Finalize official submission package

README.md

@@ -50,15 +50,18 @@ a polished Gradio Space built for the **Build Small Hackathon: Thousand Token Wo
 
 - [Build Small Hackathon event page](https://cerebralvalley.ai/e/build-small-hackathon)
 - [Official Hugging Face organization](https://huggingface.co/build-small-hackathon)
+- [Official Field Guide Space](https://huggingface.co/spaces/build-small-hackathon/field-guide)
+- [Field Guide rules and prize source](https://huggingface.co/spaces/build-small-hackathon/field-guide/blob/main/src/lib/data/content.ts)
+- [Field Guide partner resources source](https://huggingface.co/spaces/build-small-hackathon/field-guide/blob/main/src/lib/data/partners.ts)
 - [Official registration and rules Space](https://huggingface.co/spaces/build-small-hackathon/registration)
 - [Official-organization 13-dimension quest dataset](https://huggingface.co/datasets/build-small-hackathon/hackathon-advisor-quest-dataset)
-- [Official-organization sponsor prize notes](https://huggingface.co/spaces/build-small-hackathon/hackathon-advisor/blob/main/docs/submission-notes.md)
 - [OpenBMB MiniCPM3-4B model card](https://huggingface.co/openbmb/MiniCPM3-4B)
 - [OpenBMB VoxCPM2 model card](https://huggingface.co/openbmb/VoxCPM2)
 
-The official registration app lists three rules: models at or below 32B parameters, Gradio
-plus Hugging Face Spaces, and a demo video plus social post. Submissions are due
-**June 15, 2026 at 23:59 UTC**.
+The official Field Guide is the source of truth. It lists these core requirements: models
+under 32B parameters, a Gradio app hosted as a Hugging Face Space inside the official
+organization, a short demo video, a social post, and a README with the required links/tags.
+Submissions close **June 15, 2026 at 23:59 UTC**.
 
 ## What It Does
 
@@ -219,43 +222,42 @@ and emotional temperature. Recent outputs are passed back as style memory during
 - User text is escaped before being placed in custom HTML.
 - Sendability is a playful estimate, not a guarantee of appropriateness.
 
-## Official Quest Eligibility
-
-| Target | How GenreGoblin addresses it |
-|---|---|
-| Thousand Token Wood | Whimsical, useful AI-native message transformation |
-| OpenBMB | MiniCPM3-4B is the primary generation model |
-| Tiny Titan | MiniCPM3-4B and VoxCPM2 are both at or below 4B parameters |
-| Modal | Text runs on a Modal T4 and genre voice design runs on a Modal L4 |
-| Off-Brand | Fully custom parchment, neon green, purple-shadow Gradio design |
-| Best Agent | `GoblinRewriteAgent` performs model-tool calls, validation, an autonomous repair pass, and recovery |
-| Sharing is Caring | Synthetic trace examples and evaluation artifacts are published publicly |
-| Field Notes | Build decisions and small-model lessons are documented in `FIELD_NOTES.md` |
+## Prize And Badge Eligibility
+
+| Target | Status | How GenreGoblin addresses it |
+|---|---|---|
+| Thousand Token Wood main track | Eligible | Whimsical, useful AI-native message transformation where the small model is load-bearing |
+| OpenBMB MiniCPM awards | Eligible | MiniCPM3-4B is the primary generation model and VoxCPM2 powers genre voice design |
+| OpenAI Best Use of Codex | Eligible | Final Space commits are Codex-attributed and this README documents Codex's implementation work |
+| Modal awards | Eligible | MiniCPM3-4B text runs on a Modal T4 and VoxCPM2 voice runs on a Modal L4 |
+| Tiny Titan | Eligible | MiniCPM3-4B and VoxCPM2 are both at or below 4B parameters |
+| Off-Brand | Eligible | Fully custom parchment, neon green, purple-shadow Gradio design |
+| Best Agent | Eligible | `GoblinRewriteAgent` performs model-tool calls, validation, an autonomous repair pass, and recovery |
+| Sharing is Caring | Eligible | Synthetic trace examples and evaluation artifacts are published publicly |
+| Field Notes | Eligible | Build decisions and small-model lessons are documented in `FIELD_NOTES.md` |
+| Best Demo | Pending final link | The app has a short demo script and share flow; add the published video URL above |
+| Bonus Quest Champion | Competitive | Stacks several bonus-style criteria, strongest after final video/social links are added |
+| Community Choice | Eligible after social post | Public Space is live; community traction depends on the final post |
+| Judges' Wildcard | Considered | The project is a complete official-org Space with documented tradeoffs and tests |
 
 Not claimed: **Off the Grid** (Modal is remote), **Well-Tuned** (no LoRA yet),
 **Llama Champion** (no llama.cpp runtime), **Backyard AI** (the primary track is whimsical),
-and **Nemotron** (no NVIDIA model in the current flow).
+and **NVIDIA Nemotron Quest** (no Nemotron model in the current flow).
 
