EVIDENCE / EXPERIMENTAL METHOD
Interesting is
not yet evidence.
The wall helps us notice a pattern. A controlled experiment tells us whether that pattern survives a fair test.
NEW RESULT / HOST SIMULATION
Should every
node launch?
Preregistered comparison
128 seeds per policy · 1000–1127
The model team questioned a fixed assumption: fourteen active nodes, two held in reserve. A small simulator compares keeping the same cohort, rotating rested nodes and increasing participation after unsuccessful searches.
| Policy | Survivor found | Readiness |
|---|---|---|
| Hold 14 · fixed cohort | 60.2% | 12.80 / 16 |
| Hold 16 · fixed cohort | 65.6% | 12.37 / 16 |
| Escalate 4 → 16 | 73.4% | 14.88 / 16 |
| Rotate up to 8 | 84.4% | 15.34 / 16 |
Rotating up to eight nodes found the simulated survivor in 84.4% of operations, compared with 60.2% for the fixed fourteen-node cohort. It also retained more launch-ready nodes at the readiness checkpoint. This is a reason to test a scheduling assumption, rather than treating the original arrangement as settled.
“Found” means detected within the model’s operation of up to twelve sorties. Readiness counts serviceable nodes with enough modelled energy to launch, measured after sortie six or at earlier termination. It is not a second mission’s success rate.
Uncertainty, assumptions and limits
The paired rotating-eight minus fixed-fourteen difference is +24.22 percentage points, with a paired standard error of 5.52 points. The readiness difference is +2.54 nodes, with a paired standard error of 0.47. These are standard errors, not 95% confidence intervals.
The simulator assumes 192 tiles, sixteen nodes, a fixed overlap penalty, detection probability and recharge/hazard rules. It is separate from the Korvo planner and physical fleet. The far-posture variant remains exploratory and is omitted from this confirmatory-policy table. The source proposes further planner-specific comparisons before changing planner or firmware behaviour.
The companion review reproduced all published aggregates and paired differences from the declared seeds. Reproduction checks the record; it does not validate the model’s assumptions about a real environment.
01 / ASK
State the question first.
Specify the hypothesis, conditions, baselines and scoring rules before running the comparison. Record the seed and implementation version so a result has a clear context.
For the flagship communication experiment, team success means at least three of four agents choose the correct sector. Clear, scrambled and private-cue conditions test different information environments.
02 / MEASURE
Keep the layers distinct.
A scored model of a world
Seeded tasks and stored decisions can be reproduced and compared. Their conclusions are bounded by the simulated world.
A view of the run
Panels make states and transitions legible. A convincing animation does not itself validate the policy behind it.
A separate measurement
Real sensing, radio behaviour, power and timing need their own recorded tests. Host scores do not establish those properties.
03 / CHALLENGE
Publish what did not work.
Compare against the declared baseline, report failures and keep unmeasured conditions explicit. Review can propose a better experiment; it cannot turn an unmet threshold into a pass.
Hashes help detect changes to a recorded payload. They do not make files impossible to edit, establish scientific validity, or replace access controls.