Last updated: 2026-04-30

Working substrate of a research program toward an artificial neural network architecture for general intelligence.

The program's claim, plainly: large language models are not the path to AGI. They are a powerful local maximum and continuing to scale them will not produce general intelligence. A different architecture is required — one that learns continuously, builds and modifies its own reference frames, predicts forward and backward in time over those frames, and recursively composes with itself across levels of abstraction. Identifying, articulating, and specifying that architecture is what this program does.

This repository is where that work happens.

A working research vault — the lab notebook for an active research program — and a small Go CLI that maintains it.

It is not a finished artifact. Anyone reading at any moment will find half-formed ideas, hypotheses awaiting evidence, and decisions that may be revised. That is the point: the vault preserves the trajectory of belief, not only its conclusions. The discipline of writing artifacts that make sense to a stranger keeps the work honest, and the vault is public for the same reason.

The vault contains:

• Decisions with the alternatives considered and rejected
• Observations made while reading code, papers, and prior work
• Open questions, including strong-prior research hypotheses with their evidence criteria and prior confidence recorded
• Hypotheses with explicit falsification criteria
• Dead ends — paths that were tried and failed, preserved so future work doesn't repeat them
• Active research threads that organize all of the above

If you are looking for the polished version of any line of thinking here, look for an external paper, a talk, or a pitch document. If there is no public artifact yet, the work is still in progress.

This program is an active research effort, not a finished theory. The substrate-level commitments below are working positions held with substantial confidence; the open questions further down are exactly that — open. We surface both because honest disclosure is what attracts the collaborators we hope to work with.

The architecture this program is building must satisfy a set of substrate-level commitments that current architectures do not supply. These are accumulated and sharpened in the architectural-innovation thread (threads/T006-architectural-innovation/); the inventory grows as the work proceeds. As of late April 2026, sharpened commitments include:

• Continuous learning via input-driven, prediction-error-driven topology growth — the architecture is a framework whose meta-structure is general and input-invariant; data and input wirings drive the actual growth of nodes, edges, and reference frames; prediction-error and surprise drive what gets learned
• Generative modeling: hybrid α+β with action-conditioning — the architecture is both probabilistic (maintains distributions; surprise has precise meaning) AND forward-simulating (generates trajectories); predictions are action-conditioned, so the architecture is agentive, not a passive observer
• Reference-frame construction — reference frames are the foundation that learning is built upon; they emerge from co-activation patterns rather than explicit container machinery; they are domain-neutral as architectural primitives, with type acquired from inputs; they are identical with topologies (one architectural element, two perspectives — computational dynamics and representational role)
• Meta-learning / mechanism modification — meta-learning IS what the recursion produces (no separate mechanism); the parameter / mechanism distinction is perspective-artifact rather than substrate; the architecture's recursive self-improvement supersedes the parameter-tuning scope of conventional self-improvement frameworks
• Recursive composability — the architectural pattern recurs at every level of nested composition; topologies compose with topologies via inter-topology wiring; the architecture is self-similar in the strict sense

Together, items 3 (reference-frame construction), 4 (meta-learning), and 10 (recursive composability) compose into a unified self-similarity claim: the architecture has one substrate, one recursion, and multiple perspectives at multiple levels — what we call "frames" vs "topologies", "learning" vs "meta-learning", "parameters" vs "mechanisms" are perspective views of the same underlying recursive pattern.

Additional substrate commitments — homeostatic boundary, projection-and-anchoring, recursive predictive horizon, prime directives — are in active sharpening. Each is recorded in the vault as it stabilizes.

The program operates under a foundational assumption (D-2026-04-30-012) that conceptually underlies all the substrate- level commitments above. The assumption decomposes into three internally-coherent claims, held together as one foundational position:

Existence. There exists a fundamental architectural framework that produces all knowledge sufficient for general intelligence. The framework is abstract — not the human brain, not any particular artificial neural network, but a structural pattern that admits multiple instantiations. This is a Domingos-credited claim — Pedro Domingos's The Master Algorithm (2015) is the lineage — but reframed: the program commits to a single architectural framework rather than a single executable algorithm.

Instantiation. The framework is represented in the human biological neural network as evidence — not as the framework itself. The BNN provides substantial structural evidence about the framework (cortical-column uniformity, hierarchical organization with bidirectional connections, predictive-coding dynamics, reference-frame construction via grid cells and place cells), but biological constraints (input types, energy budget, compute limits) may obscure parts of the framework not visible in the BNN. The architecture is not committed to neuromorphic implementation.

