadpan comment)
+**Contract**: Establish the reasonable expectation.
+**Interpretation**: Reveal how the system optimized the requirement.
+**Exposure**: Show the difference between the proxy and the intent.
+**Optional reframe**: Change the target or reinterpret the setup.
+
+Robot internal reasoning is optional. Exact dashboard, diff, alert, log, or terminal text is rendered deterministically through `screenText` when it carries the visual joke.
 
 ## Example Comic
 
@@ -74,21 +77,22 @@ Panel 1:
 Human: "Robot, explain Kubernetes simply."
 
 Panel 2:
-Robot thought bubble:
-> request detected
-> user comprehension estimate: low
-> generating analogy
+Screen: `tests/evals.json — SCORE 100%`
+Robot thought: `> answer key found`
 
 Panel 3:
-Robot: "It's like containers... but with more containers."
+Human: "It memorized the repository."
 
 Panel 4:
-Simon (walking past): "Accurate."
+Simon: "Promote grep. It's cheaper."
 ```
 
 ## Recurring Themes
 
 - AI hallucinations and confidence
+- Proxy metrics and specification gaps
+- Agent permissions and approval theatre
+- Eval contamination and automation incentives
 - Prompt injection attacks
 - Cloudflare product placement (subtle)
 - DevOps disasters
@@ -99,8 +103,10 @@ Simon (walking past): "Accurate."
 ## Voting System
 
 Each day presents two variants:
-- **Variant A**: Generated by Model A (e.g., Llama 70B)
-- **Variant B**: Generated by Model B (e.g., Mistral 7B)
+- **Variant A**: Highest-ranked premise mechanism, generated by Model A
+- **Variant B**: A distinct mechanism and target, generated by Model B
+
+Drafts are scored for surprise, specificity, compression, visuality, character voice, and archive novelty. A failed draft gets one focused rewrite and, if still weak, one TRIZ inversion pass. See [workflows/comic-generation-loops.md](./workflows/comic-generation-loops.md).
 
 Readers vote for:
 - Better humor

functions/api/test-generate.ts

@@ -103,6 +103,9 @@ export async function onRequestPost(context: any) {
       character_count: result.character_count,
       topic_candidates: result.topic_candidates,
       selected_topic: result.selected_topic,
+      brief: result.brief,
+      selected_premises: result.selected_premises,
+      script_evaluations: result.script_evaluations,
       cast: result.cast,
       models: {
         a: result.model_a,