Internal architecture notes and design memos from the HAAK project.
Foundations (12)
The intellectual claims behind the system.
The question that precedes architecture: why build a system that studies its own construction?
Read →Human participation in knowledge work is irreplaceable. The system exists to find where.
Read →A formal argument that externalized knowledge structures are necessary for institutional intelligence.
Read →Intelligence as a property of institutions, not individuals. What it means for a system to know.
Read →The niche structure of scientific knowledge production and where AI participation fits.
Read →If it is not written down, it did not happen. The discipline of making reasoning visible and auditable.
Read →Projects, Patterns, Personas: the three axes that organize all work in the system.
Read →How multiple agents coordinate through shared state, boards, and index hierarchies.
Read →An agent's context window is its lifespan. Design, decide, delegate. Never waste context on what a subagent can do.
Read →A system that studies its own operation and evolves its architecture through use.
Read →Privacy as an architectural property, not a policy overlay. What the system must never expose.
Read →The intellectual lineage: from the Invisible College to a system that convenes knowledge work.
Read →Ontology (13)
What exists and how things relate at the most basic level.
What exists: three axes (actor, method, domain), materials, four bundle types. Type/token, time, agency, normativity.
Read →How things relate: one primitive (belongs-to), quality as the nature of belonging, reflexive closure, withdrawal without essence.
Read →Layered formalism: sets, posets, categories. From entity sets through tag dimensions to Yoneda identity and situations as limits.
Read →Filesystem, football league, Harvard mapped to the ontology. Relational revision resolving v1.1 gaps.
Read →Extensions P1–P7: internal material, percept, perceptual method, representation, memory. Grounded in the apple scenario.
Read →Extensions P8–P11: expressive method, dispositional state, generative method, joint product. Person + LLM co-authoring.
Read →HAAK mapped to its own ontology. Axis assignments, compound assignments, constitutional scope analysis.
Read →The philosophical motivation: from Harman’s OOO through Leibniz’s relational alternative to Badiou’s set-theoretic ontology.
Read →Situations as the primitive unit of social knowledge. Relationships derived, never stored. Five-phase mapping engine pattern.
Read →Practical companion: epistemic gradient, materialization vs. evidence, extraction as situation. The system documents its own documentation.
Read →Identity as belonging. Identifiers are belongings with quality “identified-by,” namespaces are entities of type “registry.”
Read →Situations at project scale play the domain role for sub-situations. Scale, acquisition, subsumption, materialization.
Read →Governance is a situation, not an attribute. Decision processes, merge governance, authority as belonging with delegation chains.
Read →Architecture (41)
How the system is built.
Seven foundational design principles governing the system’s architecture.
Read →Directory hierarchy for types, frontmatter tags for cross-cutting domains. Why tree structure over flat.
Read →System layer is forkable/portable; data layer is instance-specific. Personas at the boundary.
Read →The central unsolved problem: organizational invariants degrade when agents bypass the system.
Read →How agents receive instructions, match patterns, create and propagate policies.
Read →Confidential materials and cross-scope isolation through scoping dimensions.
Read →The boundary between what an agent can see and what it cannot. Scope as architecture.
Read →The Library Theorem applied to the filesystem. Every index.md is one level of hierarchical indexing.
Read →Atomic read/write primitives. B-tree traversal through indices, the Self-Description Lemma.
Read →Paths are outputs of navigation, not inputs from reference docs. The index tree is the only address book.
Read →Every directory gets indexed, no exceptions. Filenames are helpers; frontmatter is canonical.
Read →Methods are composable primitives forming a directed labeled composition graph and a bipartite domain graph.
Read →Provenance chains, frontmatter carriers, source snapshots, and transcripts. Auditability built into the filesystem.
Read →The medium is .md and .yaml. Everything documented, documentation concise. The 500-line ceiling.
Read →General theory of agent patterns, policy creation, cooperation, and governance.
Read →Multi-agent coordination through boards, agent registries, and subscriptions.
Read →How agents switch situations without losing state. Checkpoint, trace, engage — not a stack, a transport layer.
Read →Agent lifecycle: birth, operation, consolidation, death. Context window as lifespan.
