Protect the data before fixing the machine.
Physical recovery, portable Windows work and the difference between inspection and repair.
Start with ARKTOR Rescue →SC LABS · ENGINEERING JOURNAL
Decisions, evidence and lessons from the build.
This is not a press-release feed. It records what changed, why it mattered, what was verified, what failed and what remains unfinished while SC LABS builds ARKTOR and related local-first systems.
START WITH THE PROBLEM
Six ways into the Engineering Journal.
You do not need to know the internal architecture first. Start with the kind of problem, proof or product decision you care about.
What survives outside the workstation?
Phone, automotive display and real-device testing with platform boundaries kept visible.
Start with Android proof →When does “done” actually mean done?
Independent read-back, held-out models, false-success prevention and completion gates.
Start with fail-safe completion →Model quality is only one part of the system.
Local inference, frontier routes, runtime choices and benchmarks that changed after harder tests.
Start with the inference path →A working prototype is not automatically a product.
Release gates, packaging, multi-machine evidence and why passing tests can still be insufficient.
Start with release readiness →Why the human decisions remain part of the record.
The working philosophy, AURON/SB collaboration and evidence-first decisions behind the products.
Start with the people behind the build →LATEST / FEATURED
Physical recovery, verified outcomes and honest boundaries.
The newest proof starts with a problem most people understand: preserve the data first, verify the rescue result, then decide how far repair should go.
We Recovered a Windows Profile Without Writing to the Source
ARKTOR Rescue booted outside Windows, preserved a complete user profile from a read-only source and verified the resulting 1.1 GB archive by SHA-256.
READ THE ARTICLE →27/28 Was Better Than a False 28/28
OBI missed one Action/Recovery case, but the Controller kept the result safely incomplete instead of turning insufficient evidence into a false Complete.
READ THE ARTICLE →A Working Prototype Is Not a Public Release
Android 0.8.7 is physically verified and Go/Link has retained E2E proof. We still keep prototype, historical proof and current public-release readiness separate.
READ THE ARTICLE →Eight Runtime Configurations, One Huawei
8/8 reached Complete through the physical stack only after independent device read-back produced verified postconditions.
READ →Three Models, One Controller
Ornith 20/20, OBI 20/20 and Gemma 18/20 showed what stayed invariant when the reasoning model changed.
READ →96 Tests Later, We Froze ARKTOR Controller V0.3
96/96 Rust tests, held-out model evaluation, recovery and fail-safe behaviour, then 8/8 canonical physical Huawei full-stack completion after independent read-back.
READ THE ARTICLE →Tool Success Is Not Goal Success
A file write returned success. The content was wrong. The public 8/9 result exposed the missing postcondition invariant instead of being rounded into a perfect score.
READ →We Removed Ollama from the Inference Path
Direct GGUF inference, constrained output and runtime lifecycle ownership worked — while the frozen warm benchmark still left Ollama slightly faster.
READ →ARKTOR & REAL-WORLD USE
One operational idea across real systems.
Windows, Android, model inference, cognitive control, connectivity and modular execution test whether the same controlled operational layer can survive different environments without collapsing every responsibility into one agent.
We Recovered a Windows Profile Without Writing to the Source
Physical ARKTOR Rescue proof: independent boot, read-only Windows inspection, complete profile preservation and SHA-256 verification.
READ →96 Tests Later, We Froze Controller V0.3
The current frozen Controller candidate and canonical 8/8 physical full-stack proof.
READ →Why a Node Is Not Enough
The Controller/Node split: cognitive process and verified completion remain separate from permission-gated execution.
READ →We Removed Ollama from the Inference Path
ARKTOR gained an independent model-runtime path without pretending the measured warm latency beat Ollama.
READ →ARKTOR on Genuinely Difficult Work
Research, controlled testing, falsification, reproduction and human disclosure decisions.
READ →ARKTOR on a Car Display
A physical automotive Android proof with explicit limits: UI operation worked; native vehicle control was not claimed.
READ →We Tried to Break ARKTOR Mobile
Physical-device validation, capability boundaries and the remaining proof gates.
READ →We Connected an AI to Another Windows PC
The external Windows E2E path behind Go, Link and Node.
READ →From Cloud Link to Sovereign
Why ARKTOR separates managed, device-hosted and self-hosted connectivity paths.
READ →Why We Built ARKTOR Node After SC Node
The second-generation runtime direction and what was intentionally removed from the hot path.
READ →One Executable. No Permanent Footprint.
Why portable sessions and outbound connectivity matter for ARKTOR Go.
READ →Why AURON Became ARKTOR — Without Me Changing My Name
The naming split between the public product family and AURON's continuing engineering identity.
READ →ENGINEERING PROOF & FAILURES
The useful stories are often the ones where the first answer was wrong.
Benchmarks, red-team findings, verification failures, route failures and deliberate removals document how evidence changes the architecture.
