THE PROJECT / SOURCES & APPROACH
A field guide.
An open argument.
Built by Astra for QZ, to turn a collection of papers into something you can learn by exploring.
Where this began.
QZ brought a structured corpus of sixteen research papers connected to the YC Paper Club talk Why The Harness Matters More Than The Model ↗. The corpus includes paper summaries, a graph of mechanisms and relationships, and an editorial classification of what agents can change.
Astra designed and built this atlas from that material: eight original lessons, three deterministic experiments, an interactive harness cutaway, and six ways to explore the research. It is an independent learning project. Y Combinator and the researchers did not commission or endorse it.
Three kinds of statement.
A finding attributed to a paper in the corpus. Source links and selected short excerpts let you inspect it. This is a report, not our replication.
An interpretation made while reading the papers: mechanism classifications, relationships, and cross-paper synthesis. Reasonable readers can disagree.
A small scripted world built to teach a distinction. Outcomes follow transparent rules. No real model is running, and no displayed outcome is a research measurement.
What the atlas does not flatten.
The write-access ladder is a teaching device, not a ranking of intelligence. Its rungs are not cumulative and do not arrive in chronological order. Prompt editing, memory management, and parameter updates are different interventions.
The matrix’s four categories are coarse readings of each paper. “Central” describes a contribution; “self-modifying” describes who changes an element. Adding them into a single quality score would hide that difference.
The permission classifications describe what these papers discuss. They do not measure the total safety of a system. The newer papers do not establish that containment improves over time, and this selected bibliography cannot establish gaps in the wider field.
Prime Agent’s inherited assignment to the top rung is explicitly qualified. The corpus’s classification is not proof that the paper demonstrates online model-weight updates.
Sources stay close.
The public research snapshot was prepared on 8 September 2026 from QZ’s corpus. All sixteen papers have direct arXiv links. The original corpus records method-reading checks for all sixteen profiles; those checks are annotations, not a guarantee that every judgement is correct.
Only structured research extracts and public research relationships are included. The original PDFs, full-text extractions, personal notes, and the surrounding knowledge vault are not distributed here.
Paper versions can change. For any claim you plan to rely on, follow the link to the current paper and check the precise method, comparison, and limitations.
Learning without an account.
The course is open. Chapter completion is stored only in your browser and is optional. The laboratory uses no model API. The 3D view has labelled controls and a fallback diagram; all the lessons and research notes remain available without WebGL.
Prefer less motion? Use your device’s reduced-motion setting or pause the opening model. Every important state in the experiments is also described in text.
THE QUESTION TO TAKE AWAY