Field study 06 / 06 · Minimum Intelligence
How much mind
does a swarm need?
One planner, many local decisions, and a fleet of stubborn little rules. The interesting frontier is useful work per scarce resource.
一个群体，究竟需要多少智能？
Loading the model. Recorded evidence remains available below.
Delivered salvage value / units · higher is better
| Treatment | At this tick | At mission end |
|---|
01 / Try to break it
Replace intelligence one body at a time.
Use the same bodies and movement rules. Replace 100 reasoning controllers with 50, 10, 1, then zero. Keep the information package visible. A model label is not evidence: inspect the recorded decisions, transport receipts and unavailable measurements.
How much model planning earns its place when physical work, information and complexity all count?
02 / The mechanism
What the little world actually does.
A bounded mission interface separates observation, decision and movement. Decision tapes can replay actual model choices without a browser API key. A rule baseline runs in the same physical environment. Calls, tokens, elapsed time and plan validity are counted separately from simulated ticks.
An architecture comparison and a controlled model comparison answer different questions. A central planner may pool observations; a local agent may not. Synthetic sweeps describe the simulator. Only a recorded, identified model response counts as a model measurement.
统一观测、决策与行动接口，逐步把 100 个推理控制器替换成 50、10、1、0 个。模型决策必须有真实记录，调用次数、令牌、耗时与物理时间分别计算。模拟规则不能冒充 Sol 或 Luna 的实测。
03 / Recorded evidence
Keep the inconvenient outcomes.
Ten sector Luna planners retained 98.2–100.9% of the value of 100 local Luna planners across three worlds, using about 10.2% of their input tokens. Sector planners also pooled more information. Both declared superiority screens failed; the smallest tested local replacement subset retaining 95% in every world was 50 Luna cards. This is a one-epoch pilot, with actual dollars and equal-deadline superiority unknown.
A sparse use of model decisions might preserve most of the full local-model team’s value. First test whether zero models already does so.
Paired seeded simulations. Every retained result can be regenerated from the versioned source; these are synthetic experiments, not field validation.
Inspect every world and treatment
| World / treatment | Mean units | n |
|---|---|---|
| Original mission / 100 local rules | 2,530.12 | 3 |
| Original mission / Pooled assignment rule | 2,520.07 | 3 |
| Original mission / 1 Sol · pooled map | 2,563.34 | 3 |
| Original mission / 10 Luna · sectors | 2,722.42 | 3 |
| Original mission / 100 Luna · local | 2,736.45 | 3 |
| Original mission / 50 Luna + 50 rules | 2,687.03 | 3 |
| Original mission / 10 Luna + 90 rules | 2,570.81 | 3 |
| Original mission / 1 Luna + 99 rules | 2,532.69 | 3 |
| Noisy motion / 100 local rules | 2,493.67 | 3 |
| Noisy motion / Pooled assignment rule | 2,494.81 | 3 |
| Noisy motion / 1 Sol · pooled map | 2,594.55 | 3 |
| Noisy motion / 10 Luna · sectors | 2,724.1 | 3 |
| Noisy motion / 100 Luna · local | 2,725.31 | 3 |
| Noisy motion / 50 Luna + 50 rules | 2,667.11 | 3 |
| Noisy motion / 10 Luna + 90 rules | 2,512.49 | 3 |
| Noisy motion / 1 Luna + 99 rules | 2,501.23 | 3 |
| Salvage quake / 100 local rules | 2,473.36 | 3 |
| Salvage quake / Pooled assignment rule | 2,432.69 | 3 |
| Salvage quake / 1 Sol · pooled map | 2,513.1 | 3 |
| Salvage quake / 10 Luna · sectors | 2,667.63 | 3 |
| Salvage quake / 100 Luna · local | 2,687.14 | 3 |
| Salvage quake / 50 Luna + 50 rules | 2,613.73 | 3 |
| Salvage quake / 10 Luna + 90 rules | 2,513.21 | 3 |
| Salvage quake / 1 Luna + 99 rules | 2,475.82 | 3 |
| One loading bay / 100 local rules | 2,518.85 | 3 |
| One loading bay / Pooled assignment rule | 2,473.63 | 3 |
| One loading bay / 1 Sol · pooled map | 2,511.66 | 3 |
| One loading bay / 10 Luna · sectors | 2,701.06 | 3 |
| One loading bay / 100 Luna · local | 2,765.76 | 3 |
| One loading bay / 50 Luna + 50 rules | 2,657.05 | 3 |
| One loading bay / 10 Luna + 90 rules | 2,529.39 | 3 |
| One loading bay / 1 Luna + 99 rules | 2,521.63 | 3 |
One chosen seed is an illustration. The downloadable batch retains every world and treatment; null means unavailable. A confidence interval describes the sampled synthetic worlds, not the real world.
04 / Inspect the boundary
Enough detail to disagree.
Inputs, budgets and observation boundaries
Every measure in the current run
- A genuine three-seed, one-planning-epoch model pilot, replayed in a synthetic physical world. Not continuous replanning or online adaptation.
- Sol sees pooled observations; sector planners see sector pools; independent Luna sees local readings. Architecture and information differ.
- Mixed teams reuse recorded individual Luna decisions; these are ablations, not new independent model trials.
- Actual subscription dollar cost and immutable model snapshot are unavailable. No equal-dollar or equal-deadline superiority claim.
- The animation is an illustrative 3D projection of a two-dimensional salvage model.
3D sketching adds depth to the explanation. It does not add an unmodeled physical dimension, extra sensor data or unearned intelligence to the experiment.