Framework Library
Judgment frameworks and mental models built during teaching sessions. Each is a standalone reference document — useful without the conversation context. Obsidian-compatible with backlinks and tags.
Frameworks
| Framework | Session | Core Question It Answers | Tags |
|---|---|---|---|
| research-transfer-judgment | 15 (2026-04-14) | "Does this research finding apply to my specific use case, or only to the benchmark context?" | #research #evaluation #agents #judgment |
| research-to-infrastructure-threat | 15 (2026-04-14) | "Is this research breakthrough a real infrastructure disruption or just an interesting paper?" | #research #infrastructure #investment #hardware |
| ai-company-moat-evaluation | 12 (2026-04-09) | "Does this AI company have a durable moat, or will it get commoditized?" | #investment #moats #saas #strategy #evaluation |
| ai-self-improvement-hierarchy | 9 (2026-04-04) | "Where does the new information come from, and what's the ceiling on self-improvement?" | #self-improvement #training #evaluation #reasoning |
| who-controls-the-dial | 8 (2026-04-03) | "Is this safety capability also a misalignment tool? Who controls it, and is there an audit trail?" | #safety #governance #evaluation #alignment |
| compound-reliability-agent-workflows | 7 (2026-04-02) | "Is this agent tool actually reliable in production multi-step workflows?" | #agents #reliability #evaluation #pricing |
| reward-signal-evaluation | 6 (2026-04-01) | "What behavior does this training approach actually incentivize?" | #training #RLHF #safety #evaluation |
| revenue-per-compute | 4 (2026-03-25) | "Why did this company kill/launch this product?" | #strategy #business #compute-economics |
How to Use
- When evaluating an agent tool or startup: start with compound-reliability-agent-workflows
- When evaluating a safety capability or interpretability breakthrough: start with who-controls-the-dial
- When evaluating a training announcement or safety claim: start with reward-signal-evaluation
- When evaluating a company strategic move: start with revenue-per-compute
- When evaluating an AI/SaaS company for investment: start with ai-company-moat-evaluation
- When evaluating whether a research finding applies to your deployment: start with research-transfer-judgment
- When evaluating whether a research paper threatens existing infrastructure: start with research-to-infrastructure-threat
- When evaluating a chip or hardware announcement: see ai-compute-evaluation-framework (in parent directory)
- When evaluating an agent architecture: see agent-architecture-skeleton (in parent directory)
Inclusion Criteria
A framework qualifies when it:
- Answers a recurring judgment question — something you'll ask again when new signals arrive
- Was validated in conversation — not a theoretical construct, but something that produced a useful evaluation when applied
- Is transferable — applies to a category of signals, not just the specific one that triggered it
Source: frameworks/README.md