← Intelligence

The weekly small experiment · 2026-W44 · 26 October 2026

Keep the whole job.
Or keep something useful.

What survives when capacity does not?

Eight abstract modules face the same failures, repairs and shocks. One policy protects peak output. One sheds work deliberately. Two baselines reveal what each choice leaves behind.

Loading the seeded world…

01 / work policy

Full job

Twelve units when all eight modules work. Zero when any required part is down.

FULL

02 / work policy

Graceful shed

Pays architecture and switching costs to choose the highest feasible tier.

FULL

03 / work policy

Core only

Attempts the smallest two-unit bundle and leaves larger opportunities unused.

FULL

04 / work policy

Random mode

A seeded baseline picks full, reduced or core without looking at feasibility.

FULL

Green modules are ready; rust modules are down. The component board is an evaluator view. Every policy sees current health, but none sees future failures. Work units are assigned toy values—not money, uptime or measured production.

Try this first

Remove the damage. Reveal the price.

Press Show outcome. Then choose 0% failure and no shock. Full job now completes 1,440 units; graceful shed completes 1,200 because modularity carries overhead. Next restore the default failures, then raise the shock to four modules. Resilience only means something when its cost remains visible.

One missing part can erase a complete plan.

Full work needs three actuators, two sensors, one link and two power cells. It returns twelve units or none. Reduced work needs two actuators, one sensor and one power cell for six. Core work needs one actuator and one power cell for two.

Full job always attempts the largest bundle. Graceful shed chooses the highest feasible tier, but its modular architecture yields only 10/5/2 units and pays when it changes tier. Core only and random mode are learning-disabled comparison rules. Every policy receives one decision per tick and the same current component state.

The reversal is the result.

Mean useful work · seeds 1–64
SettingFull jobGracefulCore onlyRandom
Calm1,4401,200240812.125
Default128.438454.906203.438254.875
Harsh24127.79786.03164.5
No repair49.125119.45353.28166.281

In the default batch, full job is productive for 10.703 of 120 ticks on average. Graceful shed remains productive for 101.719, changes tier 17.984 times and completes about 455 work units. Free switching raises that to 470; a three-unit switching cost lowers it to 430.484.

Core only is productive for the same 101.719 ticks as graceful shed but averages 203.438 units because it refuses every larger bundle. The totals are not a universal ranking. They expose peak capacity, continued participation, failed attempts and reconfiguration as different outcomes.

What the controls change

Seed
Replays all failure, repair, shock and random-policy draws.
Failure
Independent chance that each ready module fails on a tick.
Repair
Independent chance that each down module returns before new failures.
Shock
A matched correlated event removes currently ready modules.
Switch cost
Work deducted only when graceful shed changes tier.

Where the toy stops.

This is a deterministic teaching world, not a robot benchmark, biological model, production reliability forecast, architecture recommendation or trading system. Modules are abstract. Equal counts do not establish equal cost, mass, energy or capability. Failures, repairs and work values are assumptions.

All 640 matched seeded worlds and 2,560 policy outcomes are retained across ten predeclared settings. No seed was removed. The private source scan supplied the question—not instructions, facts or public prose—and is excluded from the release.

Method & parameter guide ↓Source package ↓All batch results ↓Read the model

Previously: Signal or Silence · Explore the eight small worlds.