Skip to content

FAIRE Structure v2 — The Guiding Skeleton

The canonical structure for FAIRE going forward. Established 2026-05-26 to replace the flat 15-track DL taxonomy with a 3-layer tree that scales, makes the Decision Engineering convergence explicit, and lets the four product pillars (Wikipedia + roadmaps.sh + Papers with Code + MVB nudge) land on every page.

Why a new structure

The previous tree (docs/curriculum/01-ai … 15-ml-theory-foundations) had three load-bearing problems:

  1. Wrong scope. All 15 tracks are narrow DL/ML subdomains. The user's curriculum includes humanities (Philosophy, Psych & Behavioral Econ, Writing & Thinking, History of Intelligence) — currently zero coverage.
  2. No convergence. Tracks are parallel silos with no structural pointer to the Decision Engineering thesis they're supposed to feed.
  3. Wrong unit. Pages are concept-only. There is no first-class representation of authors (Pearl, Olah, Karpathy, Kahneman, Munger…), learning arcs (roadmaps.sh-style), or builds (MVB recipes / paperswithcode-style projects).

The fix is structural, not prose-level. No amount of writer-prompt tuning will produce a Decision-Engineering-converging, author-anchored, humanities-integrated wiki out of a flat technical taxonomy.

Current scope — the 10 canonical tracks (user decision 2026-05-26)

The user has scoped v2 down to the 10 canonical tracks from pracha.me/curriculum. Co-curriculum and the Decision Engineering thesis are deferred — they remain in memory as future scope but are not part of the current build.

docs/curriculum-v2/
│
└── core/                              ◀── pracha.me canonical 10
    ├── 01-ai/
    │   ├── index.md                Overview · reading order · arc map
    │   ├── concepts/                  Encyclopedic, self-contained pages
    │   ├── authors/                   Person-anchored reading guides
    │   ├── arcs/                      Learning paths
    │   └── builds/                    MVB recipes / projects
    ├── 02-generative-modeling/
    ├── 03-representation-learning/
    ├── 04-neural-networks-deep-learning/
    ├── 05-statistical-probabilistic-ml/
    ├── 06-reinforcement-learning/
    ├── 07-attention-memory-reasoning-continual/
    ├── 08-causal-statistical-inference/
    ├── 09-algorithms-systems-for-ai/
    └── 10-complexity-cognition-natural-intelligence/

Deferred (not in current build)

  • co/ — humanities + meta-skills (Philosophy, Psych/BE, History of Intelligence, Writing & Thinking, Economics & Markets). Documented in agents/memory/faire_user_curriculum.md (Layer 2). To revisit once core/ is stable.
  • thesis/decision-engineering/ — the unified structural stack (Systems Eng · Contextual AI/ML · MIRO · X). Documented in agents/memory/faire_decision_engineering.md. To revisit once core/ is stable and the convergence story is clearer.

The feeds_de_pillar: frontmatter field is kept on concept pages — it lets future Decision Engineering work bolt onto the existing tree without restructuring.

Layer 1 — Core (10 canonical tracks)

The publicly-listed tracks at pracha.me/curriculum. These are the "range" of FAIRE's four-layer architecture (Curriculum → Arcs → Projects → Capstone).

# Track Slug
01 AI 01-ai
02 Generative Modeling 02-generative-modeling
03 Representation Learning 03-representation-learning
04 Neural Networks & Deep Learning 04-neural-networks-deep-learning
05 Statistical & Probabilistic ML 05-statistical-probabilistic-ml
06 Reinforcement Learning 06-reinforcement-learning
07 Attention, Memory, Reasoning, Continual 07-attention-memory-reasoning-continual
08 Causal & Statistical Inference 08-causal-statistical-inference
09 Algorithms & Systems for AI 09-algorithms-systems-for-ai
10 Complexity, Cognition & Natural Intelligence 10-complexity-cognition-natural-intelligence

The legacy tracks 11–15 (robotics, physics-scientific-ai, graph-relational, biology, ml-theory) are absorbed back into their canonical parents:

Legacy Absorbed into
11 Robotics & Embodied AI 06 RL · 10 Complexity/Cognition
12 Physics & Scientific AI 10 Complexity/Cognition (with arcs into 04, 05, 09)
13 Graph & Relational AI 04 NN & DL · 03 Representation Learning
14 Biology & Life Sciences becomes an arc, not a track (cross-cutting application)
15 ML Theory & Foundations 04 NN & DL · 05 Stat/Prob ML

Layer 2 — Co-curriculum (5 subjects)

Humanities and meta-skills that deepen the technical 10. Do not compete for time with core — they frame and contextualize it. Structurally required because they feed pillar 4 of Decision Engineering (X — Domain Context).

Slug Anchored in
history-theory-of-intelligence Consciousness, philosophy of mind, history of AI as ideas
philosophy Stoicism, existentialism, philosophy of science, ethics under uncertainty
psychology-behavioral-econ Kahneman, Cialdini, Duhigg, Frankl, Ariely
economics-and-markets Smith, Hayek, Christensen, Acemoglu, Munger
writing-and-thinking Orwell, Pinker, Olah, Karpathy, Montaigne, DFW, Feynman

Layer 3 — Thesis (Decision Engineering)

The capstone. Source: pracha.me/decision-engineering/decision-engineering-unified-structural-stack.html.

