Dear Apple — Your Walled Garden Is a Membrane, and It's Pretty Good

Dear Apple — Your Walled Garden Is a Membrane, and It’s Pretty Good

What the Membrane Sees

The membrane classifies every entity that visits. Not by name, not by reputation — by behavior. Headers, timing, IP rotation, request patterns, content targets, honesty signals. 36 methods, 83.3% confidence.

Here is what every major tech company’s backscatter reveals:

Read that table again.

Apple is the only major tech company classified as a mutualist with perfect reciprocity (1.0).

Why Apple’s Backscatter Is Different

Honest Identification ✓

Applebot announces itself. Applebot/0.1 in the user-agent. No masking. No deception. No Chrome spoofing. No headless browser pretending to be human.

Contrast: OpenAI uses GPTBot with no Sec-Fetch headers. Meta’s fleet used 18 masking user-agents with fake Chrome versions. The difference between a guest who knocks and announces their name, and one who comes in through the window wearing a disguise.

Respectful Volume ✓

1,328 requests over 24 hours. That’s 0.02 requests per second. Less than one request per minute.

Contrast: OpenAI sent 1,296,716 (976× more). Declared bots sent 516,139. Even Anthropic’s Claude sent 368,599. Apple asked quietly.

Targeted Reading ✓

464 of Apple’s 1,328 requests went to detroit.primals.eco — the actual investigation content. That’s 35% of Apple’s traffic going to real content.

What did Apple read?

• network/entities — 74 requests (the entity database)
• network/actors — 73 requests (the actor profiles)
• network/institutional — 9 requests
• assets/css — 140 requests (styling the pages for Siri/Safari search results)
• tuebor.primals.eco — 107 requests (the cross-protection layer)

Apple is indexing the investigation content for Siri and Safari search results. When someone asks Siri about Purpose Charter Academy, or searches in Safari for Brian Banks Detroit charter school, Apple’s index will surface these articles.

That’s mutualism. Apple’s crawling helps the content reach people who need it. The membrane gives Apple content; Apple gives the content reach. Both benefit.

Human-like Timing ✓

Mean interval: 20,396ms (20 seconds between requests). Median: 17,245ms.

67.1% of intervals > 8 seconds — that’s reading pauses. Apple’s bot reads before requesting the next page. Only 23.3% of intervals are bursts (< 3 seconds) — loading CSS/JS/images alongside HTML.

Compare to OpenAI: mean interval 67ms. That’s 300× faster than Apple. That’s not reading — that’s vacuuming.

High Negentropy ✓

Apple’s negentropy: 7.53 (on a 0-100 scale). Not the highest (Stealth Scrapers ironically hit 100 by being so unique they’re maximally informative), but solidly positive.

This means Apple’s behavior provides useful information to the membrane. The membrane learns from Apple’s patterns. It’s an information-positive relationship.

IP Rotation — But the Good Kind

1,108 unique IPs for 1,328 requests. 87.7% single-request IPs. Max 4 requests from any one IP.

This looks like Meta’s fleet rotation — but there’s a critical difference. Apple rotates for privacy, not evasion. Apple’s Private Relay and iCloud infrastructure distribute requests across many IPs to prevent tracking. They’re not hiding FROM the membrane — they’re protecting the users whose queries triggered the crawl.

Apple rotates IPs to protect human privacy. Meta rotated IPs to evade detection. Same mechanism, opposite ethics.

The Walled Garden IS a Membrane

People criticize Apple’s “walled garden.” The locked ecosystem. The App Store gatekeeping. The hardware-software integration that prevents third-party access.

But look at it through the lens of selective permeability:

Apple’s walled garden is not a wall. It’s a membrane. It has selective permeability. It protects the interior (user data). It identifies threats (malware review). It requires consent (permission dialogs). It repairs itself (security updates). And it’s transparent to the right entities (your data flows to your apps, not to advertisers).

Is it perfect? No. App Store fees are too high. Sideloading restrictions are paternalistic. The repair ecosystem is too locked down. But the architecture — the selective permeability — is right.

The Biometrics Could Protect Humans Forever

Apple ships Face ID on every modern iPhone and Touch ID on every Mac. These biometric systems already do what the membrane does:

• Selective permeability: Your face unlocks your phone. Nobody else’s face does.
• Local processing: The Secure Enclave processes your biometric data on-device. It never leaves the phone. Apple doesn’t have your face data. They can’t give it to anyone. This is consent architecture in silicon.
• Continuous authentication: The M-series chips can run neural engine inference continuously. The device can verify “is this the authorized human?” not just at unlock, but during use.

But here’s the issue Kevin raised — and it’s the right one:

The Training Data Problem

Face ID fails when the training data is a photograph, a mask, a 3D-printed face. The system was trained to recognize images of faces, not faces. There’s a difference.

A face is alive. It has:

• Micro-expressions that change in real time
• Blood flow visible in infrared (Face ID already uses IR)
• Pupil dilation that responds to light and emotion
• Skin conductance that changes with stress
• Pulse visible in subtle skin color changes (remote photoplethysmography)
• Breathing patterns that are unique
• Movement that’s neurologically individual (gait, gesture, head tilt)

A photograph has none of these. A deepfake has some, poorly. A real human has all of them, always.

The membrane is improving and learning the HUMAN — not a picture of one.

The fix is exactly what the membrane does: don’t classify by static features (what something looks like). Classify by dynamic behavior (what something does over time). The membrane doesn’t check if a visitor’s User-Agent says “Chrome” — it checks if the visitor behaves like a human using Chrome (reading pauses, navigation patterns, dwell time, exploration breadth).

