Inspiration
Meaningful family stories are usually split across camera rolls, voice notes, and the memory of one person who knows why an ordinary object mattered. Storage is not understanding, and generation can quietly invent details. We asked: can AI help a family enter a true story without replacing the person who told it?
What it does
Keepscape turns three to five photos, story text, and optional original audio into a source-grounded, walkable memory space.
GPT-5.6 maps claims to supplied material, separates supported facts from uncertainty, and produces a strict source-linked story blueprint. Before anything becomes an exhibit, the storyteller must confirm or preserve uncertain details and review every displayed claim and generated story field.
The host then compiles the reviewed photo sources into a safe spatial diorama. Codex sees only opaque, prose-free hotspot and photo tokens and returns a bounded interaction plan. It cannot see or author visitor copy, claims, media, asset paths, or original IDs. A final human gate displays the mechanic, targets or order, and preserved wording before entry.
Our signature Evidence Lens desaturates generated scenery while keeping the selected cited photo view legible. Every meaningful object can open a source card with an audio timecode or explicit human confirmation. The result is clearly labeled as a generated spatial interpretation, never a scan or reconstruction.
The credential-free judge build includes two distinct exhibits:
Lantern Lane, 1998: walk and turn through a three-photo CSS-3D corridor, collect three source-backed lantern details, and inspect each through Evidence Lens. Four Moves at the Repair Bench: replay a remembered bicycle-repair ritual in its source-backed order; a wrong step resets the sequence.
All included photos and narration are clearly labeled AI-generated fictional demo material.
How we built it
GPT-5.6 Structured Outputs creates typed claims, uncertainty, hotspots, source references, and narrative copy. It does not author geometry. A human grounding desk blocks the build until uncertainty and generated language have been explicitly reviewed. The host compiler creates canonical photo planes and converts only reviewed mechanic members into opaque tokens. The Codex SDK runs read-only and without network access in an ephemeral, sanitized environment. Its dynamic JSON schema permits only the approved mechanic, exact token enums, an allowlisted spatial preset, and booleans—no free-text output. A strict host gate rejects changed mechanics, reordered sequences, and missing, extra, duplicate, or unknown tokens before rebinding original IDs. A responsive Next.js, React, TypeScript, Zod, Motion, and CSS-3D runtime renders the validated package without evaluating generated JavaScript, CSS, transforms, or shaders. The public experience uses a clearly labeled, deterministic verified replay so judges need no API credentials; live GPT-5.6 and Codex paths remain real, tested, and documented.
Codex also accelerated the engineering workflow: it helped research idea collisions, build the GPT pipeline, implement two mechanics, design the product interface, generate tests, audit trust boundaries, and repair browser-QA failures. Our public decision log and redacted live-run receipt distinguish Codex contributions from human product decisions.
Challenges
The hardest challenge was preventing beauty from outrunning truth. A normal generative pipeline fills narrative gaps; Keepscape instead treats every factual claim as a typed reference graph and fails validation if an exhibit object loses its source.
The second challenge was giving Codex meaningful creative control without exposing family material or accepting arbitrary generated code. The opaque-token interaction language and exact-set host validation made that boundary concrete.
Accomplishments
A walkable multi-photo memory corridor plus a visibly different sequence mechanic from one evidence contract. Provenance remains inside the emotional experience instead of living in a separate report. Supported fact, uncertain memory, and generated interpretation are distinct product states. The judge experience is credential-free and usable with keyboard, touch, and reduced motion. The repository includes setup instructions, sample provenance, architectural decisions, 62 unit tests, browser flows, and a redacted successful Codex SDK receipt.
What we learned
Provenance does not need to feel like compliance. A cited photo region or a few seconds of original voice can be the most emotionally powerful part of the experience. A small typed interaction language also lets Codex be genuinely useful while keeping the result portable, inspectable, and safe.
What's next
Private encrypted family spaces, collaborative fact confirmation, more interaction primitives, and exportable offline exhibits. Museums and oral-history projects could reuse the same grounding contract for community collections.
Built With
- cloudflare-workers
- css-3d
- gpt-5.6
- motion
- next.js
- openai-codex-sdk
- openai-responses-api
- playwright
- react
- typescript
- vitest
- zod