Skip to content

Build Week Baseline

Cutoff and repository evidence

The official cutoff is 2026-07-13 00:00 America/New_York. The latest commit reachable from main and origin/main before that time was:

Evidence Value
Baseline commit 36b2a547056d40eff32f00aa59b7820f7d3e98d5
Author/committer time 2026-07-12T22:41:16-04:00
Subject feat: enforce administrative step-up authentication
Baseline tag pre-build-week-2026
Build Week branch build-week/guardian-review
Branch starting commit 36b2a547056d40eff32f00aa59b7820f7d3e98d5
Existing version 0.1.0-alpha.1
Remote https://github.com/the-vibe-dev/guardiannode.git

The working tree had no tracked modifications. It did contain unrelated, untracked WordPress compression/validation archives and scripts. They could not be attributed to GuardianNode from repository evidence and were not moved, deleted, staged, or included in the baseline commit. Because they are already untracked, no claim is made about when their content was produced.

Existing refs also included agent/beta-release-hardening at 282e00a87c33c74cb249f8289c93f0122600c752 and the prior release tag v0.1.0-alpha.1, which peels to 6ce29460932826d52afb569eb464e52f81439962.

Existing release

GitHub already hosted the public prerelease v0.1.0-alpha.1, published July 4, 2026, with a checksum manifest, Windows child and server installers, and a Linux server installer archive. The Windows executables are approximately 49.7 MB and 31.7 MB; the Linux archive is approximately 23.8 KB. Local ignored installer build outputs also existed and are not repository source.

Existing modules

At the cutoff, 448 files were tracked. The principal source areas were:

Area Tracked files Existing role
backend/ 204 FastAPI API, workers, detection, persistence, evidence, auth, audit
docs/ 53 Parent, developer, safety, deployment, and release documentation
agent-windows/ 44 Windows capture, queue, pairing, health, watchdog, and tray
dashboard/ 38 React/Vite parent dashboard
installer/ 28 Windows and Linux deployment assets
assets/ 16 Branding and release assets
tests/ 14 Repository, release, benchmark, and synthetic E2E checks
shared/ 12 Shared schemas, constants, and Python types

Existing features before Build Week

GuardianNode agent

  • Visible Windows screenshot capture with application/window context.
  • Durable local frame queue, idempotent upload, retry, heartbeat, and pairing.
  • Broker/watchdog services, tray/status experience, and hardware probing.

Backend and local detection

  • FastAPI backend with SQLite/Alembic persistence and background workers.
  • Tesseract OCR, deterministic safety rules, and optional local Ollama text/vision classifiers.
  • Canonical category normalization, policy evaluation, alert deduplication, notifications, offline monitoring, and classifier readiness status.

Dashboard and UI

  • Setup/login, dashboard health, devices, profiles, privacy policy, risk feed, alert details, encrypted screenshot reveal, model status, settings, and audit.
  • Existing alert review and parent feedback controls. These are general alert controls, not Guardian Review result feedback.

Evidence handling

  • AES-256-GCM encrypted evidence blobs, protected key handling, DPAPI wrapping on new Windows installations, audit-on-view, retention, wipe, encrypted export, backup, and recovery paths.

Installer and deployment

  • Windows all-in-one, server-only, and child-only Inno Setup paths.
  • Linux server installer, Docker Compose, release checksums, repair/rollback, uninstall, and qualification documentation.

Authentication and security controls

  • One-time setup token, parent password login, server-side sessions, idle and absolute session expiry, CSRF protection, recent and critical step-up auth.
  • Recovery code, device bearer tokens, secure host validation, rate limiting, webhook restrictions, security headers, audit logging, and repository governance checks.

Existing tests at the cutoff

The complete practical suite was run on the untouched baseline before Guardian Review implementation:

Suite Command Result Test time Wall time
Backend cd backend && .venv/bin/pytest -q 224 passed 46.64 s 51.12 s
Windows agent cd agent-windows && ../.venv/bin/pytest -q 59 passed 6.34 s 8.88 s
Repository/E2E .venv/bin/pytest -q tests tests/e2e 58 passed 8.82 s 11.52 s
Dashboard cd dashboard && npm test -- --run 12 passed 13.24 s 20.31 s

Unique total: 353 passed, 0 failed, 0 skipped. The E2E directory is already included by the root invocation, so it is not counted twice.

Additional passing checks included dashboard typecheck/build, Python compile, Ruff, backend mypy, feature/version/hardware synchronization, third-party notices, repository controls, Linux shell syntax, strict MkDocs, actionlint, the 196-case deterministic benchmark, backend/agent dependency audits, npm production audit, and the Docker OCR-to-alert canary.

The baseline emitted one existing Starlette/httpx deprecation warning. No test failure was silently repaired. Windows-only bundle and installer gates were not runnable in this Linux environment and remain covered only by the existing July 9 qualification evidence.

Environment: Kali Linux rolling 2025.4 x86_64, kernel 6.16.8; Python 3.13.9 in the root environment, Python 3.12.13 in the backend environment; Node 22.22.2, npm 9.2.0; Tesseract 5.5.0; Docker 28.5.2; Git 2.53. The dashboard passed under Node 22, while repository support documentation continues to require Node 24.

Existing screenshots and assisted work

Tracked assets/1.png through assets/6.png are branding boards covering brand essence, typography, colors, horizontal logo, application icon, and vertical logo. No dashboard screenshot was found. The verifiable current UI state is therefore the dashboard source and committed production bundle.

The repository owner identifies the existing dashboard/visual UI as Claude-assisted and the agent/backend/security/installer/platform work as Codex-built. Git contains one Claude-coauthored hardening commit, but does not contain consistent assistant metadata sufficient to verify the entire split. This attribution is recorded as owner-supplied rather than inferred fact.

Existing known limitations

  • Alpha detection can miss risks and produce false positives.
  • OCR and local-model quality varies with hardware and visible content.
  • Screenshots can contain highly sensitive family data.
  • Windows 11 x64 is the promoted alpha client path; Windows 10 is not promoted.
  • Separated deployments require a trusted VPN/TLS design; built-in mTLS is planned and raw public-internet exposure is unsupported.
  • Binaries are unsigned and may trigger reputation or antivirus warnings.
  • Current Windows 11 smoke, Windows 10 qualification, and suspend/resume gates remain incomplete after the July 9 lab run.
  • GuardianNode is not an emergency service, a comprehensive content-control product, or a substitute for parent-child communication or professional help.

Added during Build Week

The baseline tag/branch, evidence documents, submission audit corrections, README disclosure, and Guardian Review schema contract were added after the cutoff. The Guardian Review service and GPT-5.6 runtime were not present in the baseline and are not claimed as implemented by these documents.