Window and screen observation paths, benchmark frames, game profiles, and screenshot evidence.
Detailed Project Profile / Manna-Gamer
Active local prototype
Vision-first
Brain Monitor
Execute gated
A local game AI learning from what it can actually see.
Manna-Gamer is the long arc toward a personal AI that can learn to play games through pixels, bounded keyboard and mouse actions, episode logs, replay review, and careful evidence gates instead of hidden game APIs or fake progress.
The flagship story is Hardware vs Ender Dragon, but the architecture is intentionally broader than Minecraft: observe the screen, choose a bounded action, log the result, review what happened, and only trust improvements that survive held-out validation.
What Exists Now
The project has a real evidence loop
The important part is not that an AI can press keys in a toy course. The important part is that Manna-Gamer now records enough evidence to learn from successes, failures, and tempting false positives.
Bounded keyboard and mouse primitives with emergency-stop thinking before live control.
Episode artifacts, visible rollout labels, outcome reports, and replay comparison files.
Brain Monitor panels for graph state, scenario families, weight transfer, replay training, and validation.
Latest Batch Reality
What the June 8 work actually proved
- Generated multiple synthetic obstacle-course layouts instead of one timing variant
- Target-reaching examples included north-zigzag and airport-detour
- Harder layouts exposed looping and regression behavior worth preserving
- Compared each scenario's best learned weights against the others
- Showed that held-out-only averages can reward misleading generic weights
- Kept balanced transfer score as the safer rank signal
- Moved from action matching toward actual outcome score
- Scaled replay training looked promising at 320x240
- Full-resolution validation failed at 640x480, so weights stay unpromoted
Design Doctrine
Why this is not just a Minecraft bot
Manna-Gamer can use Minecraft as the story people understand, but the core needs to stay general: pixels, evidence, review, and bounded action across games.
Live control should come from current pixels and after-frame validation, not hidden state or privileged APIs.
Core ruleGame-specific profiles and providers can exist, but they should declare scope and stay behind route-authority gates.
ScopedGuides and wikis can inform goals, object affordances, and curriculum, but not become hidden live control.
Planning onlyA pretty action-match score is not enough. Held-out replay outcome and full-resolution validation get the final word.
Gate ruleBrain Monitor
The project should be watchable while it learns
The monitor surfaces winners, candidate branches, rollout labels, scenario-family summaries, transfer ranks, replay-training results, and validation failures with execute authority clearly separated from explanation.
Grayson gets the fun of visible evolving-brain energy without pretending the system has earned autonomy before the evidence says so. It is both exciting and disciplined.
Boundaries
What the project should not claim yet
Single-player and local/private testing only. No multiplayer automation, evasion, or anti-cheat bypass lane.
The flagship story is long-term. The first job is truthful competence on smaller visible tasks.
Current generation-style scoring is evidence and explanation unless policies genuinely evolve and validate.
The full-resolution held-out gate failed, so learned weights stay explanation-only for now.
Next Move
The right technical direction
Diagnose scale-sensitive failures
The next build should inspect the hazard-crossing candidate rows at full resolution, add normalized/risk-path features, and retrain only after the failure is understood.
Hardware vs Ender Dragon
The dream is a personal AI that can someday learn its way through Minecraft with honest visible play. The discipline is refusing to skip the evidence ladder that would make that story real.