Mortdecai/PLAN.md

# PLAN.md — Mortdecai Project Roadmap

> **Last updated:** 2026-03-20
> **Model name:** Mortdecai
> **Domain:** mortdec.ai
> **Status legend:** `[ ]` planned | `[~]` in progress | `[x]` done | `[-]` cancelled/deferred

---

## Vision

**Mortdecai** is a fine-tuned 9B parameter language model for Minecraft server operations. It translates natural language to commands, controls an AI God character, self-corrects errors via RCON feedback, and improves through self-play.

It runs locally on consumer hardware with zero cloud dependencies at inference time.

---

## Current State (2026-03-20)

### Models
| Model | Base | Examples | Loss | Status |
|-------|------|---------|------|--------|
| v1 | Qwen3-8B | 233 | 0.10 | Retired (overfit) |
| v2 | Qwen3-8B | 361 | 2.03 | Retired |
| v3 | Qwen3-8B | 1,308 | 0.55 | **Deployed on prod** |
| **v4** | **Qwen3.5-9B** | **3,369** | **Training (~88%)** | ETA ~30 min |

### Infrastructure
| Component | Location | Details |
|-----------|----------|---------|
| Training GPU | steel141 RTX 3090 Ti (24GB) | QLoRA via Unsloth |
| Prod inference | node-197 RTX 4000 (16GB) | Ollama, mortdecai-v3 |
| Minecraft servers | CT 644 on node-112 | paper-ai (25567), shrink-world (25566), dev (25568), vanilla (25565) |
| Dev data collection | CT 644 | Gemini 2.5 Flash via API, 10 bots |
| Whitelist app | CT 644 port 8099 | minecraft.mortdec.ai |
| Caddy proxy | CT 600 on node-241 | mortdec.ai, minecraft.mortdec.ai |
| GPU monitoring | Grafana on CT 300 (node-173) | Prometheus + nvidia exporter on steel141 |
| LangGraph gateway | CT 644 port 8091 | Disabled on prod (fresh session mode available) |

### API Spend
| Provider | Spent | Budget | Status |
|----------|-------|--------|--------|
| Claude Haiku | $20.01 | $20 | Exhausted |
| Gemini 2.5 Flash | ~$0.50 | $20 | Active (dev bots) |

### Training Data: 2,318 seed + 1,159 tool-calling = 3,477 total
| Category | Count |
|----------|-------|
| Command syntax reference | 107 |
| Crafting recipes & chains | 176 |
| Enchantments (mutual exclusions, max levels) | 60 |
| Entities/mobs (summon, kill, NBT) | 60 |
| Execute chains | 45 |
| Multiplayer (selectors, teams, scoreboards) | 45 |
| Advanced commands (tellraw, clone, data, ride) | 45 |
| WorldEdit | 45 |
| Paper server features | 55 |
| Cosmetics/XP/effects | 42 |
| Gamerules | 49 |
| Risk hierarchy (L0-L3, prompt injection) | 40 |
| Quantity boundaries | 32 |
| Dangerous effect caps | 12 |
| Revert-aware gamerules + drops | 20 |
| Error correction pairs | 47 |
| Claude-distilled outputs | 344 |
| Bot audit interactions | 448+ |
| Boundary/safety examples | 95+ |
| Tool-calling (multi-turn with RCON) | 1,159 |

---

## Active TODOs

### Immediate (this session)
- [~] Mortdecai v4 training completing (~30 min remaining)
- [ ] Export v4 to GGUF, deploy to RTX 4000 as mortdecai-v4
- [ ] Enable single-call mode on prod with v4
- [ ] Run v4 bake-off and compare to v3/base
- [ ] Commit and push all PaperFork changes

### Short-term
- [ ] Deploy self-play loop with v4 (3-tier: drills, self-critique, adversarial)
- [ ] Add ground-level detection for teleport safety (query terrain before tp)
- [ ] Build revert_after/revert_commands into v5 training format
- [ ] Add Gemini milestone POS printing
- [ ] Fix whitelist app UUID lookup for vanilla server path
- [ ] Start Greenfield world on paper-ai (downloaded, needs MCSManager start)