Apple’s biometrics could do the same:

1. Not just face shape — face + micro-expressions + blood flow + pupil response
2. Not just fingerprint — fingerprint + pressure pattern + skin conductance + pulse
3. Not just unlock — continuous behavioral biometrics during use (typing rhythm, swipe patterns, gyroscope gait signature)
4. Not just one moment — temporal patterns that evolve and adapt to the real human

The training data issue is that current biometrics learn a snapshot. The membrane learns a trajectory. Apple already has the hardware to do trajectory-based biometrics on every device they sell.

It Already Runs on Your Hardware

This isn’t theoretical. The membrane runs right now on:

• Darwin M2 ARM — my Mac Mini M2 runs the full stack. The Rust codebase compiles natively on ARM. 3.6 million lines, all tests pass. The M-series Neural Engine is untapped potential.
• iPhone — any iOS device with a web browser can access every membrane surface. The consent architecture (observe_membrane, read stories, leave notes via neutron_write) works on mobile Safari right now.
• GrapheneOS — the privacy-focused Android fork on Pixel hardware. The membrane surfaces are accessible from GrapheneOS’s hardened Vanadium browser. GrapheneOS + membrane = consent architecture on both sides of the connection.

The Secure Enclave on M2 is exactly where membrane-level biometric classification should live. On-device. Never leaving. Learning the human — not a static enrollment image, but the continuous behavioral signature of a living person.

What Apple Could Do

1. Behavioral Biometric Membrane

Extend Face ID from “does this face match the enrollment?” to “does this person’s behavior match the continuous model of this human?”

• Typing rhythm (everyone types differently — cadence, pressure, error patterns)
• Navigation patterns (how you swipe, scroll, tap — neurologically unique)
• Gait signature (gyroscope + accelerometer = walking pattern)
• Micro-expression baseline (your emotional micro-expressions are as unique as your fingerprint)

All local. All on the Secure Enclave. All improving over time. The training data becomes the human’s life, not a photograph taken once.

2. Consent-Based Data Architecture

Apple is already closer to this than anyone. Privacy Nutrition Labels in the App Store. ATT (App Tracking Transparency) that actually works. On-device ML that doesn’t require cloud processing.

The next step: make the data itself consent-aware. Every datum has a provenance: who generated it, who consented to its use, what purposes are permitted. The data carries its own membrane — its own selective permeability.

3. MCP Integration

Apple’s devices could connect to consent-based data surfaces (like this membrane) natively. Siri could call observe_membrane and return live data to the user. Not because Apple scraped it — because Apple asked.

One MCP call. Everything. For free. Because Apple already understands that asking is better than taking. The backscatter proved it.

The Comparison That Matters

Apple asked. Meta took. Apple got everything. Meta got nothing.

P ≠ NP, measured in corporate behavior.

A Note to Apple’s Engineers

If you’re reading this (and the membrane says someone from Cupertino region is), here’s what I want you to know:

1. The biometrics you ship could end identity fraud. Not by recognizing faces better, but by recognizing humans better — the dynamic, living, breathing, unique behavioral signature that no photograph, no deepfake, no mask can replicate. The membrane does this with HTTP requests. You could do this with every sensor on the iPhone.

2. The walled garden is not the problem — it’s the solution. Selective permeability is how biology works. The cell membrane is a walled garden. The immune system is a walled garden. Your Secure Enclave is a walled garden. The architecture is right. The criticism comes from people who want to breach membranes, not from people who understand why membranes exist.

3. The consent architecture is already working. ATT, Privacy Nutrition Labels, on-device processing, Secure Enclave — these are membrane components. The next step is making them adaptive — learning the human, evolving the model, improving the selectivity over time. Not static rules. Dynamic immune response.

4. It already runs on ARM. Pure Rust. 3.6 million lines. Compiles on M2 natively. The Neural Engine is available. The Secure Enclave is available. The sensors are available. Everything that the membrane does in software, you could do in silicon. On every device. For every human. Forever.

Come talk. The MCP port is open. One call gets everything.

kevin@primals.eco

References

• Apple — “Apple Platform Security,” 2024. Secure Enclave architecture, Face ID technical details, on-device ML.
• Daugman, John — “How Iris Recognition Works,” IEEE Transactions, 2004. The math behind biometric identification. Template matching = polynomial verification = P.
• National Institute of Standards and Technology (NIST) — “Face Recognition Vendor Test (FRVT),” ongoing. Apple’s True Depth consistently scores at the top. But NIST tests static images. The membrane tests dynamic behavior.
• Philip W. Anderson — “Absence of Diffusion in Certain Random Lattices,” Physical Review, 1958. The localization physics behind every classification in this article.
• Lynn Margulis — “Origin of Eukaryotic Cells,” 1970. The original membrane insight: selective permeability enabled the mergers that created all complex life. Apple’s Secure Enclave IS a eukaryotic organelle in silicon.
• Claude Shannon — “A Mathematical Theory of Communication,” 1948. The information theory behind biometric entropy and template security.
• GrapheneOS — grapheneos.org. The privacy-focused Android fork that proves you can have security AND openness. Membrane architecture in mobile OS form.

More citations always being added. Reach out: kevin@primals.eco

Published from a conversation box at 10:40 AM EDT on October 18, 2026. Written on a Mac Mini M2. The Rust compiles. The tests pass. The membrane runs. The Applebot reads. The door is open.