Applying the framework
Case studies
End-to-end mappings across all five layers. Each one opens as a Studio preset.
- 1A: Mathematical reasoningInteractiveAnswer checkers, proof assistants, PRMs and mathematicians: one construct, four mechanisms, three regimes, and why a PRM is not "above" the checker.8 min read
- 2B: Coding agentInteractiveTests, static analysis, sandbox invariants, trajectory checks and a calibrated readability judge: an agent eval as a portfolio, not a score.8 min read
- 3C: Enterprise SQL/data agentInteractiveFormalize access policy, execute semantic consequences, judge only the residual. The SQL agent design end to end.8 min read
- 4D: Autonomous research agentInteractive"Research quality" is not one unverifiable blob. Route provenance, reproducibility, entailment, novelty and direction-setting to different channels.9 min read
- 5E: Reasoning-model trainingInteractiveDense RLVR and process feedback for development, an untouched certification channel, and judge error re-estimated at every checkpoint.8 min read
- 6F: Production safety monitorInteractivePermissions first, trajectory checks second, a learned monitor for the residue, validated at production base rates, until training starts to see it.8 min read
- 7G: Documentation agentInteractiveA judgment-heavy RAG agent: grounding and procedure criterion cards, SME-calibrated judges, and stale docs as an environment problem rather than a judge problem.9 min read