AI governance here is not a policy document. It is a set of named rules enforced in the rendering layer, so a model that ignores them produces nothing at all:

• ground in evidence; flag gaps rather than fabricate; no inference fills
• redact before the model — not after
• an identity band is required, and profile metadata never reaches the output
• an unresolved token is a hard failure, not a blank

Verbatim-field guarantees protect the fields where paraphrase would be a defect — legal name, dates, service package, staff contact details, document hashes. Every credentialing playbook runs recommendation-only: the model recommends, a credentialed human decides.

And the piece that makes it hold: audience ceilings live on the data source itself. A mis-scoped prompt cannot leak, because the source is not resolvable for that audience at all.

The problem. Texas mandates staffing ratios that do not reconcile — nine of them, written at different times for different purposes, all binding at once. No scheduler can satisfy a rule set that contradicts itself.

The move. A 120-point caseload weighting index that expresses all nine as weights against a single budget, with quarterly recalibration built in.

What that buys is not tidiness. It is that a supervisor can now answer "can this person take one more case" with a number instead of an argument.

The print firm's operations tooling could have been built for the partner who fields the customers. That was rejected: what the business needed was leverage behind him, not administrative work in front of him.

Four rules followed — friction goes to zero for the person being helped; he experiences a person rather than a system; the triage queue stays out of his sight; and there are no privacy walls between partners and no surveillance feel either.

Design is deciding who the system is for, and it is the same judgment whether the artifact is a platform, a report, or a sign on a truck.

The situation. One vendor estimate, one owner who must approve it, and no technical vocabulary between them.

The move. The single estimate was bracketed with a conservative alternative and an aggressive one, and wrapped in a plain-language decision framework — budget ceiling, timeline criticality, risk tolerance, feature sufficiency, coordination capacity — with explicit conditions under which each option is the right one.

Scope was held at the application layer. Infrastructure was specified as a complete resource bill of materials and sent to the vendor for price only, and when their quote came back above market, the variance was published in the analysis rather than buried. Migration was sequenced by blast radius — candidates, then clients, then administrators — with staged cutover and rollback retained. Score-correctness testing was routed to psychometricians rather than developers, so the people who know whether a score is right maintain the tests that say so.

The point was not to recommend. It was to leave a non-technical owner able to exercise real judgment on her own terms.

The situation. After the agencies' T3C submission, the regulator issued a 143-item enhancement request citing a newer edition of the Blueprint. The obvious reading was 143 new requirements and a re-write.

The move. Every item was classified against the review form's own history columns and against the documents already in the regulator's possession — rather than against the panic.

The finding. None of the 143 constituted a new requirement. Approval followed. In the next wave the instinct became a governed intake step, and the count of net-new work was zero.

What two licensed child-placing agencies actually run on is seven layers, each with its own authoritative record:

The through-line is the same at every level: the record is the system, and it is versioned, reviewable and diffable rather than remembered.

Texas rewrote its foster-care service-level and reimbursement system for the first time since 1988. Leading both agencies' readiness meant turning a 496-page Blueprint into things people could actually do:

• service-package specifications with page-level citation traceability, so every requirement points back to the sentence that created it
• a program-evaluation framework — logic models, PDSA cycles, a KPI set across five domains
• workforce design: role descriptions with inline compliance tagging, and a six-month onboarding curriculum
• an IT security and data-governance policy suite, including the full joiner-mover-leaver identity lifecycle

State credentialing reviewers examined the documentation and approved it. The agencies hold interim-active credentials across the requested service packages, with further packages in process.

Outcome statistics for the underlying programs are still under development and are not claimed here.

One family of validated selection-assessment instruments, carried across four platform generations and four ownership changes. Hands-on for the first two; specification and architecture for the last two. Every successive owner needed the knowledge, which is why the thread never broke.

The capitalization differs across all four. That is not an inconsistency to tidy — they are four products' own names from four eras.

Chief API architect for client integrations — Workday, iCIMS, other applicant- and human-resource systems, and the general API clients consume.

One client's system offered nothing to integrate against. Rather than declining, a bridge integration was designed to give it one. The interesting engineering is rarely in the systems that publish good APIs.

An artifact-definition catalog that classifies every document type in the estate — 1,610 definitions, a full taxonomy, uniform fields, binder branches, entity base roles.

The point is not the count. It is that "what kind of document is this" became a question with an authoritative answer instead of a judgment call made freshly by whoever was filing it that day.

The rule. A clinical-deterioration check, specified and ready to hand to an implementing agent.

The test. Replayed against a real case before handoff — which is the step that is easy to skip, because the spec was already written and already good.

The finding. The rule was blind to a case where five new clinical problems emerged while the actionable-item count stayed flat. It would have passed a child who was getting worse.

The decision. Handoff blocked, spec fixed first. The cheapest defect to fix is the one that has not been built yet, and the only way to find it is to run the specification against reality before anyone writes code.

A business that runs largely over text needed its messages triaged. The obvious fix was to point a cloud model at the message database. That was refused — cloud models read every message they touch, including personal threads.

Instead: an on-device model, raw content never leaving the machine unless already judged confidently business, and classification running before sender identification — so the privacy gate is the model's verdict rather than the contact list, and a message that does not qualify never has its sender resolved at all.

Real business messages will sometimes be missed. That price was written down in advance and chosen.

The complete candidate-experience design package for a platform redevelopment: a designer's primer on selection regulation and applicant-reactions theory, a sourced research brief, a surface-by-surface design direction, and nine working HTML mockups — not pictures of screens. Roughly 12,000 words, sole-authored, to WCAG 2.2 AA.

