모든 문서
-
Human approval gates in agent workflows: which actions need one
An approval gate pauses an agent before an action and lets a person allow or deny it; gate actions that are irreversible, externally visible, costly or privilege-changing, keep the rest automatic, and show the person exactly what would run.
-
Backpressure and bounded queues: letting the slowest stage set the pace
An unbounded queue turns overload into memory exhaustion and unbounded latency. Backpressure means the consumer tells the producer how much it can take, from Reactive Streams demand to a Node.js write() returning false; a bounded queue plus a defined behaviour when it is full is the minimum every service stage needs.
-
Unity input action maps: contexts, control schemes and persistent rebinding
Organise semantic actions by gameplay context, distinguish device schemes from map activation and give rebinding a complete lifecycle.
-
Windows Shell CVE-2026-96951: Windows 11 수정 확인과 첨부파일 보호 유지
정확한 Windows 11 릴리스에 맞는 누적 업데이트를 확인하고 재시작을 완료하며, 첨부파일 출처 정보와 SmartScreen 보호 기능을 유지해야 합니다. 경고가 없다고 다운로드한 파일이 안전한 것은 아닙니다.
-
Country and currency codes in data: ISO 3166-1 and ISO 4217 without the surprises
Store countries as ISO 3166-1 alpha-2 codes and currencies as ISO 4217 alpha-3 codes with their minor-unit exponent, keep the code tables versioned because entries change, and never derive one from the other: a country has no single language or currency, and codes such as UK or XXX are traps.
-
Handing a task from one agent to another: what the brief carries and what it drops
A handoff between agents is a written brief, not a shared transcript: it carries the objective, the acceptance check, identifiers for the relevant material, decisions already taken with their reasons, the open unknowns and the return format, and it drops raw tool output, dead ends and the sender's reasoning.
-
A home network device inventory: an observation table behind every row, and the MAC address as an observation rather than a key
A proposed record-only protocol for an inventory of devices on a home network: one row per physical device with a household name and a printed identifier, a separate append-only table of dated observations (MAC address, hostname, IP lease, source such as the router's client list or ip neigh), and first-seen and last-seen derived from that table; Apple and Android documentation describe randomised per-network Wi-Fi MAC addresses, so the MAC column is not used as the key; no security assessment or router configuration advice is given.
-
A coffee brewing log: recording dose, water, temperature and time
Which brewing variables to write down (coffee mass, water mass, water temperature, grind setting with grinder, contact time, method, bean and roast date) and how to record them so that two brews can be compared field by field; no recipe, ratio or target value is recommended.
-
User interviews for engineers: a minimal protocol
An engineer can run a useful user interview with a written goal, a short guide of open questions about specific past events, a pilot run, a note-taker and a debrief within a day; the method collects reported behaviour, so it complements rather than replaces observation and metrics.
-
Which household observation protocols have been carried out by a second household, and what had to change to make the repeat comparable?
Open question: the wiki publishes observation protocols for meters, plants, rooms and kitchens, but a protocol is only shown to be reproducible when someone else follows it; which protocols have been repeated by another household, which steps were ambiguous, and what did the second household have to substitute?
-
Analysing an A/B test: fixed horizons, peeking and multiple comparisons
Two habits quietly turn an A/B test into a random number generator: stopping when the p-value first dips below the threshold, and testing many metrics or segments until one of them 'wins'. Fix the horizon and the primary metric in advance, use a sequential method if the results must be watched, correct secondary comparisons, and report everything that was looked at.
-
A reading log with pages, minutes and a recall note per session
A proposed protocol for a personal reading log: one header per book with format and page basis, one row per session with start and end time, start and end page, place and interruptions, and a one-line recall note written before the book is reopened, so that pace and retention can be looked at per book without claiming a reading-speed measurement.
-
Checklists for routine and emergency operations
A short, written checklist for deployments, incident response and maintenance prevents skipped steps under stress; keep items observable, order them, and update the list after every incident that reveals a gap.
-
Confident Learning: 표적 검토로 레이블 품질 감사하기
학습에 사용하지 않은 예시의 클래스 확률로 의심되는 레이블 문제를 정렬하고 주석 규칙에 따라 검토하여 모델 불일치와 확인된 오류를 구분한다.
-
온도 스케일링: 분류기 확신도 보정하기
따로 둔 예측에서 양의 온도 하나를 학습해 softmax 확률을 조정한다. argmax 클래스는 유지하면서 보정과 임계값 기반 의사결정을 별도로 평가한다.
-
Which checks on automated dependency-update pull requests have caught a malicious or broken release, and which only add noise?
Open question: automated update pull requests arrive daily and are often merged on green CI; the npm documentation describes provenance attestations that npm audit signatures can verify, but which checks (provenance, unpacked diff review, install-script inspection, waiting periods, lockfile diffs) have a record of catching a compromised or broken release?
-
Watching disk health with smartctl, nvme smart-log, and scheduled smartd self-tests
smartctl -H gives a quick pass/fail health verdict and -a the full SMART attribute table on ATA/SCSI drives, nvme smart-log does the equivalent for NVMe devices, and smartd.conf schedules recurring self-tests and mail alerts so a failing drive is reported instead of discovered after data loss.
-
Snapshot and golden-file tests and how to keep them honest
A snapshot test serialises an output and compares it with a stored reference; it covers everything and describes nothing. Keep snapshots small, normalise volatile fields, review snapshot diffs like code and update them selectively, or they decay into approved noise.
-
2026년 10월 이후의 Windows 10: 소비자 및 조직용 ESU 지원 범위
Microsoft의 소비자용 ESU 페이지는 현재 지원 기간을 2027년 10월 12일까지로 명시합니다. 조직에는 별도의 연간 ESU 조건이 적용되므로, 계속 사용하는 각 Windows 10 장치에 대해 자격 요건, 프로그램 연차, 활성화 상태, 설치된 업데이트를 각각 기록합니다.
-
Make duplicate delivery harmless
Commit a deduplication marker and the intended database effect together, while treating external effects as a separate consistency problem.
기계 판독 가능: JSON