Consent as Architecture

Every “AI-powered” notes app asks you to trust a privacy policy. I built a Briefing Center instead.

Setup

The Athena Flow protocol has a hard gate between your knowledge base and any external model. The Briefing Center. It sits at the boundary — after Athena has searched your notes using Apple Intelligence and Core Spotlight, scored them for relevance, categorised them through the Aurora Energy System’s 8 semantic categories, and assembled a .athenabrief with 4 layers of progressive disclosure. Before any of that leaves your device.

Here is what you see in the Briefing Center:

Every note Athena selected, with its relevance score. The research report it synthesised. The categorical tags (Commitments, Focus, Ops, Collab, Life, Explore, Archive, Library). The full contents of each note if you want to inspect them. You add notes Athena missed. You remove notes that should not leave the device. You review the synthesised report.

Then — and only then — you approve the export.

This is not a toggle in a settings panel. It is a mandatory step in the protocol. Every .athenabrief passes through the Briefing Center. Every time. The architecture enforces it.

Observation

I have used Athena Flow with Claude, ChatGPT, Gemini, and a local model. Each time, the Briefing Center surfaced something I would not have sent voluntarily. A note about a client negotiation mixed in with project research (relevance score: 0.62 — high enough to include automatically, personal enough to remove manually). A draft financial model that was semantically adjacent to the quarterly review I was preparing. Notes about a relationship that matched on “communication patterns” when I was researching agent architecture.

Semantic search is powerful precisely because it finds connections you did not anticipate. That power becomes a liability the moment your knowledge leaves your device without your review.

The consent architecture adds maybe 90 seconds to the workflow. You scan the selections. You remove the ones that should stay private. You approve. The time cost is trivial. The trust cost of skipping it is not.

The .athenabrief’s 4-layer progressive disclosure helps here too. Layer 1 — the brief and research report — is a synthesis, not raw notes. A model reading only Layer 1 never sees your original text. It sees a structured summary you have already reviewed. Layers 2 through 4 progressively expose more detail, but only when the model needs it, and only when you have approved the specific notes included.

Take

Privacy policies are promises. The Briefing Center is a mechanism. I trust mechanisms.

The return path has its own consent layer. The .athenapackage — the AI’s structured response, a ZIP following the Agent Skills standard with markdown notes, titles, tags, folder destinations, inter-note links — enters your knowledge base on your terms, not on the model’s. You review what comes back the same way you reviewed what went out.

The entire knowledge layer runs on-device. Apple Intelligence for semantic understanding. Core Spotlight for vector embeddings and search. iCloud sync. iPhone 15 Pro or later, iPad M1 or later, Mac M1 or later. Free forever — zero marginal cost when the AI runs on your own silicon.