Skip to content

Devpost Submission Draft

Project name

GuardianNode — Guardian Review

One-line tagline

A local-first family safety monitor that lets a parent preview a privacy-minimized GPT-5.6 second opinion and turn an alert into a calmer conversation.

Track

Apps for Your Life

Problem

Parents can face alarming fragments of online activity without enough context. A keyword or screenshot may matter, may be a joke, or may be school research. An unexplained “high risk” label can provoke an accusation when the family first needs careful fact-finding, immediate safety separation, and proportionate next steps.

Solution

GuardianNode detects risk locally on a Windows device the parent owns or administers. Guardian Review is an optional Build Week extension: the parent opens an existing incident, supplies coarse relationship/context information, reviews the exact minimized text proposed for external processing, and chooses whether to continue. GPT-5.6 returns a strict assessment that separates observed facts from inference, surfaces uncertainty and benign explanations, and creates a practical, non-accusatory conversation plan. The parent remains the final decision-maker; the model cannot punish, block, contact, diagnose, or make legal conclusions.

How it works

  1. The visible Windows agent captures configured screen evidence.
  2. The parent-owned backend encrypts evidence and runs deterministic/local detectors and optional local Ollama models.
  3. A normalized local risk creates a dashboard incident.
  4. Guardian Review loads that authorized incident server-side; it never trusts a browser-supplied incident payload.
  5. Deterministic minimization removes paths, device/account identifiers, emails, phones, handles, URLs, precise locations, and unrelated context.
  6. The parent sees the exact outbound JSON, can remove optional fields, and must explicitly consent.
  7. The server-side Responses API request uses configurable gpt-5.6, store: false, no tools, and strict schema 1.1.0.
  8. The encrypted local result presents uncertainty, actions, and a conversation plan; versioned parent feedback remains local.

Existing-project / Build Week disclosure

The immutable pre-Build Week baseline is tag pre-build-week-2026, commit 36b2a547056d40eff32f00aa59b7820f7d3e98d5. It already contained the Windows agent, FastAPI backend, React dashboard, local detection, encrypted evidence, authentication/security controls, installers, and 353 passing unique tests.

Guardian Review is the Build Week extension: its service, strict schema, GPT-5.6 provider, minimizer and consent preview, durable persistence, complete parent communication UI, feedback, six-scenario judge demo, 55-case evaluation, and related reliability/privacy hardening were added after that tag. The owner describes portions of the pre-existing visual web UI as Claude-assisted and the pre-existing agent/backend/platform work as Codex-built; Git cannot completely verify assistant attribution.

How Codex was used

Codex preserved and tagged the baseline, traced the real incident path, defined the schema/privacy contract, implemented backend and UI changes, wrote tests and synthetic fixtures, ran release and security audits, fixed the findings, and assembled reproducible submission evidence. Important human decisions included keeping cloud review opt-in, binding consent to exact bytes and versions, keeping feedback local, refusing diagnostic/enforcement claims, and disabling a convenient coding-agent transport when its capability boundary was not safe enough for family incident evidence.

How GPT-5.6 was used

During development, GPT-5.6-powered Codex accelerated repository analysis, implementation, verification, and documentation. At runtime, Guardian Review's direct OpenAI Responses API path defaults to gpt-5.6 and returns only the versioned strict assessment object. The request sets store: false, supplies no tools, and is blocked unless the operator explicitly confirms the configured project's approved Zero Data Retention controls. We do not claim that store: false alone guarantees zero retention.

An experimental “Sign in with ChatGPT” Codex path was evaluated using only synthetic data. We disabled it after adversarial review because a coding agent can have local read tools. Parent-friendly subscription integration remains a goal, but it must first provide enforceable zero-tool isolation.

Challenges

  • Minimizing enough context to protect privacy without removing the facts that make an incident understandable.
  • Designing a strict schema that is useful for a worried parent without presenting an AI assessment as fact.
  • Making consent meaningful and resistant to stale previews, duplicate clicks, refreshes, timeouts, and restarts.
  • Being honest about model limits: a deterministic mock is excellent for the judge path but not evidence of nuanced judgment, and a small synthetic live sample is not a universal accuracy claim.

Accomplishments

  • Complete incident → preview → consent → strict assessment → encrypted local persistence → communication plan → parent feedback path.
  • Six resettable synthetic judge scenarios, usable in mock mode without real family data or an API key.
  • Fifty-five synthetic evaluation cases across concerning, ambiguous, benign, missing-context, false-positive, quoted/research/gaming/medical, and prompt-injection groups.
  • Explicit auth, CSRF, authorization/IDOR, timeout, retry, malformed output, audit leakage, redaction bypass, XSS, and failure-recovery tests.
  • A security-driven decision to fail closed rather than ship a more convenient but overly capable provider transport.

What we learned

The most important output is often not a risk label. Parents need a clear split between facts and inference, plausible benign explanations, what remains unknown, and words that help preserve trust. Privacy also needs product UX: local redaction alone is not meaningful consent unless the parent can inspect and control the exact outbound content.

Next steps

  • Qualify the unsigned Windows release candidate on fresh Windows 11 hardware, including reboot, uninstall, reinstall, firewall, and failure recovery.
  • Expand international identifier/location minimization and independent human review of the evaluation rubric.
  • Add a genuinely consumer-friendly OpenAI connection only when it can enforce a zero-tool privacy boundary.
  • Conduct opt-in beta onboarding with synthetic-first support and no silent behavior changes from individual feedback events.

Repository and judge instructions

  1. Read README.md and docs/build-week/BASELINE.md for the exact prior/new split.
  2. For a no-key demo, enable Guardian Review and demo mode with provider mock.
  3. Open Synthetic demo, select a scenario, trigger it, open the incident, inspect the outbound preview, consent, review guidance, save feedback, reset.
  4. Run python -m app.guardian_review_evaluation --provider mock from backend/ for machine-readable results.

GuardianNode is AGPL-3.0 alpha software. It can miss risks or create false positives; it is not an emergency, diagnostic, legal, or substitute-parenting service.