Most advice about using LLMs in knowledge work focuses on better prompts. These essays focus on what better prompts cannot fix.
When an LLM is doing iterative intellectual labor on a long-lived project — drafting, revising, auditing, proposing alternatives — a specific failure regime emerges that is invisible from inside a single conversation. Documents get cleaner. Scope gets tighter. Everything reads as progress. But the same substantive defects persist in relabeled form, the project quietly becomes a different project, and there is no structural mechanism to detect any of it.
This is a governance problem, not a prompting problem. The three main pieces in this series cover the framework, the identity-preserving document pattern, and the workflow architecture developed to address it. A companion architecture note maps the same system as components, authority relationships, and current build status.
The pieces — in suggested reading order
A project’s biggest risk is not that it fails — it is that it quietly becomes a different project. This piece introduces the Frozen Anchor: a four-section document design pattern that records what a project agreed to be before revision pressure can distort it.
Read essay →The systems piece. It defines the governance problem, explains why ordinary revision loops fail, frames the process as a control problem, and derives the structural requirements for a system that can actually govern research definition under weak supervision.
Read essay →A workflow architecture that breaks closed-loop validation at the process level by separating research, implementation, ideation, and authorization into different contexts with different mandates and different cannot-lists.
Read essay →A supplemental concept note mapping the framework as a system of components: identity anchor, authority map, governed bundle, audit instrument, defect register, patch layer, re-audit layer, and convergence layer.
Read companion note →About this work
Concept essays, framework notes, and architecture writing developed through direct application on a live, solo LLM-assisted research program. The structures described here were built because they were needed — each rule documents a failure mode that existed before the rule did.
These are not validated research papers. The framework has been applied and refined on one program; it has not been benchmarked across multiple projects or independently evaluated. They are presented as serious concept work and qualified practitioner reflection.
The essays and companion note are derived from working artifacts: a live research protocol, a frozen identity anchor maintained across audit cycles, an implementation log, and draft governance architecture materials that predate their publication as essays.
The adjacent literatures are LLM auditing, weak supervision, research proposal validation, and convergence theory. This work sits between them rather than fully inside any one of them, with a distinctive emphasis on project identity, authority control, bounded revision, and cross-cycle defect movement.