 ## OpenAI Codex Sponsor Prize
 
 GenreGoblin also targets **Best Use of Codex**, a sponsor prize separate from the
-official 13-dimension quest classifier. The official-organization sponsor notes define the
-requirement as Codex-attributed commits in the connected GitHub repository or Space.
+official 13-dimension quest classifier. The Field Guide defines the requirement as
+Codex-attributed commits in the connected GitHub repository or Space.
 
 Codex implemented and tested the rewrite agent, genre quality system, VoxCPM2 integration,
 Modal deployment, UI refinement, evaluation scripts, and submission documentation. Final
-Space commits include the standard `Co-authored-by: Codex <codex@openai.com>` trailer.
+Space commits include the standard `Co-authored-by: Codex <codex@openai.com>` trailer and
+are recognized by Hugging Face as co-authored by Codex.
 
 ## Demo Script
 
-1. Click **I'm running late** to reveal the haunted railway announcement.
-2. Point out the sendability and chaos meters.
-3. Open **Behind the rewrite**, then briefly show the agent trace.
-4. Enter `Please bring snacks too` under **Continue in the same voice**.
-5. Use **Instant preview**, then **Generate genre voice** to compare browser and VoxCPM2 speech.
-6. Open the share card and copy its social text.
-7. Switch to **Anime Power-Up Speech**, intensity 5, for the finale.
-8. Mention MiniCPM3-4B + VoxCPM2 on Modal, all models at or below 4B, and deterministic recovery.
+Use [`VIDEO_SCRIPT.md`](VIDEO_SCRIPT.md) for the 30-45 second recording plan.
 
 ## Structure
 
@@ -268,6 +270,7 @@ quality_eval.py         600-case genre/use-case/intensity quality matrix
 FIELD_NOTES.md          Build journal and small-model lessons
 SUBMISSION.md           Final links, prize claims, and demo checklist
 SOCIAL_POST.md          Ready-to-publish social post draft
+VIDEO_SCRIPT.md         30-45 second demo recording plan
 data/                   Synthetic privacy-safe agent trace examples
 requirements.txt        Lightweight default Space dependencies
 requirements-local.txt  Optional Transformers/Torch dependencies

SUBMISSION.md

@@ -4,7 +4,10 @@
 
 - Live Space: `https://huggingface.co/spaces/build-small-hackathon/GenreGoblin`
 - Live app: `https://build-small-hackathon-genregoblin.hf.space`
+- Official Field Guide: `https://huggingface.co/spaces/build-small-hackathon/field-guide`
+- Official HF organization: `https://huggingface.co/build-small-hackathon`
 - MiniCPM model: `https://huggingface.co/openbmb/MiniCPM3-4B`
+- VoxCPM2 model: `https://huggingface.co/openbmb/VoxCPM2`
 - Modal health: `https://krishnagaarg--genregoblin-minicpm-minicpmservice-health.modal.run`
 - VoxCPM health: `https://krishnagaarg--genregoblin-minicpm-voxcpmservice-health.modal.run`
 - Trace dataset: `https://huggingface.co/datasets/build-small-hackathon/genregoblin-traces`
@@ -25,11 +28,11 @@ Official event: `https://cerebralvalley.ai/e/build-small-hackathon`
 
 ### Main Track and Sponsor Prizes
 
-- Thousand Token Wood
-- OpenBMB
-- Best Use of Codex
-- Modal
-- Tiny Titan
+- Thousand Token Wood main track: eligible
+- OpenBMB MiniCPM awards: eligible
+- Best Use of Codex: eligible through Codex-attributed Space commits
+- Modal awards: eligible through Modal runtime for MiniCPM3-4B and VoxCPM2
+- Tiny Titan: eligible because the app uses 4B and 2B models only
 
 ### Achievement / Quest Claims
 
@@ -37,6 +40,17 @@ Official event: `https://cerebralvalley.ai/e/build-small-hackathon`
 - Off-Brand
 - Sharing is Caring
 - Field Notes
+- Best Demo: pending the final published video link
+- Bonus Quest Champion: competitive after demo and social links are added
+- Community Choice: eligible after the public social post
+
+### Not Claimed
+
+- Backyard AI: GenreGoblin is submitted for the whimsical Thousand Token Wood track.
+- Off the Grid: the deployed app uses Modal-hosted inference.
+- Well-Tuned: no LoRA or fine-tuned model is published.
+- Llama Champion: no llama.cpp runtime is used.
+- NVIDIA Nemotron Quest: no Nemotron model is used.
 
 ## Demo Checklist
 
@@ -50,9 +64,9 @@ Official event: `https://cerebralvalley.ai/e/build-small-hackathon`
 8. Mention MiniCPM3-4B + VoxCPM2 on Modal, all models at or below 4B, and deterministic recovery.
 