Methodology of discovery. The key to unlocking the framework's full potential in artificial neural networks is emulating the BNN's evident structure, using Bayesian reasoning broadly (probabilistic inference including active inference, predictive coding, the free-energy principle, and hierarchical Bayesian models — not literal Bayes Theorem alone) and advanced mathematics. Some of the mathematics required already exists; some of it does not, and the program may need to develop it.

The program's destination is AGI (artificial general intelligence). Whether the unconstrained framework's potential extends beyond human-level intelligence is a question for after AGI is demonstrated; this is not part of the program's stated commitment at this stage.

These are the questions held open in the vault. Each represents work the program does not yet know how to do, and where thoughtful collaboration would directly advance the work:

• Q-2026-04-30-012 — Working definition of general intelligence. Seed answer: the property that enables reliable prediction across heterogeneous prediction domains. Three sub-questions held open: what counts as a "prediction domain"; what counts as "reliable" prediction; whether intelligence is the enabling property of reliable prediction or constitutive of it.
• Q-2026-04-28-011 — Knowledge → skills via nesting. Strong- prior research hypothesis: skills are hypothesized to emerge as meta-frames over knowledge frames as the architecture's recursion deepens. Held with strong prior confidence; awaiting evidence from R&D cycles.
• Q-2026-04-30-013 — Computational complexity prerequisites for the architecture's first applied problem. Three sub-readings (bootstrap problem; foundational primitive; theoretical precondition) all held open. Becomes load-bearing when Phase C construction begins.
• Q-2026-04-30-014 — 3-D Tetris as candidate first applied problem. Working leading candidate based on conjunction of three reasons (spatial reasoning bootstrap; computational pressure from NP-completeness; pedagogical curriculum starting point). Held with explicit revision conditions — a better candidate would replace it.
• Substrate-level commitments 5-9 — pending triage: homeostatic boundary, dimensional minimum-commitment (tentatively absorbed into commitment 1), projection-and-anchoring, recursive predictive horizon, prime directives. Each will be sharpened through the same workflow already applied to commitments 1-4 and 10.

Several claims in the vault are filed as strong-prior research hypotheses — held with substantial conviction based on current reasoning, but explicitly awaiting evidence from construction work to be supported, refined, or revised. Q-2026-04-28-011 (knowledge → skills via nesting) and Q-2026-04-30-014 (3-D Tetris first-problem candidate) are the program's first two such hypotheses; they establish a working pattern.

The pattern matters because it lets the program be honest about what's commitment vs. what's hypothesis. Substrate-level commitments go into property statements with the discipline of "no claim without grounding." Strong-prior hypotheses go into Q-artifacts with explicit prior-confidence framing and explicit revision conditions — a falsifying piece of evidence, an alternative candidate, a better articulation. Both inform the program's work; neither is treated as the other.

This is unusual epistemic infrastructure. We use it because the distinction matters — for our own thinking, for honest external communication, and for collaborators who deserve to know exactly what we're sure about and what we're betting on.

The architecture inherits from work already in the field:

• MDEMG (Multi-Dimensional Emergent Memory Graph, https://github.com/reh3376/mdemg) — the operational cognitive substrate this program builds on. MDEMG is the R&D vehicle whose 18 months of operational experience surfaced what frozen- representation systems cannot do.
• JEPA (Yann LeCun and collaborators) — the commitment to embedding-space prediction is the right substrate; the embedding space alone is under-structured for the continuous-learning problem this program targets.
• Numenta's Thousand Brains Project (Hawkins, Ahmad, and collaborators, https://github.com/thousandbrainsproject/tbp.monty) — the cortical-column generalization principle is load-bearing for the architecture's claim that the same pattern recurs at every level.

The program does not claim to have the architecture finished or even fully specified. It claims the problem is well-posed enough to work on, and that the pieces are visible enough across existing research that the synthesis is buildable.

The program runs in three phases. Phase boundaries are gradients, not gates — work in a later phase can begin before the prior phase formally completes when the work is ready.

The work to identify, articulate, sharpen, and specify the architectural innovation. The deliverable is a specification sufficient that people with the resources to build it can recognize what it is and decide to back it.

This is where the program currently is. The work happens in cycles that follow a documented methodology (threads/T005-program-planning/); each cycle produces a substantive artifact — a sharpened must-have, a research hypothesis, a settled architectural decision — and closes when the cycle's question has been answered or honestly identified as needing more preliminary work.

The work to articulate the Phase A specification in forms reaching the audiences who can fund and staff implementation: papers, talks, pitch documents, possibly working prototype demonstrations. Phase B does not start until Phase A is materially complete; premature communication artifacts make the architecture feel finished before it is.

