---
title: "Evidence-First Interfaces: Pairing REA with Vercel UI SDK"
description: "How REA evidence and Vercel UI SDK controls work together, plus a published settings-screen benchmark where the bridge used 76% fewer tokens than Base UI with browser tools."
canonical: "https://askgina-blog.pages.dev/posts/evidence-first-ui"
---

# Evidence-First Interfaces: Pairing REA with Vercel UI SDK

[HTML version](https://askgina-blog.pages.dev/posts/evidence-first-ui)

Published November 20, 2026 · REA Research × UI Systems

How REA evidence and Vercel UI SDK controls work together, plus a published settings-screen benchmark where the bridge used 76% fewer tokens than Base UI with browser tools.

**Abstract.** Reverse engineering produces graphs of disassembly, memory reads, IPC hops, and residual unknowns. Static blog prose flattens that structure. By pairing REA’s evidence-backed analysis model (REA records every tool observation as an immutable Evidence record and tracks unresolved gaps as explicit residual unknowns rather than silent assumptions. Source: [github.com/morluto/rea](https://github.com/morluto/rea).) with Vercel UI SDK’s accessible, agent-addressable primitives (@ui-sdk/core separates focus, keyboard, and ARIA state from styling, and registers every control with createAgent() so agents drive the same UI as humans. Source: [ui-sdk.dev/docs/agents](https://ui-sdk.dev/docs/agents).), every figure in this report becomes an inspectable instrument for both human readers and coding agents.

## Case I: DX-Ball Sound-Panning Reconstruction

In the DX-Ball showcase (DX-Ball 1.07 checkpoint, 7 October 2026 (REA 4.1.0): 3,205 oracle comparisons checked return values. A separate VC4.0 replay matched 63 compiled bytes. Source: [rea.tools/showcase/dx-ball](https://rea.tools/showcase/dx-ball/).), REA traces the x86 DirectSound pan routine at `0x00406400`. At `0x406409`, `MOV EAX, [EBP+0x8]` loads the horizontal coordinate. The routine then multiplies by `1.5625`, subtracts `500.0`, applies the pan scale, and truncates via `__ftol`. Step through the assembly and probe the reconstructed pan formula below. The browser runs the formula in JavaScript, not the original x86 binary.

Interactive figure: [DX-Ball Sound-Panning: x86 Assembly to Verified C](https://askgina-blog.pages.dev/posts/evidence-first-ui#dx-ball-trace).

## Why Headless + Agent-Ready Matters

Vercel UI SDK separates interaction state, roving focus, and ARIA semantics from visual styling (Vercel Labs reports three-run medians for one project settings task. The benchmark compares the agent bridge with browser tools that read the accessibility tree and then click or type. Source: [ui-sdk.dev/docs/agents](https://ui-sdk.dev/docs/agents).). Because every control registers with `createAgent()`, an agent can inspect and drive the same figures a human reader uses. The published benchmark below measures the bridge on a separate task, not on this blog or on REA investigations.

Interactive figure: [Browser-Only vs. UI SDK Agent Bridge (3-Run Medians)](https://askgina-blog.pages.dev/posts/evidence-first-ui#agent-bridge).

## Case II: Notion Desktop 7.6.1 Clipboard IPC Bridge

Static JavaScript analysis in REA does not require a native disassembler. Inspecting Notion Desktop’s `app.asar` (REA 4.1.0 located the exposed API and renderer IPC call in the desktop preload. The clipboard format comes from separate cached web-asset probes. Source: [rea.tools/showcase/notion](https://rea.tools/showcase/notion/).) maps the full trust boundary from renderer preload wrappers to the sender-checked `ipcMain.handle` clipboard sink.

Interactive figure: [Notion Desktop 7.6.1 Clipboard Write IPC Chain](https://askgina-blog.pages.dev/posts/evidence-first-ui#notion-ipc).

## Target Capability Matrix & Investigation Trace

REA routes each artifact to a specialized static reader, bring-your-own disassembler, or live runtime bridge (JS/Electron, .NET, and archive inspections run without native engines; deep native decompilation requires a local Ghidra, Hopper, or Rizin backend. Source: [github.com/morluto/rea](https://github.com/morluto/rea).). Filter the target matrix by engine requirement and inspect the investigation waterfall below (The 5-span DX-Ball walkthrough follows the published showcase. Its 940 ms timeline uses synthetic timings, not measured host latency. Source: [rea.tools/showcase/dx-ball](https://rea.tools/showcase/dx-ball/).).

Interactive figure: [REA Target Surface & Investigation Span Waterfall](https://askgina-blog.pages.dev/posts/evidence-first-ui#target-coverage).

## Notes & References

1.  Vercel Labs (2026). UI SDK agent docs, project settings screen benchmark (vercel-labs/ui, benchmarks/agent). [ui-sdk.dev/docs/agents](https://ui-sdk.dev/docs/agents)
2.  REA Project (2026). “Reverse Engineer Anything: One MCP for reverse engineering across binaries, applications, and runtime behavior (v6.8.0, 229 tools).” [github.com/morluto/rea](https://github.com/morluto/rea)
3.  REA Showcase. DX-Ball 1.07 project checkpoint, 7 October 2026 (REA 4.1.0). [rea.tools/showcase/dx-ball](https://rea.tools/showcase/dx-ball/)
4.  REA Showcase (2026). Notion Desktop 7.6.1 clipboard IPC trace (REA 4.1.0) and separate cached web-asset probes. [rea.tools/showcase/notion](https://rea.tools/showcase/notion/)
5.  Editorial design reference: SID-2 Technical Report (three-column reading grid, inline figure toggles, margin sidenotes). [sid.ai/research/sid-2-technical-report](https://www.sid.ai/research/sid-2-technical-report)
