Lab notes, not press releases.
We publish what we build and what we learn — architecture, method, results. Including the negative ones.
LAB NOTE // 001
Architecture of a local ACI: what we learned running an autonomous cognitive system on four GPUs at home
For five weeks in 2026, an autonomous cognitive system with persistent memory, gated self-improvement and a five-model deliberation council ran continuously on a home-lab server. The architecture, the numbers, and the honest limitations.
Jul 19, 2026 • 6 min
LAB NOTE // 002
Validation first: twelve hypotheses, zero survivors
We built a validation-first trading harness, tested 12 classic systematic strategies across FX, gold and crypto, and rejected every one with statistical evidence. Here is the method, the numbers, and what failure taught us.
Jul 19, 2026 • 6 min
LAB NOTE // 003
Hydra: composing specialized models instead of scaling one
Our systems aren't one model doing everything. A high-capacity core coordinates models dedicated to specific roles — orchestration, audio, vision — over an internal nervous system we call Tesseract. What we built, and what is still design.
Jul 19, 2026 • 6 min
LAB NOTE // 004
Embedded AI in a real operation: designing an assistant that acts, but only with permission
Atlas PM puts an AI operator inside a field-service platform — reading operational data, flagging what needs attention, and taking actions only a human confirms. What we built, and the honest limit of confirm-by-prompt.
Jul 19, 2026 • 6 min
LAB NOTE // 005
When the win rate lies: killing a prediction-market strategy with our own gate
An 80% win rate that meant nothing. We built a prediction-market research system, pre-registered a validation gate, and watched the gate catch our own false positive before a dollar of judgment rode on it. A rigorously tested negative.
Jul 21, 2026 • 6 min
LAB NOTE // 006
Making a publisher unable to lie: what changed after it recorded a post it never sent
Our social distribution engine once wrote a post to the database as published, with a fabricated platform ID, when no account was connected. The fix was not a patch — it was an explicit publish mode, the removal of the false-success path from all three network clients, and an atomic budget gate. The honest status: the service is deployed and tested, and no real post has ever left it.
Sep 6, 2026 • 7 min
TECH EVOLUTION // 08
We Benchmark Agents at 50%. We Deploy Them at 99%.
The field publishes agent time horizons at 50% and 80% reliability. Serious production decisions are made at 98%+ — and METR says in writing that the 99% number cannot be fit at all with the benchmarks that exist. The public ruler stops exactly where the decision begins.
Sep 2, 2026 • 8 min
TECH EVOLUTION // 07
The Attack Surface Moved Into the Tool List
A backdoored LiteLLM release, a 9.6-severity flaw in mcp-remote, and 200,000 vulnerable MCP server instances: the 2026 CVE record shows prompt injection and tool poisoning have moved from lab scenario to production incident — faster than governance is catching up.
Aug 26, 2026 • 9 min
TECH EVOLUTION // 06
The Bottleneck Moved to the Spec
Developers now spend more time reviewing AI-generated code than writing it — 11.4 hours a week versus 9.8, a reversal a fresh Q1 2026 survey confirms. The industry's answer is spec-driven development, adopted by eight major tools in a year. Early evidence says it hasn't proven itself: agents follow specs at roughly 70% fidelity, not the 95%+ that would actually compress the review burden instead of just relocating it.
Aug 20, 2026 • 9 min
TECH EVOLUTION // 05
Smarter Models, Same Spreadsheet
Model capability rose fast and verifiably — Humanity's Last Exam from under 10% to 38.3% in a year, SWE-bench Verified to 76.8% by February 2026 — while value capture inside organizations did not keep pace; buyer data now demands proof of P&L over productivity anecdote, moving the race from who has the smartest model to who can show the result.
Aug 13, 2026 • 10 min
TECH EVOLUTION // 04
You Can't Audit a Belief
Agent capability is scaling like software while accountability still scales like headcount — 82% of executives believe policy covers them and only 21% have runtime visibility into what their agents are doing; the bottleneck of this phase of automation isn't what agents can do, it's whether anyone can enumerate (and answer for) what they CAN do.
Aug 5, 2026 • 9 min
TECH EVOLUTION // 03
Attackers Don't Have to Be Right
AI's security asymmetry isn't offensive capability — offense tolerates being wrong and defense can't, so the cost of verification per action, not model quality, decides where autonomous defense lands.
Jul 26, 2026 • 9 min
TECH EVOLUTION // 02
Code Got Cheap. Trust Didn't.
AI scaled code generation ~3x but verification scaled 0x, migrating the bottleneck (and the senior-engineer cost) into review/release — and standard metrics measure the wrong scarcity.
Jul 22, 2026 • 8 min
TECH EVOLUTION // 01
Every Agent Is a Principal Now
Machine identities outnumber humans 109:1 and 79 of them are AI agents; identity is the unglamorous layer that decides whether enterprise agent adoption compounds or stalls.
Jul 19, 2026 • 7 min
WHEN, NOT IF // 04
What Alberta Law Actually Requires the Day You Find Out
The day you find out, you may have three sets of legal obligations, not one — different thresholds, different clocks, different lists of who must be told. PIPA, PIPEDA and HIA mapped against each other from primary sources: who levies the fine (it isn't the Commissioner), and the 24-month breach log most Alberta operators attribute to the wrong statute.
Aug 25, 2026 • 13 min
WHEN, NOT IF // 03
The cheapest developer you ever hired
Canadian courts have been kind to breached companies — but Setoguchi describes a fact pattern, not a shield, and Owsianik left negligence, contract and statutory claims alive. Negligence asks whether your conduct was reasonable, and that is the one question cheap code makes unanswerable: no logs, no record of what you hold, no way to run the test PIPA s. 34.1 requires.
Aug 18, 2026 • 4 min
WHEN, NOT IF // 02
Your backup is a belief, not a system
90% of security leaders are confident they can recover from a cyber incident; among organizations actually hit by ransomware, 28% got all of their data back — the distance between those two numbers is the distance between a belief and a system, and the only thing that closes it is a restore you have actually performed, timed and written down.
Aug 11, 2026 • 2 min
WHEN, NOT IF // 01
Alberta doesn't require you to tell your customers
Under Alberta's PIPA, the legal duty after a breach is to notify the Commissioner (s. 34.1) — notifying customers is a decision the law leaves with the Commissioner (s. 37.1); yet 80%+ of organizations notify on their own first, because a prompt, documented response is the real defence and concealment is what turns an incident into a lawsuit.
Aug 4, 2026 • 3 min
Personal notes from the research journey — first entries coming.
HawkClaw
Research in multi-agent orchestration for security validation. Rules of engagement and concept of operations as first-class artifacts, deconfliction between agents, MITRE ATT&CK used as a classification framework, not a target list. Scored 7/8 on the XBOW benchmark (Hard L3) — the benchmark validates the orchestration layer, not an invitation to an offensive service.
Built on the open-source Decepticon agent framework (Apache 2.0, PurpleAILAB); the operator dashboard and the cryptographic governance layer are the lab’s own work.
Hydra & Tesseract
Hydra is our name for composing specialized models into one system — a high-capacity core coordinating models dedicated to specific roles (orchestration, audio, vision). Tesseract is the internal nervous system that connects them: a message bus as synapses, a vector store as long-term memory.
Both are described here as designed and, where measured, as measured — the lab note says which is which.
→ Lab note: Hydra, composing specialized models