Delivery approach
Match the governance to the risk.
Every proposal and programme needs a different amount of process: too much slows down low-risk work, too little lets real risk through. I made that call for years in pre-sales and delivery, most visibly on the Chrome Autofill retrieval flow, governed by a formal Definition of Done, staged production rollout, and Google's own certification testing. I bring the same judgment to the two AI-directed systems I run solo today, just scaled to a different kind of risk.
OpenCAM and my job-search tooling sit at opposite ends of that same call. OpenCAM is built to sit upstream of real credit decisions — even at this early, one-analyst-validated stage, it runs against a real analyst's real deal data, so it gets the same rigor I've applied to regulated delivery work for years: independent verification, a paper trail, nothing shipped without review. My own job-search tooling only has to answer to me, so it runs lighter: direct commits, no backlog, my own daily use as the test.
Proof in practice
The model changed the result
Controlled · OpenCAM
Reopening a governance gap before it mattered
I'd marked Issue #31 resolved once Maker and Checker could run on different models in code. A later review — the kind strict governance forces — caught that the setting was never actually switched on in production, meaning the audit hadn't been independent at all. I reopened it and only re-closed it once verified end to end.
Read the OpenCAM case studyLightweight · AI Job Search
A same-day fix, no PR queue required
A role I'd already been rejected from got scraped three times under three different job IDs, and the surviving copy nearly went out again in a batch of applications. The fix shipped the same day I caught it — direct to master, because on personal tooling the fastest path to "this can't happen again" was also the right one.
Read the job-search case studyTwo operating modes
Same discipline. Different weight.
Mode 01 · Controlled Delivery
Guardrails before velocity
For regulated, shared, or irreversible work. The process creates evidence, makes accountability explicit, and fails closed.
- Commit Policy
- 100% PR-based, zero direct-to-main
- Issue Tracking
- Formal GitHub Issues (19-issue gap audit)
- QA Model
- 439+ automated tests, regression-gated
- Target Audience
- Other institutions (forkable, not yet forked)
- Risk Profile
- Production-adjacent, third-party dependent
Mode 02 · Lightweight Delivery
Learning before ceremony
For bounded, reversible, personal-scale work. The process protects the essentials, then removes friction so learning stays fast.
- Commit Policy
- Direct commits to master
- Issue Tracking
- None — solo, no backlog overhead
- QA Model
- Manual verification against live runs
- Target Audience
- Personal use only (N=1)
- Risk Profile
- Low-risk personal tooling, fast iteration
How I decide
Three questions before process
01
What happens if this is wrong?
I map the blast radius first. OpenCAM is built to sit upstream of real credit decisions — a wrong answer there has financial and reputational consequence. My job-search tooling's worst case is a missed application. Consequence sets the minimum control level.
02
Who needs to trust the result?
A bank adopting OpenCAM needs to verify what was checked and by whom, without me there to explain it — that needs traceable, independent evidence. My own job-search tooling only has to earn my own trust, in real time.
03
What's the cheapest honest proof?
For OpenCAM, that meant 439+ regression-gated tests and an independent Maker-Checker loop before anything reached a real analyst. For personal tooling, the test was simpler: does it work on my own real job search, today? When a scraping bug let a rejected role resurface under a different ID, catching it in daily use and shipping the fix that same day was proof enough.
This is the same judgment I'd bring to any team's delivery process — proportional control on regulated or shared work, and the discipline to strip ceremony away from everything else, rather than defaulting to one process for every kind of risk.