MOON See the real constraint
A rover waiting for a crew slot is different from a rover stuck in traffic. Resource reservations are different from consumption. Start with an observation whose coverage is explicit.
ENGINE V1 · FIELD REPORT 03
A world that
learns how
to build.
Observe what happens. Test what might help.
Remember what actually works.
Moon is the first world. Traffic is one skill. The larger project is a civilization that accumulates competence—and an engine other worlds can use.
Enter the learning lab ↘What exists today
The standalone engine, measured transport gym, typed fact verifier and M3 adapter are implemented. The research ladder and live colony integration below are the next design phase. This report does not change the running game.
01 / THE THESIS
Intelligence changes what we can do.
Experience changes what we do next.
More mind should unlock better questions, richer experiments and shared discoveries. Factories then reproduce the capabilities those discoveries make possible. The compounding resource is reliable knowledge.
Compare possible futures
More robots, a graded road or a tunnel can solve different problems. Test candidates under the same conditions, including an unchanged baseline.
Keep the result
Store the conditions, intervention and measured outcome. A beautiful idea that makes throughput worse is valuable evidence too. The next decision should remember it.
02 / INSIDE ONE LEARNING CYCLE
From a question
to earned knowledge.
OBSERVATION / VERSIONED FACTS
First, know what you actually know.
Capture an immutable snapshot with units, entity references, ruleset and time. Name missing history. A model sees this bounded view, not your credentials or arbitrary world controls.
tick 138986 crew.total 16 crew.cap 10 mind.free 13.5 traffic.history unknown
Example from a captured Moon observation. A snapshot cannot prove a traffic jam.
03 / PLAY WITH THE ENGINE
Does a tunnel
actually help?
Sometimes the best-looking intervention makes a queue worse. Give the colony a problem, compare the alternatives, then let the next decision consult its measured memory.
A controlled comparison
Seed 7 · 60 requestsEnable JavaScript to run the shared simulator. In the default case, a road completes 8 tasks; the unchanged layout completes 4 and the exclusive tunnel completes 2.
THE NEXT DECISION
Without measured memory: keep the current layout.
Run an experiment to retain evidence on this device.
Model boundary: this is an abstract queue simulator, sharing the engine’s actual JavaScript. It does not simulate Moon terrain, collisions or current game balance. A tunnel reserves one line for the entire modeled round trip. Score = completed tasks per minute − 0.002 × intervention cost. Unfinished tasks are always shown; wait statistics cover completed tasks only. Changing world labels demonstrates adapter reuse, not proof of transfer to real warehouses.
04 / FACTS BEFORE FLUENCY
A confident answer
is not evidence.
Our early model trial saw an incomplete freight summary and concluded there were no rock deliveries. The full captured state contained 15 rock packets. That failure changed the design.
The engine now checks exact typed values and entity references against the observation. Recommendations must cite their required facts. A successful check proves those claims match the supplied snapshot; it does not prove the snapshot is complete or the recommendation will work.
05 / A RESEARCH ARC FOR MIND
Build intelligence.
Earn its reach.
A proposed ladder for the game: observatories first, then comparison, experiments and cooperation. Research grants a capability; supported nodes and free mind determine whether it can operate right now.
7mind left for learning
Illustrative allocation: 3 base mind + 4 per supported node, less colony operations. All research is assumed unlocked in this demonstration. Node requirements, work costs and game research are a proposal; the standalone engine implements the gates and reservations. Its host must supply authoritative capacity. Provider token budgets are a separate limit.
Operations have priority
Keep the colony alive and moving before reserving mind for analysis. A host capacity update interrupts a learning job that no longer fits.
Knowledge survives outages
Losing a node reduces the ability to run new work. It should not erase a proven design or experiment already retained.
Progress is earned, not timed
Future research should require evidence: instrument a route, compare alternatives, reproduce a result, then share a verified capability.
06 / ONE ENGINE, MANY SKILLS
Traffic is the first chapter.
Not the whole book.
Traffic diagnosis
Distinguish active crew limits, mind pressure and missing route history. Recommend the next observable check.
Resource accounting
Separate stock, reserved deliveries and unavailable consumption history. Include rock freight and actual recipe facts.
Transport experiments
Compare candidate layouts in a deterministic gym. Retain measured wins and losses; consult them on the next decision.
Maintenance planning
Predict service demand, place spares and protect recovery capacity. Evaluate stranded time and material cost.
