# Childhood Confidence Learning Lab

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Albert Tumblin](https://cerebralvalley.ai/u/RockStarsushi3)
- **GitHub:** https://github.com/GoldSpur/ccc-learning-lab-hackathon
- **Demo video:** https://www.youtube.com/watch?v=YQKBWFPpU00
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1277

What was pre-existing vs built this week.
The Confidence Wheel, the Confidence Cycle, the Appendix B parent scripts, and the safety copy are the founder's own published framework content from the book Create Childhood Confidence (Albert Tumblin III), labeled and hashed in the repository's PROVENANCE.md. Everything else — the Learning Lab application, the model gateway and budget safeguard, the deterministic validator, the safety screen, the local script matcher, the evidence trace, the user interface, and the evaluation work — was built during the hackathon (October 26 – November 4, 2026).

What the video shows.
A fictional scenario ("I'm stupid" — homework) runs through the governed path: deterministic safety screening, a Muse Spark 1.3 interpretation labeled RECORDED LIVE (a real model response recorded on October 26, 2026, and replayed — the hosted demonstration makes no live calls), deterministic validation of every identifier against the published framework, and local, auditable selection of a verbatim book script. A parent override re-validates without changing the script by a word. A fictional safety test stops before any model call (trace: 0 model calls) and shows fixed safety wording labeled PROVISIONAL — independent safety review is still pending. A scenario with no honest match returns a labeled fallback instead of an invented script. The closing section reports the preregistered Phase 1 live validation run, conducted after the build was frozen: 21/21 live dispatches completed with no retries, 7/7 governed proposals valid, the safety screen stopping 7/7 registered fictional cases at 0 model calls, script selection local in every case, at an exact measured spend of $0.10335210 — every figure verified by a project-internal audit, including a separate recomputation of the machine records, with zero material discrepancies.

Labels used in the demo.
LIVE = a real model response from the controlled Phase 1 experiment. RECORDED LIVE = a real model response recorded earlier and replayed. MOCK = a demo stand-in for the live model. FALLBACK = the system honestly reporting that no approved script matches. The video's narration is a produced voiceover of the approved script; the video is a prerecorded demonstration and is labeled as such throughout. This project was built by Albert Tumblin III with AI assistance (Muse, Meta's AI agent) used as a research and development worker under the founder's direction; the framework, decisions, and approvals are the founder's.

Limitations, stated plainly.
The Phase 1 agreement labels are provisional and single-rater — never adjudicated — and human quality ratings (H1–H4) were not collected, so no accuracy or superiority claim is made. One comparison arm (Arm B) used a self-contradictory prompt and is reported only as a confounded baseline, never as an ablation. The safety screen is a fixed, registered rule set — a bounded list, not a safety guarantee — and its wording remains provisional pending independent review. All scenarios are fictional; no real child or parent data is used anywhere. The budget ceiling cited ($120) is the project's own application-side ledger control; the provider's $150 spend notification played no role.

Verify it yourself.
Public repository (application code, PROVENANCE.md, and the curated Phase 1 evidence package): https://github.com/GoldSpur/ccc-learning-lab-hackathon — Live demonstration (Cloudflare Access email-PIN login; any email you control works): https://lab-staging.createchildhoodconfidence.com/ — The Phase 1 experiment was preregistered and verified by a project-internal audit (including a separate recomputation of the machine records); its curated evidence package is published in the repository at https://github.com/GoldSpur/ccc-learning-lab-hackathon/tree/main/evidence and its figures are stated above at the strength that audit supports.

What this is not.
The Learning Lab is a framework-constrained parenting-education prototype. It is not a medical or mental-health advice system: it does not diagnose children or advise on mental health, and inputs matching its registered screening rules are stopped before any model call (the rules are a fixed, registered list — not comprehensive danger or mental-health detection). It is also not a generic education chatbot: there is no free-form chat, and the model never generates parent-facing text.

Hosted demo: https://lab-staging.createchildhoodconfidence.com/ — the site is protected by Cloudflare Access. Enter any email address you control, and a one-time login PIN will be emailed to you; enter the PIN and the application loads. Sessions last 24 hours. The hosted demo makes no live model calls: it replays a recorded live result and deterministic local paths, all labeled in the interface.

To try the live demo:
Open https://lab-staging.createchildhoodconfidence.com/ — you will be redirected to the Cloudflare Access login ("Log in to CCC Learning Lab - Hackathon Staging"). Enter any email address you control (no account or password exists; the access policy is intentionally open and labeled as such) and request the login code. Enter the one-time PIN sent to that inbox (it arrives within a minute and is valid for about 10 minutes). The Learning Lab then loads in your browser. Your session lasts 24 hours; if it expires, repeat the same steps. Please start from the full staging URL above — the bare cloudflareaccess.com root shows an unrelated "App Launcher" notice.

## More from Meta Global AI Developer Hackathon

- [Project Saiga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1274)
- [ClearFirst](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1275)
- [EarCaster](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1276)
- [LoadMaster (Ghost Developer Studio)](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1278)
- [Breakfast](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1279)
- [DPM Agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1280)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1277. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt