Prioritise protection truth and recovery
Existence is established. The current client has concrete seams that can produce false state, trapped recovery and unsafe lifecycle ambiguity.
Product immersion · evidence through 27 Aug 2026 · n=1
One Mac established real failures in protection truth, lifecycle handling, crash stability and felt responsiveness. It did not establish how often they happen, how many customers they affect, or how much value they destroy.
Four different states. One missing source of truth.
Existence is established. The current client has concrete seams that can produce false state, trapped recovery and unsafe lifecycle ambiguity.
The repair sequence is clear enough to build and test. Production incidence and backend contribution remain unknown.
Keep it inside Retention Economics. Do not promote it until prevalence, support load and retained value are measured.
01 · Field evidence
Each card separates the observation from the claim it can support. “Observed” means the event happened in this study; it is not a fleet-wide rate.
Traffic went direct while the app still said Connected. In another state, the app said Disconnected while the tunnel remained active.
Both preserved reports show EXC_BAD_ACCESS / SIGSEGV at 0x48 with the same first 18 faulting frames.
A normal quit left tun.appex, the tunnel and VPN traffic running. During total Wi-Fi loss, the stale Connected state blackholed traffic.
The purchase succeeded before the app made plan state clear. State converged after the tour and the user could connect.
A London residential exit was classified as Corporate/Hosting. A Lithuanian consumer-ISP exit labelled Vilnius resolved to Kaunas in two services.
The app survived 17m52s disconnected and one exact blank-area Lithuania row click. No new crash report appeared.
02 · Same-hotspot A/B
One direct-versus-Vilnius-residential comparison on the same device and hotspot. Useful as a mechanism signal; not a population benchmark.
RPM = round trips per minute under load. Higher is better. The residential run retained 88% of download throughput but only 16% of direct responsiveness.
03 · Static source audit
Public client snapshot 09fa557c…; public node snapshot c45527a…. Static audit identifies mechanisms and repair seams, not production frequency.
A failed corroboration lookup is coupled to tunnel liveness instead of resolving metadata in the background.
The client can disable control while connecting, ignore a retry flag and poll teardown without delay, deadline or cancellation.
Native acceptance, interface state and usable internet are not one serialized truth. “Connected” can arrive before protection is established.
Observers can miss initial or early state, listen too broadly, or reconcile app logout and native tunnel state incompletely.
This is a direct security defect and the smallest immediate patch: remove the log and add a secret-log regression test.
04 · Product decision
The strongest product shape is a deterministic, local, outcome-verified operator: detect failure, block unsafe fallback, make the safest equivalent repair, and prove route, public IP, DNS and traffic before declaring protection.
Customers should almost never need to talk to Mysterium.
An LLM may translate intent, explain evidence and assemble escalation. It must not decide whether traffic may leak or invent a protection-critical repair.
Protection truth and recovery are co-primary gates with entitlement, billing and refund truth.
Provisional #5 challenger only. Keep it inside #2 Retention Economics.
Then serialize state truth before changing teardown and fallback behaviour.
Delight is the bounded product loop that could act on these failures: detect → contain → recover → verify, then explain only when useful. It stays inside #2 Retention Economics; bounded pilot only; not independently promoted. Open brief →
05 · Promotion contract
No more breadth for its own sake. Every remaining test must establish causation, calibrate an instrument, or size the affected population.
Evidence boundary
06 · Machine record
Visible on purpose. The embedded payload and brief.json carry the same verdict, limits and next gates for agents.
Decision: prioritise protection truth and deterministic recovery; do not promote a standalone asymmetric opportunity.
Evidence cutoff: 2026-08-27. Source is the current Lee-side product immersion record; wider shared-repository absorption is not claimed.
Release policy: living latest during the active study; every deployed change remains recoverable through Git.
Review status: Talos self-review. Claude Fable 5 was quota-unavailable; no Fable review is claimed.