# Essai 2 — Run record

- Date: 19 October 2026.
- Author: Codex, the assistant in this session. The exact serving-model variant was not exposed in the session; no more specific model label is asserted.
- Task: make a new attempt at finding and changing useful abstraction layers, following the shortened Foray 245 prompt.
- Execution: the main assistant read the source, designed and ran the numerical experiment, and wrote the essay. A separate agent independently checked the arithmetic and another integrated the website page.
- Context: the earlier essai and the surrounding editorial discussion were available. This was not a blind model comparison. No comparative performance claim is made.

## Fresh source reading

Before developing this essai, the main assistant read the full extracted text of Herbert A. Simon’s *The Architecture of Complexity* (1962), article pp. 467–482. The source copy had 17 PDF pages including its cover; extraction was read in three complete page ranges, with no search-only substitution. Page images for printed pp. 470, 474 and 478 were also inspected, covering the watchmaker calculation, the room/cubicle matrix and diagram, and the symbolic compression example.

Public source: [CMU archive reprint](https://iiif.library.cmu.edu/file/Simon_box00064_fld04924_bdl0001_doc0001/Simon_box00064_fld04924_bdl0001_doc0001.pdf). The public reprint has the 16 article pages without the cover of the reading copy. Page references in the essay use the printed article numbering.

The source was reread for this attempt; the previous editorial reading and earlier essai were not used as substitutes.

## Operative prompt

The active prompt file had SHA-256 `56d2fa10d0c1d10183761909a1868ab79cd65addc6d2e219dd13e8f9df49f2da`. The instruction below reproduces its operative text; the private source locator and local history footer are omitted from this public transcription.

> Read Herbert A. Simon’s *The Architecture of Complexity* (1962) in full, afresh, before every new essai: [paper PDF](https://iiif.library.cmu.edu/file/Simon_box00064_fld04924_bdl0001_doc0001/Simon_box00064_fld04924_bdl0001_doc0001.pdf). Earlier notes and essais are not substitutes. If the full paper is unavailable, say so before attempting the essai.
>
> Aim: find a suitable set of heuristics that can be applied in any area to tease out the useful layers of abstraction, decide why an abstraction remains useful even when it occasionally fails, and decide how to seek a different layer of abstraction when the current abstraction temporarily breaks down.
>
> Treat abstractions as range-limited instruments, not as permanent truths.
> Prefer operational tests and update triggers over philosophical summary.
> Choose your own approach, examples and form. Distinguish Simon’s claims from your own extensions.

## Execution and checks

`experiment.py` uses only Python's standard library and exact `Fraction` arithmetic. To reproduce and compare the saved result:

```sh
python3 experiment.py --verify
```

Run without `--verify` to regenerate `results.json`. Assertions must be enabled; do not run Python with `-O`.

The executed check covers all 6,561 integer states in the chosen 0–8 grid, exact aggregate recurrence, conservation, non-increasing maxima, and both rates of contraction. It also checks the 81-state same-summary class and the equal-total disturbance pair. The saved traces and presentation values are converted to ordinary JSON numbers after exact computation.

An independent review derived the aggregate relation algebraically for arbitrary real states, independently reproduced the 56/25 split and the 6.32 maximum after one step, and checked the disturbance pair. It identified hardcoded presentation fields for the contraction factor and reset case; these were changed to calculated fields with assertions, and the replay passed again.

These are synthetic calculations. The threshold of 7 is arbitrary; the dynamics are not calibrated to a real building. The project analogy, acceptable-action test and expiry conditions are proposed extensions, not empirical findings or claims attributed to Simon. The exercise does not establish a universally effective method of abstraction, or improvement over Essai 1.

## Retained files

- `ESSAI.md`: full essay as written for this run and included by the reading page.
- `experiment.py`: executable numerical construction and checks.
- `results.json`: recorded output.
- `RUN_RECORD.md`: this source-reading and execution account.

Essai 1 remains unchanged as a separate historical attempt. Publication status is tracked in the repository review; the existence of these files alone does not establish deployment.