English | 한국어

Validation record

Date: 2026-09-20. Source preview 0.2.0. Local checks and GitHub-hosted platform checks are distinguished below; publishing the source does not constitute a packaged release.

Product checks

• Recorded Python regression suite: 146 tests before localization. Candidate contracts, confidence gates, HTTP reuse/concurrency, context preparation, dashboard server, Native Messaging, installation backup/recovery, and browser loop boundaries.
• Recorded JavaScript controller suite: 4 tests before localization. No ON indication before host confirmation, duplicate toggle prevention, uncertain failure state, settings/OFF changes.
• uv build: wheel and source distribution generated. Packaging does not prove extension installation or account connectivity.
• Shell/Python syntax, skill frontmatter, and Git whitespace checks.

Browser coverage includes six page transitions, page/action cycles, unchanged pages with different refs, stale observations, step/time budgets, exact completion URLs, OFF/confidence changes during decisions, invalid IDs/abstain/model errors, relevant content near the end of large pages, aggregate context preservation, and candidate-description collisions after sanitization. Aside initialization is included in the overall time budget.

Actual Aside runtime

scripts/verify_browser_runtime.py opened a localhost fixture tab in a persistent REPL through actual aside mcp. It clicked links six times, refreshed observations each time, and verified completion text plus exact URL at the seventh choice. The test tab and HTTP server were closed afterward.

Recorded first run: 3,124.27ms, six executions, verified. The decision function was a fixture stub; no actual Jev API call was made. This does not establish Jev quality, inference latency, or general website success rates.

Native Messaging used temporary profiles and registration directories to exercise generated wrappers and length-prefixed JSON. Actual user Aside profiles and browser registration files were not changed.

Latency

scripts/benchmark_latency.py --samples 30 connects the actual TypeSafe SDK to a localhost HTTP fixture without artificial delay.

Strategy	p50	p95	TCP connections
New SDK client per request	1.157ms	1.695ms	30
Reuse client and connection	0.540ms	0.747ms	1

Transport-path p50 overhead fell approximately 53.3% in this environment. Internet latency, TypeSafe server processing, and actual inference are excluded. timeout_s applies separately to HTTP phases; SDK automatic retries are zero.

The new-client group ran first, followed by reuse, with 30 requests each including the initial connection. There were no repeated trials, randomized ordering, or concurrent loads. Baseline calls the SDK directly, while reuse includes choose_live validation/context preparation, so client lifetime is not the only controlled variable. Per-request closing is included in baseline timing; final shared-client closing is outside the reuse interval. Raw samples were not retained; the video uses these aggregates.

Speed duel video

Source and reproduction · English MP4 · English GIF · Korean MP4

A 1920×1080, 30fps, 20-second data visualization. Both lanes use elapsedMs = (frame - 60) / 150 × 1.157, stopping at their respective p50. Baseline 1.157ms expands to five screen seconds; reuse finishes after approximately 2.334 seconds. It does not represent recorded browser work, the total time of 30 requests, or default Aside model versus Jev inference.

The original video passed TypeScript/ESLint, representative-frame review, H.264 decoding, and metadata checks (600 frames / 20 seconds). The GIF is 960×540, 240 frames, looping for 20 seconds. Computer Use verified the local README GIF and error-free MP4 playback through the end. Local links and loopback-only binding were checked. Remote GitHub rendering was not tested.

UI evidence

Computer Use verified local demo decisions, results/probabilities/timing, invalid JSON, settings cancellation, and a 390px viewport. Without a key, Live selection is disabled. One observed demo round trip was 6.9ms; this is not a benchmark or Live latency measurement.

The popup HTML preview covered an unconfirmed host, disabled toggle, and reconnect guidance on localhost. Actual extension-context Native Messaging remains unverified.

Original Korean UI evidence

English-default localization

• Full regression run after localization: 152 Python tests and 20 JavaScript tests passed. Includes locale persistence/fallback, blocked storage, matching dictionaries/placeholders, HTML accessibility labels, real UI event flows, edited-input preservation, and stable native request contracts.
• TypeScript/ESLint passed. Both video locales decoded successfully as H.264, 1080p, 30fps, 600 frames / 20 seconds. Each GIF has 240 frames / 20 seconds. English/Korean frames were reviewed together.
• Computer Use confirmed fresh English UI, Korean switching, saved Korean after reload, translated result/status text, and preservation of a custom goal. Popup verification used localhost HTML with an unconnected host, not a real extension registration.
• Local English/Korean README previews displayed the matching GIF. Both language screenshots were refreshed from the actual demo UI. The English selector and labels fit a 390px viewport without horizontal overflow; the temporary viewport override was cleared afterward.
• Documentation artifact mappings and local links passed scripts/check_i18n.py. Wheel and source distribution rebuilt; packaged dashboard language modules were confirmed.
• CLI-owned messages default to English, with --lang ko for Korean. Original backend/provider diagnostics and generated agent instructions are outside the translated UI contract; see localization boundaries.