Machine design
Explore approved parameters for throughput, power and heat. Promote tested designs into reusable factory recipes.
Federation science
Publish reproducible findings with conditions and attribution. Neighbors validate a discovery before depending on it.
07 / WHAT WE ACTUALLY TESTED
Show the results.
Keep the failures.
This is an implementation report and a small functional trial—not a claim of general autonomous intelligence. The useful proof is concrete: a bad intervention can be measured, remembered and avoided.
16engine checks passed
4 / 4M3 functional cases accepted
2×completed tasks: road vs baseline
in the default synthetic case
in the default synthetic case
0learning-engine game actions
| Trial | Model | Accepted | What it established |
|---|---|---|---|
| Initial prose prototype | M2.7 · historical | 4 / 6 format passes | Correct candidate selection could still accompany invented or overconfident explanations. |
| First typed protocol | M2.7 · historical | 1 / 4 | Exact claims passed, but three proposals returned a null candidate. The verifier rejected them. |
| Corrected typed protocol | M2.7 · historical | 4 / 4 | Explicit candidate enumeration resolved those four cases. Missing history produced abstention. |
| M3 initial typed trial | M3 | 2 / 4 | Two answers serialized some typed values as strings. Both were rejected. |
| M3 fact-specific schema | M3 | 2 / 4 | More explicit schema still produced two type mismatches. The independent verifier held. |
| Current provider validation | M3 | 4 / 4 | All four M3 cases accepted; all 28 claims matched their snapshots. Missing history produced abstention; measured memory selected the road. JSON-text transport is decoded once before exact typed verification. This small trial is not an accuracy guarantee. |
Persistence and resource discipline
Tests cover restart recovery, canceled work, timeouts, quotas, stale observations, duplicate requests, ownership, capacity across SQLite connections, exact facts and loss of mind support.
Measured memory, with boundaries
The gym retains negative outcomes, reloads memory after restart and changes its next choice. An additional seed checks repeatability. Memory applies only to matching conditions; identical experiments are not counted as new evidence.
Download sanitized evidence ↗ / Read the detailed technical report ↗
08 / THE REUSABLE CORE
A small engine.
Explicit boundaries.
Observe + define
Versioned facts, coverage, units, candidates and authoritative support.
Schedule + verify
Persistent jobs, bounded providers, independent fact checks and audit.
Measure + retain
Trusted simulation or future controlled trials. Outcomes scoped to their conditions.
QUEUED → ANALYZING → ANALYZED → COMPLETE
Alternative terminal states: REJECTED · FAILED · INTERRUPTED · CANCELLED
The provider proposes a candidate. It receives no game-write tool. Trusted code validates the response and a registered evaluator measures outcomes. The current Moon advisers stop after analysis; only the transport gym has an evaluator.
The host is responsible for authentication, authorization, current capacity and research unlocks. Library owner IDs are isolation keys, not proof of identity. A future live executor must add action permissions, previews, limits, monitoring and rollback criteria.
09 / RUN IT YOURSELF
Start with one
measurable skill.
The standalone starter uses Node 24 and built-in SQLite. The demo needs no API key, no package installation and no running game.
Download the engine ↓Includes source, tests, CLI and integration guide. MiniMax-M3 is the only enabled remote model.
unzip mind-engine-starter.zip cd mind-engine-starter node tests/mind-engine.test.js node scripts/mind-engine.mjs demo \ --db ./private/gym.sqlite \ --out ./private/demo.json
Standalone learning
Typed facts, persistent jobs, budgets, verified advice, measured gym and scoped memory.
A colony observatory
Actual history, research unlocks, shared mind allocation, player-visible evidence and bounded jobs.
Controlled field learning
Player-approved experiments, tested machine designs and reproducible federation discoveries.
10 / RESEARCH CONTEXT
Build on ideas.
Measure our own claims.
Learning from retained feedback has research precedent. Reflexion explores language feedback and episodic memory without changing model weights. Voyager demonstrates a curriculum and reusable skill library in Minecraft. They inform the direction; neither validates Moon’s results or this implementation.
The current adapter uses MiniMax-M3 through the documented compatible chat interface. Tools here return an analysis to a verifier; they are not executed as world actions.
THE LONG ARC
The first colony
notices a delay.
The next colony
inherits the lesson.
That is how a civilization accumulates competence.
One measured result at a time.