Simulation Log

Small simulations written to probe design questions. Each is published with its code and output, and with a reading of what it can and cannot show.

PurposeShow the lab’s probing work as it is, including tests that turned out to show less than their titles suggest.

StatusExploratory — none of these tests the four-channel architecture, and none has been reviewed by an outside party.

How to read itThese are simulations of simple baseline-tracking logic, written to expose design questions. They are separate from the contamination benchmark and from isolation under adversarial pressure, which run against the architecture’s own code.

AttributionCode AI-generated (tool not recorded).

StandardEvery entry states the question, the code’s own claim, what the data show, and what is not shown. Results that overstate are labeled as such rather than removed. Files submitted more than once are archived once.

What they led toThe questions raised here (a floor set by an early noise peak, a floor rule that false-alarms on stable data, a misattribution test with no real execution noise) are the useful output. They are design questions for the real system.

The log

Determined by setup Adversarial Scenarios Forced max scaffolding, a wrong prerequisite graph, a malicious verifier. Each headline number comes from a value the script sets. Does not show harm Genesis Poisoning A floor locked early never updates. Three scripts confirm it stays low; the formula means a low floor cannot lower the baseline. Partly circular Boiling Frog A rolling mean follows slow drift. A check added here finds the running-maximum floor would false-alarm on stable data. Cannot fail Compound Attacks Both scripts print success on conditions the setup guarantees; mastery does not depend on accuracy. Cannot show improvement Environment Noise The new rule blames the C channel for network trouble half as often, but the sample has no real execution noise. Cannot fail Combined Stress Mastery follows a fixed formula that the stressors do not touch; the “remained functional” line is printed unconditionally.

What counts here

The earlier verification pages report tests whose result could have come out the other way. These could not have, and say so. More simulations will be added as they are run, each with the same four-part reading. A test is not listed as support for the architecture unless it acts on the architecture’s own code and states its pass condition in advance.