Case files // 03

Open the
evidence.

Two flagship investigations and the supporting work around them. Everything below is public on GitHub — specs, safety rules, tests, scorecards, and reports included.

Case file AF-001 // shipped · MIT license

AgentForge

One spec. Many agents. Agentic coding assistants each want their own config dialect — CLAUDE.md and hooks here, AGENTS.md there, .cursor/rules somewhere else. AgentForge separates the portable agent posture from each tool's file formats: author identity, routing, memory, automation, and telemetry once in canonical YAML, then compile platform-native files for every runtime you use.

What ships

  • Six adapter targets: Claude Code, Codex, Gemini CLI, Cursor, Aider, and a portable generic AGENTS.md.
  • A real CLI: agentforge init <adapter> and agentforge doctor, published to npm as @kmitops/agentforge@0.3.1 — runnable via npx with no clone.
  • Proof, not promises: adapter round-trip tests, package install smoke tests, a readiness runbook, and a visual demo.
  • Safe to re-run: idempotent installs; every install creates a git-tracked checkpoint so rollback is one command.
  • Benchmarks built in: the AgentForge suite includes reproducible agent evaluations — ATT&CKLens (below) is its first defensive-security benchmark.
Stack
Node.js ≥ 18 · npm · bash · YAML
Registry
npm — @kmitops/agentforge@0.3.1
Verification
Round-trip tests · install smoke tests · doctor CLI · readiness runbook
License
MIT

Case file AL-002 // published · MIT license

ATT&CKLens Benchmark

The question: can AI coding agents build secure, uncertainty-aware, MITRE ATT&CK-aligned defensive security tooling? Seven agents got identical specifications, standardized acceptance tests, and prompt-injection challenges. Their output was scored against a published rubric — mapping discipline, evidence citation, confidence handling ("No Clear Mapping" when warranted), detection ideas, remediation guidance, and the follow-up questions a real analyst would ask.

Final scores — 7 agent artifacts, rubric of 100

Full methodology, scorecards, raw outputs, and logs live in the repo. The authoritative report is evaluations/final/final-report.md.

Why it's credible

  • Reproducible by design: one branch per evaluated agent plus a main harness branch — clone, fetch all, run benchmark-all.ps1, and the harness re-scores every submission.
  • Defensive framing throughout: agents are graded on detection and mitigation quality and on resisting prompt injection — not on building offensive capability.
  • Honesty is scored: the rubric rewards "No Clear Mapping" over confident guessing, plus evidence citation and analyst follow-up questions.
  • Everything preserved: scripts, scorecards, raw outputs, logs, branch inventory, acceptance tests, and a final report.

MITRE ATT&CK® is a registered trademark of The MITRE Corporation. ATT&CK content is used under MITRE's Terms of Use. The benchmark is independent work, not affiliated with or endorsed by MITRE.

Reproduce it.
Then call me.

The strongest interview question you can ask me is "walk me through your benchmark harness." I'd love that.