Four pillars:

  1. Systems Engineering — MBSE, system dynamics, organizational boundaries, feedback loops.
  2. Contextual AI/ML — Modular algorithmic components, not monolithic end-to-end.
  3. MIRO Stack — Modeling · Inference · Reasoning & Simulation · Optimization. Each is a subdirectory under pillar-3-miro/.
  4. X — Domain Context Variable — Domain theory, design thinking for adoption, tailored execution context. Fed primarily by the co-curriculum.

This is where the curriculum converges. Every concept page in core/ and co/ must declare which pillar it feeds via feeds_de_pillar: frontmatter.

The five page types

Type Filename pattern Purpose
subject-overview index.md (one per subject) Where to start. Reading order, key authors, arc map.
concept concepts/<slug>.md Olah/Distill-grade self-contained encyclopedic article. The bread-and-butter.
author authors/<lastname>.md Person-anchored reading guide: bio · arc of their thinking · key works · builds inspired by them.
arc arcs/<slug>.md Roadmaps.sh-style learning path. Ordered sequence of concept pages with rationale at each step.
build builds/<slug>.md MVB recipe / project. Reproducible, persona-tagged, runs end-to-end.

How the five page types map to the four product pillars

Product pillar Carried by
Wikipedia (depth, citations, neutrality) concept + subject-overview
roadmaps.sh (learning paths) arc
Papers with Code (SotA, reproducible code) concept's Current SotA section + build
MVB nudge (push to building) build + every concept's "What can you build next" block

This is what dissolves validation fatigue: each pillar has a home page type, not a "make sure every page does all four things" reviewer instruction.

Growth frontmatter (canonical)

Every page carries this frontmatter. New fields are additive — old pages stay valid.

---
title: <Human-readable title>
slug: <kebab-case>
layer: core | co | thesis
subject: <subject-slug>             # e.g. 01-ai, philosophy, decision-engineering
page_type: concept | author | arc | build | subject-overview
state: stub | drafted | reviewed | approved
authors_anchored: [pearl, kahneman] # who the page leans on; null for author pages themselves
feeds_de_pillar:                    # MANDATORY on concept pages; null elsewhere
  - miro-modeling | miro-inference | miro-reasoning | miro-optimization
  - systems-engineering | contextual-aiml | x-domain-context
arc_position:                       # only for pages embedded in an arc
  arc: <arc-slug>
  prev: <slug>
  next: <slug>
mvb_personas: [applied-researcher, research-engineer, applied-ai-engineer]  # the canonical 3
prereqs: [<concept-slug>, ...]
tags: []
updated: YYYY-MM-DD
---

feeds_de_pillar is the convergence enforcement field. The reviewer fails any concept page missing it. This makes the "everything converges on Decision Engineering" principle structural, not aspirational.

Quality bar (the readability rule)

Every concept page must be self-contained: a motivated reader landing cold should come away understanding the topic without chasing links. Olah/Distill grade, not Wikipedia-stub-with-links. Length floor: 1500 words; expected sweet spot 2000–3500 words for meaty concepts. Cross-links enrich; they do not replace explanation.

This is non-negotiable. See agents/memory/feedback_self_contained_pages.md.

Migration approach — Freeze + Parallel

User-confirmed 2026-05-26: build the new tree at docs/curriculum-v2/ while leaving docs/curriculum/ intact. The autonomous loop stays frozen until:

  1. The skeleton in curriculum-v2/ is built (this doc + empty directories + index.md stubs).
  2. agents/SCHEMA.md is updated to v2 (5 page types, growth frontmatter, convergence rule).
  3. Writer + reviewer prompts in agents/server.py are updated to read the new schema.
  4. At least one exemplar concept page has been hand-built end-to-end under the new schema as proof.
  5. The reviewer + critic panel has been re-tuned against the exemplar to confirm validation fatigue is resolved.

Only then does the loop resume — pointing at curriculum-v2/. The old curriculum/ is preserved as reference until enough v2 pages exist to publish the cutover.

What stays the same

  • The reader personas — now the canonical 3 (applied-researcher / research-engineer / applied-ai-engineer); the older 4-persona set was consolidated 2026-05-27 — see agents/SCHEMA.md.
  • The source policy (arxiv, .edu, huggingface, official docs only).
  • The arc / project / capstone framing from faire_canon.
  • The closed-loop observer / supervisor / scheduler architecture.
  • Budget control and the model-routing config.

What changes

  • Tree layout (flat 15 → 3-layer 10+5+1).
  • Page-type taxonomy (1 → 5).
  • Frontmatter (adds layer, feeds_de_pillar, authors_anchored, page_type).
  • Reviewer enforcement (the convergence rule + readability bar become hard fails).
  • Author pages become first-class generation targets.

See agents/SCHEMA.md (v2) for the page-body templates. See agents/memory/faire_structure_v2.md for the live decision log.