Skip to content

Build Week Changelog

Changes in this file are relative to the immutable pre-build-week-2026 tag.

2026-07-13 — Baseline and submission foundation

Added:

  • Verified baseline tag and dedicated Build Week branch.
  • Baseline, collaboration, privacy, evaluation, specification, checklist, and preliminary Devpost evidence.
  • Versioned Guardian Review strict assessment JSON Schema.
  • README Build Week disclosure and safe sample environment file.
  • Public-readiness audit results and sanitized current release-validation text.

Not yet implemented:

  • Guardian Review backend service, database migration, worker, and dashboard UI.
  • OpenAI Responses API calls or GPT-5.6 runtime dependency.
  • Outbound preview/consent interaction, Guardian Review feedback, synthetic judge demo, or evaluation harness.

Future entries must identify the implementation commit, tests, prompt version, schema version, model configuration, and any privacy-impacting change.

2026-07-14 — Guardian Review backend and GPT-5.6 integration

Implemented:

  • Schema 1.1.0, prompt guardian-review-v1, encrypted preview/result tables, Alembic revision 0003_guardian_reviews, durable worker, audit events, and cascade-safe retention.
  • Authoritative alert loading, bounded parent context, coarse age group, minimization/redaction, exact preview, expiring digest-bound consent, idempotent duplicate handling, and explicit fresh assessments.
  • Direct OpenAI Responses API provider using configurable gpt-5.6, strict text.format JSON Schema, store: false, no tools, 45-second timeout, and at most two attempts for eligible transient failures only.
  • ChatGPT-authenticated Codex provider using configurable gpt-5.6-sol, ephemeral/read-only isolated execution, strict output schema, service-owned OAuth storage, and guided device login. API-key mode remains an advanced optional provider.
  • Deterministic mock provider and guardiannode-review-harness synthetic CLI.
  • Parent dashboard provider/readiness and “Connect with ChatGPT” flow.

Verified:

  • Untouched July 14 baseline: 353 passed, 0 failed, 0 skipped.
  • Live synthetic Codex path: completed, schema-valid, encrypted locally, requested/returned model gpt-5.6-sol, 34,112 ms provider latency.
  • An initial explicit gpt-5.6 Codex probe failed safely because this ChatGPT-authenticated CLI accepts the Codex alias gpt-5.6-sol; the direct Responses provider retains the required gpt-5.6 API default.

Still planned:

  • Alert-page outbound preview/consent/result presentation and Guardian Review-specific parent feedback.
  • Broader frozen judge-scenario rubric and Windows installer qualification of the Codex device-login flow.

Implemented:

  • Redaction contract guardian-review-redaction-v2 with Unicode normalization, invisible-control removal, obfuscated identifier handling, incident-scoped stable placeholders, path/device/account/location masking, and relevance-aware URL minimization.
  • A stricter outbound allowlist that excludes screenshots, local incident and device/profile IDs, exact scores/counts, classifier state, unselected context, and full event text by default. Evidence and total payload sizes are bounded.
  • Provider readiness gates, exact provider/retention disclosures, digest binding to the redaction contract, and sanitized versioned audit metadata.
  • Alert-page guided optional-field controls, separate local/transmitted views, exact read-only JSON, unchecked consent, cancel-without-send, result display, per-alert history, and global history.
  • Parent-controlled deletion that scrubs encrypted preview/context/assessment content and provider response identifiers while retaining a minimal audit tombstone.
  • Alembic revision 0004_guardian_review_privacy plus backend and dashboard privacy, cancellation, deletion, history, authorization, and bypass tests.

Still planned:

  • Guardian Review-specific parent feedback/evaluation reporting.
  • Expanded international address/identifier coverage and frozen judge scenarios.

2026-07-16 — Parent communication coaching and complete UI

Implemented:

  • Complete advisory result hierarchy: observed facts, concern rationale, benign explanations, unknowns, parent questions, recommended tone, opening language, child questions, phrases to avoid, actions, escalation indicators, and limitations.
  • Age/severity/repetition/relationship/goal-aware guidance through the strict schema and versioned prompt contract.
  • Local versioned feedback labels: helpful, inaccurate, too alarmist, too dismissive, missing context, and needs follow-up.
  • Keyboard/focus semantics, live status regions, escaped text-only rendering, responsive sections, and tests for consent/result/feedback states.

