Skip to content

INDEX.md

Latest commit

 

History

History
193 lines (155 loc) · 9.48 KB

INDEX.md

File metadata and controls

193 lines (155 loc) · 9.48 KB
Raw

Index — Current State of Understanding

A fresh visitor (human or AI) should read this second (after README.md). It is the working-memory entry point and is rewritten periodically as understanding evolves.

Last reviewed: 2026-04-28 (post-PR-#5)

What we are working on

Canonical statement: D-2026-04-28-009 records the program's primary objectives and research strategy in three layers (north star, mid-level capabilities, operational target) plus six methodological commitments. Read that artifact for the full framing; the paragraph below is the one-paragraph summary.

The research program is to develop novel artificial neural network topologies that more closely mimic neocortical computation than current transformer-based systems, with the long-term aim of architectures capable of artificial general intelligence. The operational target is a stateful, deterministic, hybrid Bayesian prediction machine supporting continual learning, reference-frame adaptation, world-model construction, and attention-modulated prediction. The method is deep mechanism-first understanding of MDEMG (operational substrate) and tbp.monty (canonical Thousand Brains Theory implementation) before any integration proposal.

The current proposed framework name is AHH (Adaptive Hebbian Hierarchy). The framework name may evolve.

What we have decided

Thirteen decisions to date (eleven on 2026-04-28, one on 2026-05-21, one on 2026-08-10). Grouped by purpose:

Vault & repository foundation (bootstrap)

  • D-2026-04-28-001 — Vault substrate is markdown + YAML frontmatter, one artifact per file, six artifact types, in git.
  • D-2026-04-28-002 — Repository at github.com/reh3376/ann-research, public, MIT licensed, Go CLI base + Python sidecar.
  • D-2026-04-28-003 — Branch protection: PRs required for assistant contributions; @reh3376 admin bypass; CODEOWNERS = @reh3376.
  • D-2026-04-28-004 — Working agreement: no integration / interop / mapping work until T001 and T002 are complete.

Research strategy

  • D-2026-04-28-005 — Trace target: five_lm_no_threading.yaml config in tbp.monty (5 LMs, fully connected vote matrix, deterministic, bounded). Read-until-ambiguous methodology with run-and-instrument as authorized fallback.

Workflow operationalization

  • D-2026-04-28-006 — Assistant direct push via fine-grained scoped GitHub token (Contents:rw, Pull requests:rw, Metadata:r); replaces the originally-assumed patch-and-apply mechanism in D-003.
  • D-2026-04-28-007 — Every assistant-authored PR gets an activity summary comment posted after creation. Captures trajectory the PR description doesn't (sources, granularity decisions, confidence calibration, things considered and rejected).
  • D-2026-04-28-008 — Merges via admin bypass rather than approval-then-merge, because GitHub's platform-level "PR-author-can't-approve" rule prevents the approval workflow when the assistant operates under @reh3376's identity. Substantive control (Roger reads each PR before merging) is preserved as behavioral commitment.

Strategic

  • D-2026-04-28-009 — Program objectives and research strategy. Canonical statement of what this program is for. Three-layer objectives (north star, mid-level capabilities, operational target) plus six methodological commitments. Read first by any fresh visitor.

Architectural

  • D-2026-04-28-010 — Hybrid superstructure with LLM as language plug-in. Canonical statement of what this program is building. The architecture is a learning, predicting, world-modeling superstructure with an LLM as one of its faculties (specifically the language faculty). The superstructure is the cognitive core, the seat of self, the locus of continual learning. The relationship between superstructure and LLM is teacher-pupil — default non-intervention, override when warranted, continuous training of the LLM's LoRA adapter library based on the teacher's evaluation. Inter-LLM communication uses a token-efficient AI2AI protocol. Read alongside D-009.
  • D-2026-04-28-011 — V-JEPA family as candidate visual faculty. First specified non-language faculty in the hybrid architecture. Predict-in-representation-space architecture with demonstrated LLM integration, action-conditioning, and open-source availability. Adopted at family level (V-JEPA 2 / 2.1 / successors), for visual modality only. Long-horizon prediction limitation surfaces Q-010 as known gap.

