Mortdecai
5a2a02e483
This repo opens with the design-discovery work completed before any product code is written. Two model bakeoffs against gemma4:8b/26b/31b on a local Ollama established that: - Whole-puzzle generation in the Connections shape is unreliable on Gemma 4 (gemma4:31b ~50% structural-pass, gemma4:26b ~20-30%); 31b is intentionally out of project scope, so the generation route is harder still. - Atomic semantic-judging skills are reliable: 87.5%/93.75%/100% (8B/26b/31b) on JUDGE; *all three models* scored 10/10 on CREATIVE_ACCEPT — fair judging of player-INVENTED categories. That is the structural unlock vs static hand-curated word games. The README contains the full writeup, the test bench, and a brainstormed bank of 10 distinct game-mechanics ideas across the fast/medium/slow tempo range, plus a primitives table for recombination. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2.5 KiB
IDEA.md — seth_semantic_game
What is this?
A daily word game based on NYT Connections, powered by a locally-hosted Gemma 4 model. Connections gives the player 16 words that have to be sorted into 4 hidden groups of 4 by shared semantic category. The twist for this project — what makes it worth building rather than just playing the original — is whatever Gemma 4 enables that NYT's hand-curated static format cannot.
That twist is not yet decided. That's what brainstorming is for.
The base mechanic is fixed:
- Connections-style grouping puzzle (semantic categories, not letters)
- Gemma 4 in the loop somewhere (puzzle generation, judging, hint system, or all of the above)
- Daily-puzzle structure with social-shareable result (the Connections / Wordle ritual — borrowed only for its sharing pattern, not its gameplay)
This is not Wordle-derived. The original draft of this file framed it as "Wordle-style"; that was wrong. The mechanic is grouping, not letter-guessing.
Problem it solves
Mostly fun and a real use of the local Gemma 4 stack. NYT Connections is hand-curated and ships one puzzle per day; a generative version could ship infinite puzzles, accept fuzzy or creative groupings, generate themed/seeded puzzles, or do other things the hand-built version structurally cannot. Secondary: a daily-puzzle hook for sethpc.xyz alongside other homelab games.
Constraints / preferences
- Self-hosted: Ollama with Gemma 4 on commodity GPU (a single 24 GB card is enough)
- Web frontend, dark theme with orange accents
- If a puzzle is generative, output must be deterministic per day (every player on a given date gets the same puzzle). Likely a date-seeded prompt with cached output rather than a fresh generation per request.
- Per-guess judging cost should be cheap — at most one Gemma call per submission, and ideally answers are precomputed when the daily puzzle is generated, so judging becomes a cheap lookup.
- No login required for casual play (cookies/localStorage for streak)
NOTE on history: this brief was originally a "Wordle-style" framing. That was wrong — the seed game is NYT Connections (16 words → 4 hidden groups of 4). But after the model bakeoffs (see README), the direction shifted again: rather than cloning Connections, the project pivots toward gameplay that uses Gemma's per-call CREATIVE_ACCEPT ability to fairly judge player-INVENTED categories — a thing static curated games structurally can't do. The brainstormed game ideas in the README are what came out of that.