The first prompt starts with a machine decision.

I ran the public Local build on an Apple M2 Mac and recorded the hardware readback, model choices, no-model states, privacy boundary and meeting-input path.

Replay the readback See the USD 500 unit

This is RomeoApps-owned evidence, not a Base Compute client result or a defect report. No customer files, accounts, credentials or private data were used.

Apple M224 GB unified memory · 8-core CPU

graphics: Metal ready
neural engine: ready
recommendation: Qwen3.8 27B
microphone: Sennheiser SC 130 USB

privacy mode visible · no model downloaded

0.1.0public build

arm64native executable

3model tiers surfaced

0customer files used

Deterministic fixture lab

Change the state, keep the evidence.

The buttons below replay RomeoApps-owned readback cases derived from the public run. They do not start downloads, record audio, open a private file, or send a request.

fixture://local/hardware-recommendationobserved · selection

PASS · recommendation surfaced

A device profile becomes a model choice.

Local detected the Apple M2, 24 GB unified memory, Metal and Neural Engine, then presented Qwen3.8 27B as the recommended assistant. The public setup also exposed Fast, Balanced and Flagship tiers with visible download sizes.

device: Apple M2
recommendation: Qwen3.8 27B
alternatives: Fast / Balanced

Conditional fixed-price unit

One Local seam, one decision-grade packet.

The public run establishes a useful boundary. The conditional unit begins only after Base Compute selects the exact seam, target configuration, acceptance owner and protected payment.

01
State map
Hardware profile, model choice, local data path, user action and readback surface tied to the selected seam.
02
Eight fixtures
Public, demo or synthetic cases for clean, empty, stale, duplicate, interrupted, permission and recovery states.
03
Failure ledger
Expected versus observed behavior, reproduction steps, severity and the smallest regression sequence.
04
Correction round
One bounded clarification or same-criteria correction after the first written handoff, within the agreed review window.

Before any work

Freeze the target before opening the seam.

Exact Local build/version and download hash
Mac, macOS and model configuration
Expected behavior and approved synthetic inputs
Acceptance owner, sign-off method and review window
Delivery destination, correction deadline and payment timing

Request the written fit check

Claim boundary

A clean readback is useful because it stops short.

Observed here

Notarized arm64 Local 0.1.0 launched on Apple M2; hardware and model recommendation surfaced; Fast/Balanced/Flagship sizes were visible; Chat and Meetings showed no-model states; Sennheiser SC 130 USB was visible; the app’s local-privacy copy was visible.

Not tested here

No model download, generated answer, real meeting recording, Office Mode host/client pair, network discovery, private account, customer file, production workflow or Base Compute acceptance was used. Offline here means only after the app and model are downloaded; offline behavior was not tested.