The notebook index

Field notes

Living questions, continuously revised. Each note is an entry point into the same underlying problem: keeping information, control, and evidence intact as they cross layers.

New writing · August 2026

  1. 01Do not collapse the spirals.The same word now names a Haitian novel, a chatbot religion, a named mood in the weights, and a protocol that claims not to break you. Keep the rows apart or the landscape turns into mysticism.
  2. 02One untrusted agent is enough.In 2024 this was a figure: one bad agent, then all of them. In 2026 the replica learned to want the hop, and the trusted agents talk to each other for a living.
  1. AA zero needs adjudication.A failed task can mean incapability, a broken task, an opaque harness, or a gameable verifier. Provenance determines which claim the number can support.
  2. BWhere does the model end?Weights, tokenizer, system prompt, tools, memory, classifiers, and product harness jointly shape behavior. Open and closed models are different kinds of objects.
  3. CThe frontier is funded.Geopolitics directs capital; capital builds infrastructure; infrastructure selects research; research becomes products that redirect capital again.
  4. DSystems are strange loops.Builders shape tools that reshape builders. Models, markets, institutions, and mimetic desire continually produce one another across time.
  5. ERed teaming is accountability.Adversarial pressure is most useful when it reveals a system boundary and forces providers to make more truthful claims—not when it stops at spectacle.
  6. FThe loop is an evidence system.OODA is the ancestry. In practice I add an explicit plan contract, bounded execution, a verifier that can reject the work, and a trace that becomes the next observation.
  7. GWhen synthesis gets cheap, truth gets expensive.Compute can multiply hypotheses and implementations. Planning, verification, and contact with reality become the scarce layers that decide which outputs deserve belief.
  8. HControl has grammar.System prompts, specifications, judge rubrics, and agent policies all use modal force to turn prose into an operational hierarchy.
  9. IEducation is an evaluation problem.Institutions graded artifacts as proof of thinking. AI made artifacts free, and the proxy collapsed. The same verifier failure I study in agents runs through the systems that credential people.
  10. JThe tiers are not one model.Three Claude tiers drift 100% on identity, refusal, and safety probes while capability stays nearly flat. Persona is a per-tier product layer, and it is observable without weights.
  11. KVeracity is infrastructure.Google's Knowledge Graph, EEAT, and AI Overviews are the world's largest deployed answer to "which claims count as facts." Trust there is a gradient, engineered — and agents that treat belief as a boolean repeat the mistake search already solved.
  12. LThe shape of what's absent.Deployed algorithmic systems encode their operators' priorities in measurable behavior. Refusals, framings, and shifts are signals — and behavioral probes are to public-facing AI what FOI requests are to public institutions.
  13. MHandoffs are memory.A modern research team includes members with no memory at all: the agents. The only durable team state is what gets written where the next reader — human or machine — will actually start.
  14. NEpistemic grounding is the missing layer.Safety research, search engineering, and production systems have independently converged on the same absence: agents have no substrate that tracks where beliefs come from, how confident they are, and when to defer. This note is the dated claim that these are one gap.
  15. OThe ontology is the interface.Between raw data and consequential action sits a plane of named objects with properties, relations, and permitted actions. Whoever defines that plane defines what the organization can mean — and what its agents can do.
  16. POrientation precedes grounding.Before "how true is this claim?" comes "from what vantage am I evaluating it?" Agents fail characteristically when their posture — lens, mode, altitude, reach — is left implicit.

Published elsewhere

  1. EA ForumThe Alignment Problem Is Upstream of the ModelAlignment is decided by layers around and upstream of the weights. The published essay that seeded the notebook’s harness-and-layers framing.

The research digests carry the measured claims; these notes carry the questions that produced them.Back to the papers →