JD DavenportFounder & Managing Director, BYU AI Foundry

A calibrated multi-persona LLM-as-judge that any MCP client can call as a tool to score a document against a rubric.

Why it mattersIt fixes the four ways LLM graders fail (mid-band compression, severity bias, score and verdict drifting apart, run-to-run noise) and ships the eval harness that proves it.

Five working controls for putting Claude agents inside an enterprise: a policy chokepoint, a destructive-command deny-hook, an eval gate, a redacting audit log and pluggable secrets.

Why it mattersThe questions every enterprise deployment asks, answered in code instead of a slide.

The production orchestration core of the system above: bounded fan-out, a durable message bus, circuit breakers, best-of-N with an adversarial judge, and permission tiers, with the eval harness that tests them. clawd-agent-os is an earlier, smaller extraction of the same core.

Why it mattersParallel agents that cannot silently drop a message, and a judge that has to be convinced before work counts as done.

PreToolUse and Stop hooks for autonomous Claude sessions: irreversible commands wait for tests, credential stores stay off-limits, outbound GitHub actions get judged.

Why it mattersIt enforces at the tool boundary what a system prompt can only ask for.

A governor for autonomous fixes. An action runs unattended only if it can be undone, and it undoes itself if the health signal does not go green.

Why it mattersAutonomy admitted by reversibility, not confidence. Every test runs offline.

Target-behavior specs for the Claude Code CLI, written from production incidents. Each pairs observed and expected behavior with a reproduction you can run and a proposed eval.

Why it mattersWhat breaks when the CLI runs unattended, at concurrency, for months, written so a product team can act on it.

Point Claude Code at a repo and a bug report. It triages, fixes in an isolated worktree, proves the bug goes red to green, and runs the fix through a layered gate.

Why it mattersSafe by default: it proposes and gates, and never merges until you turn that on.

A research loop that searches 9 sources in parallel until a knowledge-gap check says it is done, argues against its own draft, then fact-checks every claim with a 3-vote refute-or-survive pass.

Why it mattersIt treats a draft as a hypothesis to attack, not an answer to trust.

Treats a job search as a verification problem. It merges your resumes into one experience bank, tailors a resume per posting, then reads the finished PDF back and runs three independent judges before it will say yes.

Why it mattersIt measures its own output instead of trusting it, and it deliberately has no submit path.

Four Markdown templates and five skills that give Claude Code a durable model of you: who you are, how you decide, how you write, and your hard rules.

Why it mattersStop re-explaining yourself every session. One-command install.