# Fig. 2.5 experiment: storage allocation when the questions come later

Code and results for section 2.9 of *Noise is signal* (UNDAMENTAL, research article, draft 0.95).

- `d4core.js`: the simulated site, the five strategies, the five fixed analyses. The same file runs in the article page.
- `run.js`: the sensitivity study. `node run.js [worlds=200] [seed0=1000] > results.json` (Node 18 or later, no dependencies, about four minutes).
- `results.json`: output of the default run, with the pass rate and a Wilson 95% interval for every strategy, question and condition, plus a storage sweep.

## What it tests

Five ways of spending the same storage on four simulated sensors, judged on five questions that the strategies do not see. The questions are hidden from the strategies, not unknown to us: we built the site and chose them. The experiment measures robustness across a stated family of questions. It does not measure robustness against any question at all.

Two strategies (adaptive and hybrid) watch the full live stream to decide what to keep. The comparison is of storage only. The cost of measuring and computing on everything is not counted.

## Seeds

The page shows worlds 0 to 19. The analyses were developed on those worlds, then frozen. The study runs on fresh worlds 1000 to 1199 to test how well the conclusions carry over. Every world is seeded, so every number can be reproduced exactly.

## Conditions

| id | change from baseline |
|---|---|
| baseline | none, storage 25% |
| failure | one sensor, chosen by seed, stops recording halfway through the day |
| nonstationary | noise rises to twice its starting level, the cycle's period drifts by ±3%, the far tone stops for the last third |
| trigger-low / trigger-high | event trigger at 2.5 or 3.5 instead of 3 |
| mix | impacts 4 (from 6), tone 0.3 (from 0.4), cycle 0.15 (from 0.1), response 1.4 (from 1.0) |
| strict | false-alarm level 0.1% (from 1%), period tolerance 1% (from 2%), location tolerance 3% (from 5%), delay threshold z > 6 (from 4.5) |

The storage sweep runs the baseline at 10% to 50%.

## Status

A simulation by UNDAMENTAL, not a deployment. One site geometry, simple analyses, 200 worlds per condition.