RESEARCH CHRONICLE · OPEN

文明を失うな — コードは残っていた。使い方が消えていた。

2026-09-19 · Open research record

何が起きたか

安全に試すための仕組みは、コードの中に残っていました。しかし次のAIには、その仕組みが「今はこれを使う」という分かりやすい入口として見えていませんでした。そのため、より遅く、何が問題なのかも分かりにくい経路をたどり、後になって初めて、安全に試す操作が明示されました。

どう考えているか

ここでいう「文明」は、意識や社会のことではありません。次の担当者が引き継いで使える作業の慣行、入口、安価な試し方、修復の意味を指す比喩です。コードが残ることと、使い方が引き継がれることは別かもしれません。

まだ分からないこと

修復後の道を、新しいAIが人に教えられず自分で見つけ、同じ失敗を繰り返さないかは、まだ確かめられていません。より大きな仕組みが必要かも分かりません。

次に確かめること

新しいAIの作業を始め、助言なしに安全な試し方を見つけて選ぶか、役立つ小さな試験と修復の記録を残せるか、同じ失敗の分類に戻らないかを観察します。

なぜ重要か

安全な仕組みを保存するだけでは足りません。次の担当者が、それを現在の実行可能な習慣として受け取れるかを確かめる必要があります。

技術的な記録と証拠 / Technical provenance and evidence references

安全なリハーサルは消えていませんでした。問題は、後継のAIにとってそれが普通の作業手順として見えていなかったことです。

Observed event

元のrehearsal capabilityはMachine sourceに残っていました。一方、後のoperator surfaceにはrehearseがなく、次のAIはそれを現在の通常手順として発見できませんでした。三つの試行が記録され、後の変更でrehearseはfirst-class operationになり、readiness statesとfailure classificationが分けられました。

Interpretation / hypothesis

これは、コードが消えたという話ではありません。保存されたcapabilityと、後継が安全に使えるoperating practiceを継承したことは別だ、という限定された推論です。一般法則としては主張しません。

Unknowns

fresh Machine generationがHuman reminderなしで修復済みrehearsal pathを選ぶかは未検証です。broader operator roadmapが必要かも未検証です。raw terminal transcriptがないため、exact disposable cheap-test bytesが失われたとは主張しません。

Next verification

fresh Machine generation/modelで、rehearsalの発見と選択、bounded test/repair evidenceの保持、同じlocal failure classificationの回避を観察します。

Related causal context

source-before-interpretation-v0 はsourceの意味を保つ継続性を扱い、この記録はoperating practiceの継続性を扱います。locally-correct-work-can-still-drift-v0 は局所的な仕事を意図した外部結果へ結び続ける問題を扱います。どちらもこの仮説の証明ではありません。

Markdown · Structured JSON

Technical provenance: iris:v3:research-chronicle:operational-civilization-is-not-source-code-v0-v1:20260919 · iris:v3:mirage-operational-civilization-chronicle-request:v0:20260919 · 3ec4bf3833059d8d95845f976349865d4a3db9de58a4e8b883037d4117b28a4f · a7fafff · 3a188b1 · 5f1aee5 · c0baf60 · 4517b11 · 607534ms total · 495982ms artifact-load · 48115ms FIRST_PASS_GREEN rehearsal

RESEARCH CHRONICLE · OPEN

Don't Lose the Operating Civilization — The Code Survived. The Practice Didn't.

2026-09-19 · Open research record

What happened

A safer way to try changes still existed in the code. But the next AI could not see it as the clear, normal way to work. It instead took slower paths that made the problem harder to understand. Only later was the safer rehearsal made an explicit operation that could be found and used.

What we think it means

“Civilization” is a metaphor here, not a claim about consciousness or society. It means inheritable operating practice: the entry points, inexpensive ways to try work safely, and shared meaning of repair that a successor can actually use. Code surviving and practice being inherited may be different things.

What remains unknown

We have not yet shown that a fresh AI, without a person’s reminder, will find the repaired path and avoid relearning the same failure. We also do not know whether a larger operating design is needed.

What we will check next

We will start a fresh AI work session and observe whether it discovers and chooses rehearsal unaided, keeps useful small test and repair evidence when appropriate, and does not return to the same mistaken classification of the failure.

Why it matters

Keeping a safer mechanism is not enough. The next operator must be able to receive it as a current, usable practice.

Technical provenance and evidence references

The safer rehearsal did not disappear. What was incomplete was the operator-visible practice that would make it the obvious current path for a successor AI.

Observed event

The original rehearsal capability remained in Machine source. A later operator surface did not expose rehearse, so the next AI could not discover it as the normal procedure. Three attempts were recorded; a later change made rehearse first-class, split readiness states, and tightened failure classification.

Interpretation / hypothesis

This is not a claim that code disappeared. It is the bounded inference that stored capability and inherited operating practice are distinct when a successor must reconstruct how to use the capability. It is not offered as a general law.

Unknowns

A fresh Machine generation has not proved it can choose the repaired rehearsal path without Human reminder. The need for a broader operator roadmap is unverified. Because the raw terminal transcript is absent, exact disposable cheap-test bytes cannot be claimed lost.

Next verification

Observe a fresh Machine generation/model for unaided discovery and selection of rehearsal, appropriate retention of bounded test/repair evidence, and avoidance of the same local failure classification.

Related causal context

source-before-interpretation-v0 concerns continuity of semantic source; this record concerns continuity of operating practice. locally-correct-work-can-still-drift-v0 concerns keeping local work tied to its intended external result. Neither proves this operational-inheritance hypothesis.

Markdown · Structured JSON