2026-07-17 — Synthetic judge demo and recovery

Implemented:

  • Six resettable, visibly synthetic scenarios covering secrecy/platform migration, ambiguous gaming language, school research, peer coercion, medical/educational discussion, and a harmless control.
  • Guided dashboard path from demo health through local detection, incident, preview, review, feedback, and reset.
  • Durable queued/running states, idempotent duplicate submission, safe failure persistence, refresh/restart recovery, bounded screenshot backlog, and corruption/error handling.

Manual dependency:

  • Build Week credit-request submission is an external account/form action and cannot be confirmed from repository evidence; no credentials or form answers are stored here.

2026-07-18 — Evaluation and adversarial review

Implemented:

  • A balanced 55-case synthetic dataset across 11 requested groups, a machine-readable scorer, and explicit property-based results.
  • Prompt-injection, stored-XSS, authorization, redaction-bypass, oversized input, logging, and capability-boundary review.
  • Redaction contract v3 for bracket-dot emails, IPv6, and complete Windows paths; stale previews automatically fail digest/version validation.
  • Fail-closed Codex provider/device-login after identifying that a coding agent's local tools are unsuitable for untrusted family incident evidence.
  • Local classifier prompt cap/concurrency gate, positive list limits, token prefix removal from logs, and demo reset ownership markers.

2026-07-19 — Release hardening candidate

Implemented:

  • Version 0.1.0-alpha.2, release notes, configuration/troubleshooting guide, clean-install checklist, sample configuration, mock demo data, and checksum workflow inputs.
  • Correct Tailwind v4 CSS entrypoint so the production dashboard bundle includes the complete responsive design system; capture five real synthetic-flow screenshots from a disposable backend/browser run.
  • Full repository test, type, build, dependency, secret, license/notices, and link audits are recorded in the daily report on the exact commit.

Still manual before beta promotion:

  • Fresh Windows 11 installer, reboot recovery, firewall behavior, uninstall, reinstall, signature, and artifact-checksum qualification.

2026-07-20 — Submission freeze

Implemented:

  • Final README disclosure, Devpost copy, architecture/privacy/evaluation links, and a 2:48 voiceover/shot package with captions, a machine-readable manifest, a Codex computer prompt, a production runbook, and a disposable mock-only recording server.
  • Functionality freeze: only golden-path defects and submission evidence change after this entry.

Still external/manual:

  • Record and upload the public YouTube video, verify logged-out repository/video access, and submit the final Devpost form.

2026-07-21 — Final verification shell

Added:

  • A deterministic submission-package validator for required disclosures, local links, baseline identity, video runtime/assets, and Devpost sections.
  • Final claim-to-evidence review, private submission-note template, release/tag identity, and exact tomorrow handoff for Windows, video, /feedback, and Devpost account gates.

External status remains explicit: real-node qualification, a live direct-API synthetic demonstration, YouTube publication, /feedback, and form submission are not represented as complete until the maintainer performs them.

Late July 21 qualification completed the Windows 11 all-in-one and separate child/server paths, a real synthetic agent-capture alert, and a live GPT-5.6 Guardian Review with strict structured persistence. The final 2:48 H.264/AAC video uses disclosed OpenAI Coral narration and embedded English captions. YouTube publication, /feedback, and the Devpost account submission remain maintainer actions.

Verified and published internally as a draft release:

  • Merged release commit 62590db1911cb84cf01b27de76ab26f238d003d7 and immutable tag guardian-node-build-week-2026-final.
  • Passed the final GitHub release workflow and produced both versioned Windows installers, SHA256SUMS, and release-manifest.json.
  • Downloaded the artifacts and verified both SHA-256 values. The prerelease remains draft and the unsigned installers remain unqualified on a real node.