Internal documents that happen to be published.
Essays and field notes on AI-native operations, constitutional agent design, and the systems behind the work. No cadence promised. Each one makes a claim and defends it.
Featured
SelectedBrowse by tag
- / 0114 October 2026
The Safe Way to Use AI in High-Stakes Work
A safe AI deployment begins by narrowing the job. That is the clearest lesson in the Superior Court of Québec's evaluation of its 98-day artificial intelligence pilot, published in April 2026. The pilot involved 22 judges and nine specialised AI agents. The court did not ask the systems to exercise judicial judgment. It restricted them to preparatory and support tasks, required judges to verify every output, and kept the deployment inside a secure institutional environment. The design also in
- / 0212 October 2026
The Safe Way to Use AI in High-Stakes Work
TL;DR: The Superior Court of Québec ran AI with 22 judges for 98 days and kept every decision with the judge. It worked because the court narrowed the job first, then verified every output, evaluated the use, and kept the deployment reversible. * Limit AI to preparatory and support tasks. Leave judgment with a human. * Require the reviewer to verify every output. * Evaluate the work around the model, not only the adoption figures. * Name the stop conditions before launch. A safe AI deploym
- / 037 October 2026
Your AI Outage Is a Business-Continuity Event. Design For It.
TL;DR: Two of the three AI outages on 3 September 2026 shared one hidden dependency, and nobody could see it while it was happening. Treat AI outages as business-continuity events. Give every AI-dependent process a named owner, a written degraded mode and a recovery time you have actually tested. What is the short answer? * A second AI vendor does not protect you if both vendors sit on the same infrastructure. * Ask four questions about each AI-dependent process: what stops, who owns the fa
- / 0430 September 2026
The ops debt machine: what Meta's AI agent collapse teaches about activity versus output
TL;DR: Meta's AI agents produced 220% more code. Users got 36% more features. Incidents rose 40%. Firefighting time rose 70%. Eight thousand people were let go, and the model still underdelivered. The failure was sequencing, not the technology. * Output and outcomes separated. More code did not mean more value reaching users. * Human capacity was removed before machine capacity existed to replace it. * The plan leaked before vice presidents were briefed. Morale dropped 19 points. * The corr
- / 0513 September 2026
Decision Record
TL;DR: Every AI rule written for financial services since 2024 converges on one demand. A record. Firms understand the obligation. They fail at the schema, the write path, and the retention decision. This piece is the build specification nobody publishes. * Show a decision your system made, prove what it was, why it happened, which control was in force, and who owned it. * APRA says it in principle-based language. The EU says it in Article 12. The SEC said it in 2022 without mentioning AI at
- / 0613 September 2026
What APRA's AI Letter Actually Means for the Executive Who Signs the Accountability Statement
TL;DR: APRA's 30 April 2026 letter moved AI governance inside the Financial Accountability Regime. The named executive on the accountability register now carries personal exposure, including disqualification and civil penalty, for AI systems they may not know are running. IOSCO, FCA, FSB, ASIC, and MAS reached the same position inside ten weeks. The informal phase is ending. * APRA's April letter places AI governance inside FAR, attaching personal liability to named executives, not to the comp
- / 079 September 2026
The four structural failures behind every AI governance program regulators are about to find
TL;DR: Regulators across four jurisdictions moved in one quarter. The institutions that filed policy documents are already behind. The ones that built runnable architecture are not. This article names the four structural failures separating the two groups, and what fixing them actually requires. * AI governance built around documents will not survive a regulatory exam that asks for raw evidence. * Assurance that runs quarterly on a system that changes daily is a dated snapshot, not a control.
- / 0826 August 2026
The call and the size
TL;DR: Elm Wealth ran eight language models through a real historical trading simulation with leverage. The models were credible at picking direction. They were catastrophic at sizing. The loss, in every case, came from the multiplier, not the view. This is a structural finding, not a statistical one. * Language models can form directional market views that beat human benchmarks. * They consistently over-leverage: average gross leverage near 30x against a defensible 7.5–15% volatility target.
- / 0922 August 2026
The Excel formula that belongs in your privacy policy
You have until 10 December to know where your software makes decisions about people TL;DR: APP 1.7 to 1.9 commence 10 December 2026. Every APP entity must disclose automated decision-making (ADM) in its privacy policy. The disclosure takes an afternoon to write. Finding what to disclose is a four-month programme. Almost no organisation has started. * "Computer program" covers Excel scorecards, rules engines, and legacy decision trees, not just AI models. * Human sign-off does not exempt you.
- / 1016 August 2026
Hidden reasoning is an information asset. You have not classified it.
Hidden model reasoning is an information asset you have not classified, cannot sanitise, and cannot verify. What the reasoning-trace extraction paper means for CPS 234, your retention obligations, and what counts as evidence in an AI system.
Get these as they’re published.
Long-form on AI governance in regulated firms — what the control gap actually is, what regulators are asking for, and what the evidence has to look like. Roughly monthly. No pitches.