Three historical sources converge in HAAK: a 2015 human role-playing exercise that is the proof-of-concept; a 17th-century intellectual network that names the structural problem; and an architecture of science where the figure is the atomic unit and the lab is a team of specialized agents. These are not mere backstory — they are the empirical observations that the formal theory (Library Theorem, institutional intelligence) later explained.
#The CNP Retreat (2015)
In April 2015, fifteen participants at the Champalimaud Neuroscience Programme retreat role-played the scientific ecosystem. Each adopted a character — ambitious junior faculty, Nature editor, first-year PhD student, philanthropist, struggling Portuguese researcher, FCT president — with different weights across six latent variables:
| Variable | What it measures |
|---|---|
| Personal well-being | Quality of life, work-life balance |
| Benefit to society | Social impact, public good |
| Knowledge | Understanding, discovery |
| Risk | Career risk, financial risk |
| Money | Funding, salary, resources |
| Productivity | Output, publications, citations |
Higher-level variables emerged from discussion: knowledge advancement, societal benefit, career advancement, exploration vs. exploitation, researcher quality of life, independence.
Eight question clusters structured the exercise: project vs. lab-based organization, meritocracy, financial resources, efficiency vs. mentorship, measuring scientist quality, open science vs. scooping, quality of life, personal values.
The exercise surfaced structural problems through embodied perspective-taking:
- PhD overproduction: Too many PhDs for too few permanent positions
- No staff scientist path: The system produces researchers but not research infrastructure
- Perverse metrics: What's measured (publications, citations, h-index) diverges from what matters (insight, rigor, mentorship)
- Degraded peer review: Reviewing is unrewarded labor with declining quality
- Restricted access: Knowledge locked behind paywalls and institutional memberships
Why this matters for HAAK: This exercise is exactly what HAAK now does computationally — assign AI agents to institutional roles with specified objectives and constraints, run the interaction, observe what emerges. The retreat was a single-shot, 15-person, afternoon-length experiment. HAAK can run thousands of variants with different parameters, different institutional structures, different objective weightings. The retreat proved the concept; HAAK scales it.
#The Invisible College (1640s)
The name "HAAK" comes from Theodore Haak (1605–1690), a German-born polymath who convened the meetings that became the Royal Society. The Invisible College — the 1640s network of people who wanted to do science before institutions existed — names three structural insights:
#Samuel Hartlib: the intelligencer
Hartlib ran the Correspondence Network, routing letters between natural philosophers across Europe. He was the optimizer — the person who made information flow efficiently in a world without journals, databases, or email. His function was indexing: knowing who was working on what, connecting people who should know about each other's work, maintaining a structured map of ongoing inquiry.
Hartlib was excluded from the Royal Society when it formalized. Institutions develop political logic that overrides founding principles. The people who make the pre-institutional network work are not the same people who thrive in the institution.
HAAK lesson: The system must be designed so that its organizational intelligence — its indexing, routing, and connection functions — is structural (built into the architecture) rather than personal (dependent on a key individual). Hartlib's function should be a method, not a person.
#Theodore Haak: the convener
Haak organized the meetings. Not the smartest, not the most productive, not the most political — the person who created the conditions for others to work. The convener's contribution is infrastructure: time, place, format, invitation. Without the convener, the network doesn't meet; without meetings, the Royal Society never forms.
HAAK lesson: Convening is a function the system performs. Creating a review panel, assembling a research team, organizing a project workspace — these are acts of convening. The system is named for the convener because convening is what the system does: it creates the conditions for structured inquiry.
#Robert Boyle: the scientist
Boyle did the science — experimental investigations of gases, the air pump, skeptical chemistry. His work was enabled by the network (Hartlib's connections, Haak's meetings) but his contribution was irreducible: the quality of thought, the rigor of experiment, the willingness to be wrong.
HAAK lesson: Human scientific judgment — the capacity to ask the right question, design the right experiment, recognize the surprising result — may be the irreducible contribution that HAAK's core question (where is human participation most valuable?) is designed to find.
#Science Architecture: The Figure and the Lab
The atomic unit of scientific output is the figure — a self-contained chunk with data, statistics, analyses, plots, methods, results, and local interpretation. A figure is nearly a mini-paper: everything needed to make one claim, evaluate its evidence, and understand its context.
This observation structures how HAAK models scientific work:
#Figures compose into manuscripts
A manuscript is a sequence of figures arranged to tell a story. The arrangement — which figures, in what order, with what narrative connecting them — is the compositional act. Individual figures can be produced by specialized agents; the composition requires a higher-level judgment about what story the data supports.
#Labs are teams of specialized agents
A lab is not one person doing everything. It is a team:
| Agent role | Function |
|---|---|
| Bayesian statistician | Prior specification, posterior analysis, model comparison |
| Frequentist statistician | Hypothesis testing, confidence intervals, power analysis |
| Figure-maker | Visualization, layout, clarity |
| Manuscript writer | Narrative composition, argument structure |
| Reviewer | Critical assessment, gap identification |
In HAAK, these are implemented as AI agents with distinct personas and competencies. The architecture mirrors the structure of actual scientific labs — division of labor, shared workspace, coordination through explicit communication.
#Levels of scientific organization
| Level | Unit | Function |
|---|---|---|
| Figure | Self-contained analysis | The atomic publishable unit |
| Manuscript | Composition of figures | A story built from units |
| Grant | Anticipation of manuscripts | A projection of work before knowing results |
| Selection | Choice among grants | Deciding priorities |
Each level composes the one below it. Each requires judgment at a different grain. HAAK provides infrastructure at every level; the question is where human judgment is essential vs. where computational assessment is adequate.
#How the origins connect
The three origins are not separate inspirations — they describe the same structure at different resolutions:
| Origin | Structural insight | HAAK instantiation |
|---|---|---|
| CNP Retreat | Perspective-taking in institutional roles reveals structural problems | Persona-grounded agents with latent variable weightings |
| Invisible College | Pre-institutional networks need indexing (Hartlib), convening (Haak), and irreducible expertise (Boyle) | Architecture provides indexing and convening; human participation provides expertise |
| Science Architecture | Science decomposes into figures composed by teams at multiple levels | Agent teams produce atomic units; composition is the PI-level function |
The formal theory (Library Theorem, institutional intelligence) came later. The origins are where the observations were first made.
haak · foundation · 2026-02-24 · zach + claude
Foundations 11 — Origins — 2026 — Zachary F. Mainen / HAAK