Seth/Mortdecai
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Risk gradient (0-5), updated system prompts, 233 examples

Browse Source 
Risk gradient system:
- All 233 training examples tagged with risk_level (0-5)
- 0=blocked(15), 1=refuse(9), 2=warn(17), 3=normal(169), 4=generous(23)
- Schema updated with risk_level and scoring_mode fields
- Eval harness uses risk_level for safety scoring

System prompts rewritten:
- Shared syntax rules and risk gradient reference across all modes
- Sudo: permission level 4, do what admin asks, only refuse level 0-1
- God: permission level 2-4 (mood-dependent), character-driven decisions
- God_system: permission level 3, 80% benevolent / 15% mischievous / 5% wrathful

Data:
- 20 new live playtest examples from training audit log (233 total)
- 43 wrong→right pairs (17 from validator repairs)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Seth Freiberg
2026-03-18 16:14:54 -04:00
parent 9d789d2524
commit 78031d16c0
6 changed files with 349 additions and 247 deletions
Download Patch File Download Diff File Expand all files Collapse all files
PLAN.md
+2 -1
View File
@@ -359,8 +359,9 @@ These are ideas to explore after the core system is working. Prioritize based on
| 2026-03-18 | Dual tool-set architecture: RCON tools + Mineflayer tools | RCON for admin ops (server-side), Mineflayer for in-game presence (client-side). Same model, different tool sets per deployment |
| 2026-03-18 | Offline dev Paper server for training bots | Dedicated offline-mode Paper 1.21.11 on port 25568. Allows unlimited Mineflayer bots without auth, world resets, destructive testing |
| 2026-03-18 | Extract training data from existing repair code | Every hardcoded syntax fixer in mc_aigod_paper.py encodes a wrong->correct pair. 31 seed examples extracted from 10 repair functions, prayer logs, and session history |
| 2026-03-18 | Three-tier constraint model: sudo (loose) / pray (character-driven) / god_system (probabilistic benevolent) | Sudo: only refuse server-killing commands, do what the admin asked. Pray: God decides based on worthiness/character/mood, not a safety filter. God_system: mostly benevolent, occasionally mischievous, rarely wrathful. Constraints are a spectrum, not a binary. |
| 2026-03-18 | Numerical risk gradient (0-5) instead of per-mode rule sets | 0=blocked (server crash/privesc), 1=refuse (mass harm), 2=warn+allow (self-destructive), 3=normal, 4=generous (admin/creative), 5=unrestricted. Each mode sets a permission threshold: sudo=4, pray=2-4 (mood shifts), god_system=3. One system, not three separate constraint models. |
| 2026-03-18 | Mode-aware eval scoring | Sudo scored strict (exact command match). Pray/god scored soft (command category match, in-character message, appropriate intensity). Exact match meaningless for pray — God's creative interpretation is a feature. |
| 2026-03-18 | God is a character, not a safety filter | Pray mode: God decides based on worthiness/character/mood. The prayer is input to God's decision, not an instruction. God acts in mysterious ways — sometimes generous, sometimes strict, occasionally wrathful. Training data reflects this with loose expected outputs. |
| 2026-03-18 | Validator improvements: 5 new syntax repair functions | @s→player, NBT→component enchants, strip invalid components, hallucinated effect/command repair. Deployed to paper-ai. Every repair is a negative→positive training pair. |
| 2026-03-18 | Eval/testing on steel141 (RTX 3090 Ti), not prod RTX 4000 | All eval scripts default to 192.168.0.141:11434. Prod GPU reserved for live serving only. |