Status: designed and delivered. The redevelopment has not shipped.

Sole responsibility for the delivery function since 2022: scope, specification and design, then handoff with expectations set at handoff. Priority and sequence for the team's work, the project board with them, and every ticket written personally. Follow-up through implementation, pre-release QA, and release sign-off.

Vendor staff rotate off as they move to other roles. The documentation and the ticket discipline are not administrative habits — they are the mechanism that carries continuity across every handover, because the people change and the record does not.

hapbi — billing generation and export, pricing with change history, client and activity reporting, and a candidate right-to-be-forgotten module built in visible phases with dry-run preview, audit trail and response letter.

Node and Express on Azure, Entra ID and Key Vault, SQL Server and MongoDB Atlas, CI/CD with staging-slot swaps. Concept to implementation by one person, human-directed and AI-assisted. One author, 195 commits.

Development paused since May 2026 pending a decision on the platform's historical data. Paused is paused.

Built from scratch in late 2025: a git-versioned policy base that serves as human reference, API payload, and AI-agent context at once, with an instruction file at its root telling every future agent session how to operate inside it. Continuous integration generates the published PDFs and the goal library from source.

Sole author, sole committer — every pull request in it. A hundred megabytes of policy that a person can read, a service can serve, and an agent can be trusted inside.

A Model Context Protocol server exposes the compliance platform to AI agents as first-class tools, and it exposes them transactionally: begin, preview, commit, roll back.

An agent seeding data does not write and hope. It stages, shows its work, and waits for a commit — and the rollback is a real path, not a promise. Giving an agent write access without giving it a transaction is how you get a mess nobody can unwind.

Rates derive from stated rules. The boundary between standard and specialty work is settled by a stated test rather than by feel. The whole model is market-validated and reviewed on a schedule, and one model renders for three audiences — the crew, the client, and the books.

A loss on a specialty job became two standing rules and a signed-agreement requirement. The systemic fix was named as such rather than left to memory, which is the difference between learning something and merely regretting it.

The platform's scoring layer carries two paradigms, both designed and specified here with the company's psychometricians: ratings applied at the score level first, then the conversion of all scoring and calibration to percentile ratings.

The Overall and Potential algorithms are of this design, along with an original multi-pass method in which composite percentiles are re-normed against the benchmark. Rating adjustment is progressive and data-driven, keyed to percentile ranges with compounding parameters — configured, deterministic, auditable.

Designed and specified as a peer to the psychometric team. The instruments themselves stay with the psychologists.

Radius — a case-management and document platform for two licensed child-placing agencies — was architected and hand-built from about 2004, in Visual Basic and later VB.NET against SQL Server, and has run continuously since, touched by no other developer. A registry-driven form model resolves and instantiates forms at runtime; per-child shared mailboxes ride on Microsoft Graph; a React-in-WebView2 bridge retired InfoPath.

Its front end is now migrating to Pulse, on a path chosen deliberately so that other people can be brought in — the exit plan is part of the design.

Elements mix and match by trait type and rendering model, then stack: custom trait labels, custom items such as problem-solving scenarios, custom components, custom reports. Narrative comes from a prioritized-trait comment mechanism that measures priority traits against a candidate's scores and returns database-configured comments in priority order.

In production for more than ten years, which is the only review that counts for a reporting engine.

The audit framework was superseded twice — by its author. Two earlier findings turned out to fail the rule that had since been written, and both were retracted.

Not deleted. The retraction was written into the playbook's own text, and the run was left standing. A finding that vanishes teaches nobody anything; a finding that is visibly withdrawn, with the reason attached, is how a governance layer earns the right to be believed the next time.

An audit layer can surface everything at once and paralyze the people who have to act on it. The governing test here decides what surfaces now: would this finding trigger a follow-up, a correction, a clarification — or would it trigger a forensic investigation?

The second kind is not dismissed. It is sequenced, deliberately and in writing, so that deferring it stays a decision rather than becoming an omission.

Visual Basic was picked up in 1998 on the founding assessment project — the OCR bubble-form system for BH2000. Within a few years it became the foundation of Radius, a mission-critical platform in an entirely unrelated industry, and it carried forward into VB.NET as the tooling changed underneath it.

A tool learned on an assessment-industry project in 1998 has now been in continuous operation, in child welfare, for over two decades — and is being migrated forward rather than retired. Each chamber is built on the chamber before it.

Nothing AI-generated becomes operative without a named human approving it — and that rule lives in the rendering layer, not in a policy document.

Internal-audit and consulting engagements across industries; a security review and analysis tool built for Oracle Financials; a development team managed for an audit-management tool; analytics and business-process documentation.

The second stint followed a decade of independent consulting — which is a different thing from never having left.

Physical: business cards, brochures, periodicals, banners, backdrops, apparel, billboards, and building signage — flat and lit. Digital: web design, mobile app design, digital magazines, email.

All current. Not an earlier, broader era that has since narrowed.

A playbook registry binds document types to executable governance rules, with real supersession history.

A tenant-specific fork was designed at the outset for clean removal at sunset: its schema droppable in one statement, and every file marked with a warning banner so an agent can grep for it later and take the whole fork out.

Deciding how something ends is part of designing it.

SonicAssess — specification and database architecture across relational and document stores, retrofitting for vendor integration and upgrade, and ETL delivering decrypted historical data to clients. Generation three of the lineage.

Curriculum, competency and delivery, designed as a system rather than a binder.