← Work
AI governance product experiment

FieldRules

Independent product experiment2026

FieldRules is an independent AI governance product experiment about the expertise layer above docs, data, and models: how domain expert reasoning at the time of judgment gets captured, reviewed, versioned, and reused before an AI system acts on it. The work connects a rule library, reasoning fields, review surfaces, evals, source-backed schemas, and an operating archive so a reviewer can inspect how trust is earned inside the product. The active build paused in May 2026; the thesis, schema, and evaluation work remain a complete record of the product reasoning.

Role
Self-directed product designer and builder across product thesis, schema, evals, artifact system, and marketing surface.
Status
Active build paused in May 2026; marketing site and product record remain live as evidence of the product thesis.
Eval signal
36 scenarios across 10 healthtech domains tested whether expert reasoning changed AI behavior.
Reasoning layer
Domain expert reasoning is treated as first-class semantic material, captured at the time of judgment rather than reconstructed later.
Why it belongs
Shows how I turn a fuzzy AI trust problem into rules, review loops, eval evidence, and a product surface an agent can safely read from.
Visit live FieldRules site →View FieldRules portfolio site →
Artifact system
AI governance product and artifact archive
Artifact systemAI governance product and artifact archive
Case study path

What this case study covers

  1. 01

    Why practitioner reasoning became the product material FieldRules was designed to protect.

  2. 02

    How expert judgment becomes a versioned dependency instead of scattered institutional memory.

  3. 03

    How rules stay owned, versioned, reviewable, and safe to reuse.

  4. 04

    How the eval work measured behavior and reasoning quality, not just polish.

  5. 05

    Where the prototype is complete, where it paused, and which caveats remain visible.

Live site
Product narrative and positioning
Live siteProduct narrative and positioning
Core insight

The system evaluates the teacher and the model.

Most AI evaluation infrastructure measures what the model produces. FieldRules starts earlier: what quality of human reasoning is the model being conditioned by? Every rule has an IF/THEN trigger and a reasoning field that holds the practitioner's judgment in her own words at the moment the rule is made. That domain expert reasoning is treated as first-class semantic material: scored, elicited, and governed differently from the structural clauses, even where it sits next to them in the schema.

Artifacts
Readable product proof across the archive
ArtifactsReadable product proof across the archive
Rule library

Rules became versioned expert judgment.

Writing the rule was only one part of the design problem. The product also had to keep that rule owned, versioned, reviewable, and safe to reuse. FieldRules treats rules as a dependency other systems can cite: not just policy text, but expert judgment with authorship, source context, version history, review states, and the ability to inspect how the reasoning changed over time.

Eval + archive
Connected evidence, evals, and rule reasoning
Eval + archiveConnected evidence, evals, and rule reasoning
Eval discipline

I tested whether the reasoning actually changed model behavior.

It would have been easy to assume that expert reasoning helps. I ran a head-to-head instead: 36 scenarios across ten healthtech domains, four context conditions each — rule with captured expert reasoning, rule only, playbook SOP, and raw ticket transcript. The rule with captured reasoning outperformed the other conditions on a composite of correctness, reasoning depth, specificity, and calibration. I would still caveat that honestly: the scenarios are healthtech-only, the scoring is LLM-rated rather than blind human, and the next test should compare practitioner-written reasoning against model-generated reasoning.

Agent archive
Notion repo for MCP-queryable FieldRules artifacts
Agent archiveNotion repo for MCP-queryable FieldRules artifacts
Operating archive

The strongest proof is the system around the product.

The archive shows strategy, schema, Notion and Obsidian knowledge work, pilot scripts, quality judges, Jira snapshots, design-system updates, component test scripting, seed-library generation, and recurring coherence checks. It was not meant to be a polished human help center. It was built as an agent-traversible Notion repo: structured so teammates could query the workspace through Notion MCP and get back the right rule, artifact, caveat, or source path without asking me to be the lookup layer.

Live site
Product narrative and positioning
Live siteProduct narrative and positioning
Why it matters

The agent needs a designed product surface too.

FieldRules is one of the clearest examples of how I think about AI product design. A person writes or reviews the rule, and an AI system may later consume it. When that second user gets an ambiguous surface — no provenance, no versioning, no reasoning, no review trail — the human gets brittle guidance back. The design work is making the reasoning usable by both. The active build paused in May 2026, but the schema, evals, and operating archive remain the record of the product argument.

Visit live FieldRules site →View FieldRules portfolio site →