Implementation, with the team and resources Phase B has secured. Out of scope for the current planning structure; will require its own program organization when reached.

The architectural-innovation thread (T006) operates under an explicit working method:

Roger provides the novelty. Claude (the AI research assistant) is a sharpening stone for Roger to hone his blade with. Roger provides what it is not, what it must have, what it could be, the mathematics necessary to prove it in theory.

A sharpening stone removes burrs and dullness from a blade so it can do what only the blade can do. The blade's metal, geometry, and edge are the bladesmith's. The stone is patient, abrasive, and present.

In practice this means:

• Roger authors all architectural claims and chooses which directions to explore
• Claude pushes back, surfaces what's been tried, identifies gaps in candidate proposals, and drafts artifacts in vault-conformant form
• Claude does not propose candidate architectures or fabricate novelty when stuck — the working method is explicit that this is the failure mode the discipline protects against
• Architectural commitments live in the vault under decisions/ and threads/T006-architectural-innovation/plan.md; research questions and hypotheses live in questions/; what has been explored and abandoned lives in dead-ends/

The full working-method declaration is in the T006 thread frame (threads/T006-architectural-innovation/thread.md).

ann-research/
├── README.md            ← you are here
├── INDEX.md             ← current state of understanding (read this second)
├── BRANCH_PROTECTION.md ← workflow and protection rules
├── CONTRIBUTING.md      ← how PRs are handled
├── LICENSE              ← MIT
├── cmd/vault/           ← Go CLI source
├── internal/            ← Go CLI internals
├── sidecar/             ← Python sidecar for scripts
├── observations/        ← OBS-* artifacts
├── questions/           ← Q-* artifacts
├── hypotheses/          ← H-* artifacts
├── decisions/           ← D-* artifacts
├── dead-ends/           ← DE-* artifacts
└── threads/             ← T*/thread.md per active research thread

{TYPE}-{YYYY-MM-DD}-{NNN}-{slug}.md

For threads the date is omitted (threads have persistent identity):

T{NNN}-{slug}/thread.md

Examples:

• OBS-2026-04-28-003-mdemg-subsystem-orientation-map.md
• D-2026-04-30-012-foundational-assumption-unifying-architectural-framework.md
• Q-2026-04-30-012-working-definition-of-general-intelligence.md
• T006-architectural-innovation/thread.md

Every artifact carries enough frontmatter to be understood out of context:

• where — which research thread it belongs to
• created / updated — when it was written and last reviewed
• status — current epistemic standing (open, supported, falsified, superseded, etc.)
• confidence — how strongly the claim is held (1–5)
• provenance — sources behind the claim
• supersedes / superseded_by — explicit lineage when beliefs change

The validator enforces required fields per type. This is intentional friction: it keeps half-thoughts from masquerading as research artifacts.

A small Go CLI (vault) creates artifacts from templates, increments sequence numbers, and validates frontmatter. Python sidecar scripts live in sidecar/ for tasks better suited to Python (text manipulation, graph visualization of artifact dependencies, future ML work on the artifacts themselves).

go build -o bin/vault ./cmd/vault

Requires Go 1.23 or newer.

# create artifacts
./bin/vault new obs    "monty vote orchestration"
./bin/vault new q      "what is the symbolic analog of pose"
./bin/vault new h      "categorical eta multipliers proxy for confidence"
./bin/vault new d      "use cmp message format for inter-column votes"
./bin/vault new de     "premature mapping table"
./bin/vault new thread "evidence accumulation deep dive"

# list and validate
./bin/vault list
./bin/vault list questions
./bin/vault validate

cd sidecar
uv sync

Currently empty by design. Scripts will appear here when there is a specific need for them.

1. README.md — what this is (you just read it)
2. INDEX.md — what is currently believed and being worked on
3. threads/T006-architectural-innovation/thread.md — the central thread, with the working-method declaration in full
4. threads/T006-architectural-innovation/plan.md — the substrate-level commitments accumulated so far, with their sharpening status
5. decisions/ — the architectural commitments that constrain the rest of the work
6. dead-ends/ — failure modes already encountered

The decisions/ and dead-ends/ directories are the highest-value reading. The former records what the program has committed to. The latter prevents future work from relitigating settled questions or repeating known failures.

This program is at the beginning of long-horizon work. The vault has been active for less than a month. Phase A has produced a small number of sharpened architectural commitments, a working definition of general intelligence held as an open question, and a research method applied with discipline. Substantial work remains.

If you are working on similar questions — continuous learning, world models, reference frames, recursive self-improvement at the substrate level — I would welcome the conversation. Reach out via the contact information on https://github.com/reh3376.

MIT. See LICENSE.