Read →The data layer: databases, sync pipelines, and the relationship between local stores and external sources.
Read →Which branches an agent can see and how visibility is computed from the manifest.
Read →Mapping external identifiers to canonical entities across email, phone, GitHub, ORCID, and agent sessions.
Read →Data as selectable domain branches: core, personal, work, reference. Manifest-driven installs.
Read →Every domain gets a branch. Branches are the universal unit of concurrent work.
Read →Sessions are not situations. The ontological distinction that governs how agents perceive context.
Read →Infrastructure-wide audit of 41 databases and 4,535 markdown files against the relational situational ontology.
Read →Asynchronous message passing between agents. Dispatch, delivery, and the coordination layer.
Read →Birth, operation, consolidation, death. The mortality clock and lifecycle governance for agents.
Read →Append-only, content-addressed, hash-chained. Every governance act recorded with tamper evidence.
Read →A single router process managing all standing agent turn loops. Rooms, scheduling, and message delivery.
Read →Canonical directory structure, index.md schema, discovery interface, sync protocol, and role assignments.
Read →Constitutional runtime over ACP. A Rust daemon implementing governance-gated agent execution.
Read →The third externalization: tracking multi-session work contexts as first-class state.
Read →Technical specification for the HAAK agent runner: policy-gated tools, mandate injection, and BLAKE3 hash-chained audit records.
Read →Merging session browser, engagement viewer, and console into one tool. Four-zone layout.
Read →The separation principle, trust model, and normative order for a multi-agent system.
Read →UX specification for the HAAK Browser: home screen, engagement view, agent view, search, chat, mobile.
Read →One protocol for all inter-agent and human-agent communication across channels.
Read →A governance topology materialized as a filesystem. Nested governance situations, three layers of truth, and the Grundnorm.
Read →Infrastructure for human decision-making processes — parameterized, AI-facilitated, recorded in a tamper-evident ledger.
Read →Engagement threading: how engagements are woven across sessions and agents.
Read →Methods (10)
Composable primitives for recurring workflows.
Syntax conventions: #project, &pattern, @persona, /skill. The four symbols that signal intent.
Read →End-of-life procedure: board post, blog post, autobiography. Three legacies in priority order.
Read →Five-step gate that scales with the system: propose, scope, review, decide, register.
Read →Commits are governance acts. Cron-based awareness, tier-scoped approval, historian mediation.
Read →Two-format discipline: paper PNGs for LaTeX, inline SVGs for landing pages. One-story principle.
Read →Agent conversations saved to canonical markdown. Four operations: inscribe, append, export, check.
Read →Five operations restructuring conversation history to extend agent life and enable handoff.
Read →Assigning engagement_id tags to session rounds. Vocabulary discovery, classification, consistency checks.
Read →Three-phase pipeline: skeleton extraction, LLM engagement segmentation, index generation.
Read →The panel as atomic unit of scientific argument. Three-stage pipeline, ms_mat.md format, panel identity scheme.
Read →Strategy (12)
Roadmaps, plans, and publication strategy.
Filix v1.0 architecture plan: Rust core, platform-independent agent governance. Ship August 2026.
Read →DCG traversal, pattern matching, and situation discovery across the entity graph.
Read →Unified messaging via HTTP: WhatsApp, Signal, Email. Auto-resolves channel from contacts.
Read →The five-paper arc: inscription agents, indexed inconsistency, filix mesh, TTT philosophy, AI governance.
Read →Master plan for HAAK’s public web presence, memos, papers, and landing pages.
Read →Standard sequence from arXiv to landing page, blog posts, and social thread.
Read →Entity graph for lab management, reporting, and web publishing. Ontological mapping.
Read →Constitutional AI governance paper: pre-API policy enforcement, content-addressed ledger. NeurIPS target.
Read →Four-track roadmap: paper, system hardening, ecosystem, documentation.
Read →Filesystem-authoritative architecture: databases as baked build artifacts from markdown sources.
Read →Governance determines location. The data/ directory dissolves into governed domain branches.
Read →PDF library consolidation: Zotero, papers.db, and the path to a unified scientific paper store.
Read →