Article 7: Test everything
Article seven of my personal coding constitution: tests are a property of the work, not a phase of it. The minimum viable test is 'does the documented behavior actually happen?' Anything less is hope.
Defense in depth: how many isolation layers do you actually need?
Most multi-tenant platforms over-build defense in depth. Network, namespace, RLS, tenant-id, per-tenant keys, the catalog feels safer than it is. The disciplined answer is fewer layers than instinct says, sized to the threat model, with the seams labeled.
DNA, lineage, and provenance: the genetic metaphor for AI artifacts
Manifests are snapshots; lineage is a graph. The genetic metaphor, descent, inheritance, mutation, gives an AI pipeline the vocabulary to track what every artifact inherited from where, and the design discipline to treat provenance as a first-class artifact instead of metadata.
Traceability as a debugging tool, not a compliance one
If you design traceability for compliance, you almost never get the debugging case. If you design for debugging, the 3am 'where did this payment go' case, the audit trail falls out for free. The reframe, what each shape of trace looks like, and why one is strictly more powerful than the other.
Pricing the service: subscription, per-resolution, outcome-based
Subscription, per-resolution, outcome-based. The pricing decision tree, the cost-of-goods math, and the free-tier question, closer for the operate series.
The customer view vs the consultant view: two surfaces, one product
One backend, two surfaces. Customer asks and gets an answer. Consultant supervises the queue, approves, denies, and mines for patterns.
Onboarding new tenants: the five-minute path from signup to working AI
When a consultant signs up, how do they get from 'I have secret sauce' to a live AI surface in five minutes? Onboarding as a first-class feature.
Bedrock model selection: pick on evidence, not vibes
Sonnet, Haiku, Opus, Llama. Picking the right Bedrock model per use case using evals, not gut feel, and knowing when to switch.
What I'd cut, what I'd keep: the actual MVP cutline
Closing the AI MVP series. What you can safely skip on day one, what you absolutely can't, and what the first 30 days of customers will teach you that nothing else can.
The downgrade pattern for cross-boundary data transfer
When data leaves a regulated universe for an analytics one, what crosses isn't the data, it's a downgraded version of it. Plain downgrade rules, enforcement, audit trail, and a human approval step for first-of-pattern transfers.
The hybrid sync pattern: how cloud and local actually talk
The wiring between a cloud AI product and a local Mac Studio. SQS for events, S3 for artifacts, EventBridge for schedules, signed manifests for the round-trip.
AI in the news — week of May 24, 2026
Google I/O ships Gemini 3.5, Antigravity 2.0 with a CLI that built a working OS in 12 hours, AI Mode at 1B MAU. Meta cuts 8,000 with a $21B CoreWeave bet behind it. BCG: AI data centres hit two-thirds of US home electricity by 2030. The Stratos campus moves; Florida's pipeline gets denser.