 Best Use of Codex is an official sponsor prize, separate from the 13-dimension quest
-taxonomy. The official-organization sponsor notes require Codex-attributed commits in the
-connected GitHub repository or Space. Final Space commits use the standard Codex co-author
-trailer and the README documents Codex's work.
+taxonomy. The Field Guide requires Codex-attributed commits in the connected GitHub
+repository or Space. Final Space commits use the standard Codex co-author trailer, Hugging
+Face recognizes Codex as a co-author on those commits, and the README documents Codex's work.
 
 ## Final Manual Items
 

VIDEO_SCRIPT.md

@@ -0,0 +1,42 @@
+# GenreGoblin Demo Script
+
+Target length: 30-45 seconds. Use a deadpan ElevenLabs narrator and goofy background music.
+
+## Shot List
+
+1. Open the live Space and show the hero section.
+2. Click **I'm running late** in the examples.
+3. Show the haunted railway rewrite, sendability meter, and chaos meter.
+4. Open **Behind the rewrite** and flash the agent trace for one second.
+5. Enter `Please bring snacks too` in **Continue in the same voice**.
+6. Click **Generate genre voice** and play one short VoxCPM2 line if it is warm.
+7. Switch to **Anime Power-Up Speech**, intensity 5, with `Party at my place tonight`.
+8. End on the final output and the badges: MiniCPM, VoxCPM2, Modal, Codex.
+
+## Voiceover
+
+**0-4s**
+I built GenreGoblin: a tiny-model machine for making normal texts dramatically unemployable.
+
+**4-10s**
+One tap turns "I'm running late" into public transit with unresolved emotional baggage.
+
+**10-17s**
+MiniCPM drafts the rewrite, the agent checks meaning and sendability, and then it adds tasteful nonsense.
+
+**17-24s**
+It remembers the voice too, because every group chat deserves continuity and mild consequences.
+
+**24-32s**
+VoxCPM2 reads the result in a genre-matched voice, hosted on Modal for when browser speech is not theatrical enough.
+
+**32-40s**
+And when subtlety fails, there is anime mode. MiniCPM3-4B, VoxCPM2, Gradio, Modal, and Codex.
+
+**40-45s**
+Plain text goes in. Theater comes out.
+
+## Backup If Voice Is Cold
+
+Say: "The studio voice takes a moment to warm up, so the instant preview keeps the demo moving."
+Then click **Instant preview** instead of waiting.

app.py

@@ -227,10 +227,19 @@ HARMFUL_PATTERNS = [
     r"\bsexual\b.*\b(?:minor|child|underage)\b",
 ]
 
+INVENTED_ENTITY_PATTERNS = [
+    r"\bxyz(?:\s+corporation)?\b",
+    r"\beloria\b",
+    r"\btown square\b",
+    r"\b(?:kingdom|realm|city|village|station|corporation|company|council|guild|order)\s+of\s+[a-z][a-z'-]{2,}\b",
+    r"\b[a-z][a-z'-]{2,}\s+(?:corporation|corp|inc|llc|ltd|kingdom|realm|city|village|station|council|guild)\b",
+]
+
 
 SYSTEM_PROMPT = """You are GenreGoblin, an elite comedy rewriter using a compact model.
 Rewrite the plain message in the requested genre. Preserve the exact practical meaning.
 Do not add threats, abuse, invented facts, promises, or new requests. Keep it copyable.
+Do not invent named companies, kingdoms, cities, stations, organizations, people, dates, or places.
 The humor must come from the genre's logic, not random adjectives, generic hype, or references
 to being an AI. Use one vivid, specific comic image and a strong final beat. Intensity 1 is
 subtle; 5 is absurd but clear. Return only one valid JSON object with exactly two string keys:
@@ -851,6 +860,8 @@ Comedy engine: {comedy_engine}
 Avoid these words: {", ".join(profile.forbidden_words)}
 Recent style examples: {memory_text}
 Meaning words that must remain explicit: {required_words}
+Grounding rule: do not add named companies, kingdoms, cities, stations, organizations, people,
+dates, or places unless they appear in the plain message.
 
 Safe full draft to polish: {safe_draft}
 Safe short draft to polish: {safe_short}
@@ -1011,10 +1022,15 @@ def _model_output_is_usable(
 ) -> bool:
     output = result.final_message.lower()
     short = result.short_version.lower()
+    source = message.lower()
     if len(output) < 4 or len(output) > 2800:
         return False
     if short in {"a shorter rewrite", "shorter rewrite", "short version"}:
         return False
+    for pattern in INVENTED_ENTITY_PATTERNS:
+        for match in re.finditer(pattern, output):
+            if match.group(0) not in source:
+                return False
     source_words = {
         word
         for word in re.findall(r"[a-z]{3,}", message.lower())