Wingspan Labs

Research & Development

Wingspan Labs publishes position papers, evidence reviews, and practical tools for people making consequential decisions about websites, workflows, accessibility, and responsible uses of AI. Each publication names its purpose, evidence boundaries, and limitations.

R&D library

Current work

The newest work appears first. Each entry names its publication type or active-research status so readers can see what it is—and what it is not.

Playable experiment · September 4, 2026

Little Wings

A green-cheek conure, a garden, and one sentence that started it all. Play here—no sign-in needed.

Choose “Let’s fly” to begin. Space, ↑, or W flaps; hold to climb and release to glide. On touchscreens, tap and hold. P pauses or resumes; M changes sound.

Accessibility limitations: Some pause-screen text has low contrast. Screen readers may not correctly announce remaining lives. The on-screen sound button cannot be tapped while paused; press M on a keyboard, or resume before changing sound.

The game needs JavaScript and uses visual movement. Keyboard controls and status messages do not provide a nonvisual way to avoid obstacles. Use Tab to leave the game. Score and sound preferences reset when this page reloads. Kevin’s story is available without JavaScript.

One sentence, two conures, and what comes next

Kevin’s story

My creation-time estimate: just over 10 minutes.

On the afternoon of September 4, 2026, Wingspan Labs got to start using OpenAI GPT 6 Astra in Codex. I asked for this game in one sentence, then used another to ask for a security-flaw check and the package that lets you play it here. There were follow-up clarifications and automated work, too—not just two messages for the entire process.

I’m a huge fan of Green Cheek Conures and have two at home. A game from a single sentence, looking this good? Someone with a thing for elves, sardines, and Australia who collects Garbage Pail Kids could ask for a game with all those elements. Wow.

If this was this easy and fast, what is the future of gaming? Personally, I’m interested in an AI story being generated as I play: townsfolk with rich backstories who don’t mind talking about them, and side quests that emerge if the dialogue leads that way. That’s a hope for future games, not a feature of Little Wings.

The recorded creation work: 4,545,684 coding input tokens and 55,769 output tokens, including the art helper. Tokens are the pieces of information the model processes and produces. Of the input, 4,355,968 tokens were cached (reused) and are already included—not an extra amount.

These creation-only coding counts exclude the later security audit, coordination, and image-generation internal usage. They are not a total for every system or a cost estimate.

Read the game’s third-party software licenses (text file)

Position paper · August 27, 2026

Custodial Passkey Agents: A Pattern for Reversible AI-Agent Succession

An author-defined operating pattern for preserving a predecessor agent's useful context without restoring its authority. Includes a plain-language explanation, fictional example, safeguards, glossary, and annotated sources.

Complete HTML publication · Tagged PDF · 10 pages

Read the complete position paper Download the tagged position paper PDF (10 pages)

Practical tool · Conversation guide

Agent Continuity Risk Checklist

A one-page Yes / No / Unknown guide for examining ownership, source custody, authority limits, verification, and closure. It is not a score, certification, compliance assessment, or security guarantee.

Complete HTML checklist · Tagged PDF · 1 page

Use the complete HTML checklist Download the tagged checklist PDF (1 page)

Active R&D initiative · July 21, 2026

Wingspan Command Center

An internal research initiative exploring how governed AI-assisted work can preserve context, limit authority, document decisions, and produce verifiable handoffs without obscuring human responsibility.

Active research · Not a public software product

Explore the Wingspan Command Center initiative

Rapid systematic evidence review · July 20, 2026

Automated Accessibility Scores Are Not Accessibility Evidence

What automated web testing can detect, what it misses, and how organizations should interpret results without turning a partial measurement into a conformance claim.

HTML summary · Tagged PDF · 12 pages

Read the accessibility evidence review