Edward Lue Chee LipEdward

Field notes / index

The notebook index

Field notes

I follow failures from memory and agents into the verifiers and institutions around them. These notes ask how the resulting feedback loops can stay open to correction. Start with Inside the loop →

The connecting argument · September 2026

00Inside the loop.The system is larger than the model. A working argument from memory and evaluation to independent verification, pluralism, and human-machine feedback.

Recent writing

  1. 03Optimization without an optimizerFrom a recurring dream to a question about evaluation: how do we keep systems correctable when they can influence the evidence used to judge them?
  2. 01Do not collapse the spirals.Reading accounts of chatbot personas without treating a metaphor, a reported experience, and an experimental mechanism as the same claim.
  3. 02One untrusted agent is enough.Three studies show how malicious instructions can propagate between agents. What do they imply for systems that pass messages and preserve memory?

Memory & agents

What does the system remember?

  1. 01Where 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.
  2. 02Systems are strange loops.Builders shape tools that reshape builders. Seven distinct layers trace how an output becomes an action, a judgment, and part of the next environment.
  3. 03The shape of what's absent.Behavioral probes can reveal changes in a deployed system. Explaining those changes requires a baseline, repeated measurements, and limits on causal interpretation.
  4. 04Handoffs are memory.A handoff document gives the next person or agent a starting point, with links back to the evidence and decisions behind it.

Evidence & evaluation

What would make me believe the result?

  1. 01A 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. 02When 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.
  3. 03Red 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.
  4. 04The tiers are not one model.An archived 44-probe comparison shows textual differences between three Claude tiers. Rechecking the public files also exposed limits in the original interpretation.
  5. 05Veracity is infrastructure.A note on carrying evidence and uncertainty with a claim, and deciding what an agent may do with it.

Tools & working methods

How do I turn a judgment into an action?

  1. 01The 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.
  2. 02Control has grammar.System prompts, specifications, judge rubrics, and agent policies all use modal force to turn prose into an operational hierarchy.
  3. 03Epistemic grounding is the missing layer.A design question connecting human agency, evidence provenance, and decisions about when an agent should defer.
  4. 04The 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.
  5. 05Orientation 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.

Institutions & incentives

Who sets the conditions?

  1. 01The frontier is funded.Geopolitics directs capital; capital builds infrastructure; infrastructure selects research; research becomes products that redirect capital again.
  2. 02Education is an evaluation problem.My experience with assessment led me to ask what a submitted artifact can establish about the person who produced it.

Try it yourself: change a file, then check the evidence → A synthetic walkthrough with commands and captured output.

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 →