Framing & evidence (2026-05 onward)

  • D-2026-05-21-013 — Spatio-temporal awareness supersedes the temporal-only framing.
  • D-2026-08-10-014 — Neuroimaging evidence pivot: the program is stalled on evidence; commits to acquiring direct measurement of the mammalian neocortex via an incremental modality ladder (OBS-2026-08-10-005), with the terminal requirement (human, mesoscale, millisecond, column-level dynamics during continuous learning) held as asymptote, not near-term build target.

What we have ruled out

  • DE-2026-04-28-001 — Producing mapping tables or interop documents from structural reading alone. Specific failure mode: pattern-matching on shape rather than tracing mechanism. Terminal.

What we are investigating

Seven active research threads:

ID Status Current question
T001-monty-trace active What does a single Monty episode actually do, mechanistically, from observation through motor action?
T002-mdemg-trace active What does an MDEMG ingest+retrieve cycle actually do, mechanistically?
T003-grounding-question active What does it mean for a symbolic substrate to have a "where am I" signal of the kind proprioception provides Monty?
T004-entity-first-principles active What is the minimal first-principles framing of what kind of entity this program is designing — from the entity's own perspective, before it has any models of self or world?
T005-program-planning active What is the next cycle's focus, what mode are we in, and what does it take to reach the next unblock gate?
T006-architectural-innovation active What is the architectural innovation that propagates the intelligence explosion, and how do we sharpen it from idea to specification?
T007-neuroimaging-evidence active What evidence about neocortical organization — acquired from existing modalities and datasets first, purpose-built instruments only where a proven gap remains — can support, refine, or falsify the program's substrate-level commitments?

T001 and T002 are mostly trace work that proceeds through reading and reporting. T003 is synthesis work that depends on the trace threads and on theoretical reading. T004 is conversational surfacing followed by deliberate capture; it is the internal-facing complement to D-009's external-facing objectives. T004 currently holds five open questions (Q-005 through Q-009) covering homeostatic boundary, dimensional commitment, projection-and-anchoring, recursive prediction-horizon expansion, and prime directives. T005 schedules the work that flows through the other threads via iterative cycles (plan→research→specification→plan→prototyping→validation→assess→unblock→plan). The plan-of-record lives at threads/T005-program-planning/plan.md and is revised at every cycle gate.

Open questions

  • Q-2026-04-28-001 — Do Monty's algorithms exploit properties of continuous 3D geometry that have no useful analog in symbolic structure? (load-bearing for whether integration is translation or substrate work)
  • Q-2026-04-28-002 — What is the symbolic-substrate analog of Monty's (object_id × pose) hypothesis space?
  • Q-2026-04-28-003 — What is the empirical grounding for MDEMG's etaConvMult / etaConfigMult / etaSameDirMult learning-rate multipliers?
  • Q-2026-08-10-015 — Is nanite-class BBB-crossing instrumentation eventually necessary to meet the terminal measurement requirement, or can intermediate modalities suffice? (strong-prior hypothesis with explicit revision conditions; nothing near-term is gated on it)

Hypotheses under test

None yet. The current state is "we have questions, not hypotheses." Hypotheses become appropriate once the trace work has produced enough mechanism-level understanding to make falsifiable claims.

What's load-bearing right now

The single most load-bearing artifact in the vault is D-2026-04-28-004 (the working agreement). It governs what kinds of artifacts the assistant produces during the trace phase. If it's violated, the program reverts to the failure mode in DE-2026-04-28-001.

The single most load-bearing thread is T003, because its answer shapes what integration work — if any — eventually makes sense. But T003 can't progress meaningfully until T001 and T002 produce evidence, so the practical near-term load-bearing is on those two.

How this index is maintained

Rewrite this file when one of the following changes:

  • A thread reaches complete or abandoned
  • A decision is committed or superseded
  • A hypothesis becomes supported or falsified
  • A new thread is opened
  • Roughly every two weeks regardless, to mark "still current as of"

The full per-artifact history lives in the vault folders and in git log. This file is the projection of "what's relevant now."