Guided installation on Windows and macOS

• Full local regression after installer changes: 215 Python tests passed, 1 Windows-only integration test skipped; 20 JavaScript tests passed. Translation parity caught one missing mirrored label during development; both dictionaries were synchronized and the JavaScript suite passed afterward.
• Final GitHub-hosted verification, source 625898b: Windows 222 Python tests passed; macOS 221 passed and 1 Windows-only test skipped; each platform passed 20 JavaScript tests, wheel building, and 115 local documentation links. Both bootstrap runs used Python 3.11.9 and verified nine packaged extension assets, English installation, Korean repeated setup, preserved settings, default OFF, and framed native responses. Windows additionally verified an isolated HKCU registration and its cleanup. Temporary files were removed on both platforms.
• Windows CI exposed a JSON-escaped path assertion, % expansion in launcher paths, an unavailable Get-FileHash command, and a non-UTF-8 output stream under isolated Python. Fixes compare path objects, reject unsupported launcher paths, use built-in .NET hashing/ZIP APIs, and explicitly select UTF-8. The final run exercises these corrected paths.
• On macOS, the actual bootstrap ran in a temporary directory with uv excluded from PATH. It downloaded and SHA-256-verified uv 0.12.17, prepared Python 3.11.16, and installed the packaged app. A subsequent install used the locked dependency export with hash verification and Korean setup output.
• Synthetic account paths included spaces and Korean characters. Account settings and other MCP entries were preserved; setup cancellation, previews, repeated setup, key secrecy, settings races, conflicting ownership, and rollback were tested. No real user account files were changed.
• The installed native-host executable accepted the fixed extension origin and returned a correctly framed status response with MCP configuration recognized and OFF retained. This proves the local process contract, not a browser-launched connection.
• Windows regression tests exercised real Win32 file handles, locks, and native .cmd execution with Korean text, spaces, and ! & ^ () in the path. % in the native launcher path is rejected before writing; Windows would otherwise expand it before entering the wrapper. Registry ownership/view conflicts and failure recovery use a fake registry. Tests do not establish arbitrary Windows ACL configurations or actual Aside browser registration.
• Installer-owned code and extension locale assets are included in the wheel. The extension public key produces the same ID independent of folder location. No private signing key is included.

Remaining boundaries

• Actual Jev API authentication, decision quality, and inference time remain unverified.
• Real-account extension installation, Native Messaging registration, and MCP configuration remain unverified.
• ON affects new-task instructions and the provided MCP; no native hook globally intercepts existing direct tools.
• Task-specific action rules and completion conditions are required. General browser-agent capability is not established.
• OFF/timeout does not undo actions already sent. Inspect unconfirmed state before manually resuming.
• Sanitization is an auxiliary size/credential-pattern filter, not universal personal-data detection.
• CI uses synthetic accounts. It does not establish actual Aside browser connectivity or a published GitHub release.

Integrated completion assessment

설치된 MCP 브라우저 도구는 최종 DOM 확인 → 고정된 질문 계약으로 Jev Noul/Score 평가 → DOM 재확인을 수행합니다. browser_verified는 브라우저 완료를 별도로 기록합니다. 전체 verified는 완료 확률과 위험 평가 확신이 현재 최소 신뢰도 이상이고, 0..2 위험 척도에서 점수가 1.5 미만일 때만 참입니다. 공급자 오류·낮은 확신·높은 위험·화면 변경·시간 초과는 전체 성공으로 반환하지 않습니다. 평가 실패나 시간 초과에도 이미 실행한 행동 기록을 보존합니다. 이 사후 평가는 민감한 행동의 사전 승인을 대체하지 않습니다. 일반 System One 질문은 instructions와 criteria를 사용하며 잘못된 question·rubric 필드는 외부 호출 전에 거부합니다.