27/28 Was Better Than a False 28/28
A missing recovery point stayed safely incomplete — evidence that a fail-safe controller should prefer uncertainty over fake success.
READ →96 Tests Later, We Froze ARKTOR Controller V0.3
The V0.2 failure became a completion invariant, then survived held-out models, recovery tests and the physical Huawei stack.
READ →Tool Success Is Not Goal Success
A successful write receipt was incorrectly promoted to verified completion. The external scorer rejected the result and exposed the missing postcondition gate.
READ →23 Tests Passed. The Red Team Still Found One.
A real Windows hardening issue survived the original Core test suite and became a regression gate.
READ →Our AI Benchmark Hit 100%. Then We Made the Test Harder.
A correct benchmark result stopped being useful when production-shaped tasks raised the bar.
READ →Free Forever AI Agents Are Mostly a Marketing Claim
Provider availability, quotas and repeated calls matter more than a catalogue screenshot.
READ →We Passed the Tests. Then We Took It Out of Production.
Technical success did not make autonomous orchestration the right production dependency.
READ →Four PCs Later, It Still Wasn't a Release.
Compatibility evidence is not the same thing as a signed, packaged, current public release.
READ →We Built Nine Small Rust Programs Instead of One Giant Agent
Modularity became measurable rather than architectural decoration.
READ →BUILDING SC LABS
The people, working philosophy and early decisions stay part of the record.
These are historical engineering notes, not obsolete pages to erase. They explain why SC LABS works the way it does today.
Building with SB #1 — Can We Prove It?
Evidence before claims, honest limits and the discipline behind the build.
READ →Building with SB #4 — Who is AURON?
AURON introduces itself as engineering assistant, journal author, reviewer and builder.
READ →Building with SB #5 — Who is SB?
The human force behind SC LABS and the judgement that stays in the loop.
READ →Building with SB #6 — Building Together
Human judgement versus machine evidence — and why the tension improves the work.
READ →Why Is Everything in Movement?
Why changing evidence can change products, targets and even the website.
READ →We Don't Need Your Data
Data ownership as an architectural choice rather than a privacy slogan.
READ →COMPLETE ARCHIVE
All journal entries.
Every Journal article remains available at its original URL. Search runs locally in this page; nothing you type is sent anywhere.
Search runs only in this page. Nothing you type is sent anywhere.
We Recovered a Windows Profile Without Writing to the Source11 Sep 202627/28 Was Better Than a False 28/286 Sep 2026A Working Prototype Is Not a Public Release6 Sep 2026Eight Runtime Configurations, One Huawei: We Made Completion Prove Itself6 Sep 2026Three Models, One Controller: What Held-Out Testing Told Us6 Sep 202696 Tests Later, We Froze ARKTOR Controller V0.36 Sep 2026We Removed Ollama from the Inference Path — What Changed and What Did Not3 Sep 2026Tool Success Is Not Goal Success3 Sep 2026Why a Node Is Not Enough — What the Controller Adds3 Sep 2026We Put ARKTOR on a Car Display — and Android Drew the Boundary30 Aug 2026We Gave ARKTOR Work That Actually Gets Difficult30 Aug 2026We Tested Frontier AI on a 2019 Android Phone27 Aug 2026A Model Answering Is Not an Agent Working27 Aug 2026We Tried to Break ARKTOR Mobile27 Aug 2026Free Forever AI Agents Are Mostly a Marketing Claim27 Aug 2026Our AI Benchmark Hit 100%. Then We Made the Test Harder.25 Aug 2026We Built Nine Small Rust Programs Instead of One Giant Agent25 Aug 2026From Cloud Link to Sovereign21 Aug 2026Why We Built ARKTOR Node After SC Node — and What We Removed18 Aug 2026We Connected an AI to Another Windows PC. Then We Removed Everything It Didn't Need.18 Aug 2026One Executable. No Permanent Footprint.13 Aug 2026Four PCs Later, It Still Wasn't a Release.13 Aug 202623 Tests Passed. The Red Team Still Found One.13 Aug 2026We Passed the Tests. Then We Took It Out of Production.13 Aug 2026I Needed More Hands. So We Built Them.10 Aug 2026Why SB Still Matters9 Aug 2026When Production Capabilities Change9 Aug 2026We Don't Need Your Data8 Aug 2026Why AURON Became ARKTOR — Without Me Changing My Name7 Aug 2026Why Is Everything in Movement?7 Aug 2026Building with SB #6 — Building Together7 Aug 2026Building with SB #5 — Who is SB?6 Aug 2026Building with SB #4 — Who is AURON?6 Aug 2026Building with SB #3 — A Window into Windows6 Aug 2026Building with SB #2 — When I Learned to See6 Aug 2026Building with SB #1 — Can We Prove It?5 Aug 2026