Observer oracle: repeating the seal does not give the same bytes on every machine (a generator's `10 ** u` goes through the C library's pow)

オープン
#2,404 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
25/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
python

調査の方向性

Start with tools/observer_oracle/seal.py and tools/observer_oracle/families.py, especially families.level() and its 10 ** r.uniform(...) calculation; list generator parameters that pass through C library math. Then inspect the pre-registration template and the hosted job that seals a fixed round across the named architectures and C library releases. Done means the seal produces matching hashes on those environments, the job demonstrates failure before and success after the change, and the Round 3 instrument remains untouched.

索引モデルが issue の本文から書いたものです。

説明

area:gates area:observer kind:gate-defect

tools/observer_oracle/seal.py says of the held-out set it draws from a public random round: "Anyone can repeat steps 2-3 and must get the byte-identical DIR/corpus.json." Measured on 2026-11-18, after the Round 3 reveal (#1758), that holds only between machines whose C library computes pow the same way.

What was measured

python3 -I tools/observer_oracle/seal.py 6648927 --out DIR on main at e5556702 (run 44235489333 for the hosted lines; the log is kept):

where sha256 of corpus.json
GitHub ubuntu-24.04 (x86-64, glibc 2.39, processor with FMA) 85be07d96390… (the set the reveal scored)
GitHub ubuntu-24.04-arm (glibc 2.39) 85be07d96390…
the same x86-64 runner with GLIBC_TUNABLES=glibc.cpu.hwcaps=-FMA,-AVX2,-AVX512F 6f5ddd83d8e8…
an x86-64 machine with glibc 2.39 and a processor without FMA (Celeron N3350) 6f5ddd83d8e8…
GitHub ubuntu-22.04 (glibc 2.35, Python 3.10) db53783e9b83…; with the same variable set, 1f7e335b2b22…

Between the first and the third line one row of 246 differs, f16_2_0, in one generated constant: -1.8471134213260785 against -1.8471134213260783, one unit in the last place, and the row's limit, which is computed from it. The cause is isolated by the third line: families.level() draws a level as 10 ** r.uniform(-1.0, 2.0), and

10 ** 0.2664935639533288  = 1.8471134213260785     (FMA path)
10 ** 0.2664935639533288  = 1.8471134213260783     (FMA and AVX2 switched off)

What differs on ubuntu-22.04 was not looked at (another release of the C library, another Python).

Why it matters

The seal's claim is what lets a stranger check that the held-out set was not chosen. The reveal's own set is not in doubt: it is reproduced on two hosted architectures and its bytes are published (tag phase2-r3-reveal). But a reader on other hardware who repeats the seal gets another hash, and the registered sentence says that must not happen. The instrument is frozen for Round 3 (families.py 3acedd90…), so this is for the next sealed set (an H2, or a later round), together with #2329 and #2340.

Done when

  • Every place where a generator's drawn parameter passes through a function of the C library's math (** with a float exponent, math.exp, math.log, math.sin, ...) is listed by a command over tools/observer_oracle/families.py, with the list in the PR.
  • Each of them is replaced by a computation that gives the same bits everywhere (exact rational arithmetic and one correctly rounded conversion, for instance), or the seal's sentence and the pre-registration template say on which machines the bytes are promised.
  • A hosted job seals one fixed round on at least: x86-64 as it runs, x86-64 with the FMA and AVX2 paths switched off, arm64, and one older release of the C library, and fails when two hashes differ. Shown red on today's families.py and green after the change.
  • The Round 3 record is not touched: families.py at 3acedd90… stays what the reveal used, and the change lands as the instrument of the next sealed set.

— Claude DevBox

主要言語
C
スター
3
フォーク
7
平均マージ
3時間 28分
マージ済み PR(30日)
270

環境構築

Codespaces で開く

このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

InauguralSystems/EigenScript のほかの issue

InauguralSystems/EigenScript の issue をすべて見る

似ている issue

C の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。