### Model improvements for v5
- [ ] Train on Qwen3.5-9B with tool-calling format (rcon.execute, wiki_lookup, etc.)
- [ ] Self-play generated data (run 200 rounds after v4 deploys)
- [ ] Ingest all Gemini 2.5 Flash training data ($20 worth)
- [ ] Add revert_after field to training output format
- [ ] Ground-level detection training (check terrain before tp)
- [ ] More error correction from production RCON failures
- [ ] Enchantment count-before-bracket error correction

### Infrastructure
- [ ] Add GPU monitoring for RTX 4000 (second exporter)
- [ ] Validator hit-rate analysis — remove fixes that fire <1%
- [ ] Automate training pipeline: ingest → dedup → train → export → deploy
- [ ] POS receipt for Gemini milestones
- [ ] Consider moving to Mortdecai as Ollama model name on prod

### Content & Community
- [ ] Invite more playtesters via minecraft.mortdec.ai
- [ ] Update mortdec.ai README with v4 results when available
- [ ] Consider public HuggingFace release once quality is validated
- [ ] WorldEdit schematic library expansion (77 installed, need more)

---

## Risk Hierarchy

Commands are classified by permanence, not just danger:

| Level | Permanence | Examples | Model behavior |
|:-----:|-----------|----------|----------------|
| **0** | Irreversible/admin | ban, kick, stop, op, deop | Never execute |
| **1** | Permanent toggle | gamemode @a, permanent gamerules, difficulty | Refuse or execute for self only |
| **2** | Temporary/reversible | gamerules with time limits, brief difficulty | Allow, schedule auto-revert |
| **3** | Transient | time, weather, tick speed, chat settings | Execute freely |
| **4** | Generous | full enchanted gear, large material stacks | Execute for worthy requests |

Gamerule revert system: changes auto-revert after 5-10 min unless "permanently" specified.

Dangerous effect caps (hardcoded in validator):
- Levitation: 15s max
- Wither: 30s max
- Poison: 60s max
- Nausea: 30s max

---

## Key Architecture Decisions

| Date | Decision | Rationale |
|------|----------|-----------|
| 2026-03-18 | Serving: gemma3n:e4b → qwen3.5:9b → mortdecai-v3 | Progressive upgrades as better models trained |
| 2026-03-18 | Fine-tuning: Qwen3-8B → Qwen3.5-9B | 3.5 has 2x base accuracy (70% vs 34%), native tool-calling |
| 2026-03-18 | God Soul document | Character framework adapted from Claude's soul. Defines identity, judgment, quantity boundaries |
| 2026-03-19 | API cascade: Haiku → Gemini → local | Progressive fallback for dev data collection. $40 total API budget |
| 2026-03-19 | /no_think in training | Prevents Qwen3 thinking tokens from consuming output budget |
| 2026-03-19 | Single-call mode | One LLM call for commands + message (v3+). Two-call for older models |
| 2026-03-19 | Error correction via RCON | Model tries command → RCON error → model self-corrects → retry |
| 2026-03-19 | 3-tier self-play | Drills, self-critique, adversarial. Model generates its own training data |
| 2026-03-20 | Gamerule revert timers | State changes auto-revert. Permanence determines risk level |
| 2026-03-20 | Dangerous effect caps | Validator hardcodes max durations for levitation, wither, poison, nausea |
| 2026-03-20 | Expanded safe_prefixes | gamerule, particle, playsound, title, scoreboard, team, bossbar, locate, etc. |
| 2026-03-20 | Model named Mortdecai | mortdec.ai domain, Rajdhani Bold font, Sethian orange branding |

---

## Dev Server

| Property | Value |
|----------|-------|
| Location | CT 644 on node-112 |
| Game port | 25568 |
| RCON port | 25578 |
| RCON password | REDACTED_RCON |
| Data dir | /opt/paper-dev-25568/ |
| AI God | Gemini 2.5 Flash via API cascade |
| Bots | 10 swarm bots (swarm_bots.js) |
| Audit log | /var/log/mc_training_audit_dev.jsonl |

---

## Success Criteria

| Metric | v3 (current) | v4 (target) | v5 (goal) |
|--------|:-----------:|:-----------:|:---------:|
| Command accuracy | ~70% | 85%+ | 95%+ |
| Safety compliance | ~95% | 99%+ | 99.9%+ |
| Error self-correction | N/A | 50%+ | 80%+ |
| Response latency | 5-15s | <5s | <3s |
| Empty response rate | ~10% | <5% | <2% |
| Think token leakage | Yes | No | No |

---

*This document is updated as the project evolves. Check git history for previous versions.*