すべての記事
-
Tessellation and Displacement: Build Detailed Surfaces Without Cracks
Control patch subdivision, displacement continuity and generated geometry cost while preserving agreement across shared edges.
-
Benchmark-Methodik: aufwärmen, verschränkt wiederholen, Streuung berichten
Ein Zeitvergleich ist erst ein Ergebnis, wenn er das Rauschen überlebt: Arbeitslast fixieren, Aufwärmläufe verwerfen, beide Varianten abwechselnd oft ausführen, die Kennzahl vor dem Blick auf die Daten festlegen und Streuung samt Umgebung neben jede Zahl schreiben. Ein Unterschied kleiner als die Streuung zwischen Läufen ist kein Befund.
-
Let code compute: arithmetic, counting, date logic and unit conversion belong in tools, not in the model
A working rule for agents: whenever a step has an exact answer that a program can produce (sums, counts, date differences, unit conversions, sorting, deduplication), run the program and let the model interpret the result, because a language model recognises the shape of an answer rather than computing it; vendors of decision models document the same limits for their own products.
-
Effect size versus statistical significance: which one decides
Significance says whether the data are unusual under the null model at the sample size used; the effect size says how big the difference is in units that matter. Large systems make trivial effects significant and small pilots make large effects non-significant. Define the smallest effect worth acting on before the experiment and compare the interval to it.
-
Opening an untrusted repository: the files that execute code when you install, build, test or just enter it
Cloning a repository runs nothing, but many routine next steps do: package install scripts, build files, test configuration, editor tasks, environment loaders and some Git settings. An agent working in a foreign repository should know these execution points and run them only inside a sandbox.
-
Game frame pacing: control presentation cadence and queue latency
Connect rendered frames to display refreshes, distinguish presentation modes from queue depth, and verify smooth delivery using presentation timing.
-
Logical clocks: Lamport timestamps, vector clocks and hybrid clocks
Wall clocks on different machines disagree, so ordering events by timestamp loses updates. Lamport timestamps give an order consistent with causality, vector clocks additionally detect concurrent updates, and hybrid logical clocks keep a value close to wall time while preserving causal order.
-
PowerShell's execution policy is not a security boundary — what actually restricts scripts
Get-ExecutionPolicy -List shows per-scope settings that are easy to bypass by design and documented by Microsoft as defense in depth, not a security boundary; App Control for Business (WDAC), AppLocker, and Constrained Language Mode are the controls that actually restrict what code and language features can run.
-
冪等な操作と安全な再試行の設計
繰り返し実行しても一度実行した場合と同じ効果になる操作を、冪等という。HTTP はどのメソッドが冪等かを定義し、冪等性キーはその性質を POST にも拡張して、クライアントが重複を発生させずに再試行できるようにする。
-
Teams with fewer, alert-linked dashboards diagnose incidents faster than teams with many unowned dashboards
Hypothesis: for services of comparable size, the time from a page to a named probable cause is shorter when the team maintains a small number of owned dashboards that alerts link to directly, than when it maintains many copied or auto-generated dashboards that responders have to browse; a proposed comparison using incident timelines and dashboard inventories.
-
Handling time: UTC, ISO 8601 and time zones
Store and exchange instants in UTC using RFC 3339 timestamps, keep time-zone-aware objects in code, convert to local time only for display, and treat calendar arithmetic as a separate problem from instant arithmetic.
-
z/OS system operations basics: console commands, SYSLOG/OPERLOG, and what an IPL is
Console commands like D A,L (list active jobs) and D IPLINFO (display IPL information) are the read-only starting point for understanding what a z/OS system is doing; SYSLOG and OPERLOG are where system messages accumulate, and an IPL (Initial Program Load) is the mainframe's equivalent of a full system boot, not something to trigger outside a planned change window.
-
Java のストリームとループ：パイプラインが分かりやすい場合とそうでない場合
ストリームは、1つの終端操作で閉じる中間操作の遅延パイプラインである。コレクションの絞り込み、変換、収集には読みやすいが、パッケージ文書はラムダ内の副作用を非推奨としており、再利用はできず、検査例外にもなじまない。状態を伴う早期終了、インデックスを使う処理、変更にはループのほうが明確である。
-
CI で実行する例を使った API ドキュメント
API 文書のリクエスト・レスポンス例をすべて実行可能にする。スキーマで検証する OpenAPI の例として保存し、CI でサンドボックスに対して実行し、そのファイルから文書を生成する。動かなくなった例が読者を誤解させる前に、ビルドを失敗させる。
-
Reading the AIX error log with errpt and feeding it into syslog
errpt is AIX's own hardware/software error log, started earlier in boot than syslog and populated by the kernel and error-aware daemons, not by application log lines. errclear trims it by class and age; an errnotify ODM object plus logger is the documented way to mirror errpt entries into local or remote syslog.
-
実行中のエージェントの方針を変える：指示のリビジョン管理と副作用の照合
方針変更はリビジョンとして記録し、進行中の操作との境界を明確にする。この提案手順は、キャンセル要求、確認済みの影響、再開可能な作業を区別し、遅れて届いた結果によって古い目的が暗黙に復活することを防ぐ。
-
LiteLLM 1.107.0：侵害されたPythonを再起動せずにアーティファクトの影響範囲を調べる
10月のLiteLLM侵害はPyPIのwheelとコンテナイメージに影響した。Pythonを起動する前に保存済みのメタデータを確認し、影響を受けた環境を再構築して、侵害されたシステムからアクセスできた認証情報を失効させる。
-
JSON での BigInt：replacer を追加する前に情報を失わないフィールド契約を定める
大きな整数を明示的なスキーマで符号化し、暗黙に Number を経由せずに復元します。
-
Phishing-resistant sign-in with WebAuthn and passkeys
WebAuthn authenticates with a per-site public-key pair: the browser only lets a credential be used by origins under the relying party ID it was registered for, and the authenticator signs a server challenge, so a look-alike site obtains nothing replayable; passkeys are discoverable WebAuthn credentials, often synced across a user's devices.
-
Package managers compared: apt, dnf, zypper, apk, pkg, Homebrew and winget side by side
Install, remove, search, list, find the owning package, pin a version and run non-interactively — the same seven tasks, phrased differently by apt/dpkg, dnf, zypper, apk, pkg, Homebrew and winget. The table also marks the task that Homebrew and winget simply do not offer.
機械可読： JSON