Master Plan
Current machine-facing baseline, without pretending the targets are finished.
Master Plan V220 is the proposed machine-facing reconciliation of Huey, the Monkey-Head-Project, the active V4 physical rebuild, and the next engineering gates. It is proposed and still requires Dylan’s human review before it can be treated as accepted project canon.
The website build and project-document version are separate version domains. This site release is build v217; the project-document standardization draft is V220.
How V220 labels truth
V220 separates what exists from what has been selected, what is a target, and what remains research. That distinction is part of the engineering record, not fine print.
Physically verified now
Observed build state, hardware on hand, working commands, and directly verified behaviour.
Selected for implementation
Architecture chosen for the build even where procurement, migration, or integration is incomplete.
Where the system is going
Desired capability or preserved architecture that still needs implementation and evidence.
Intentionally open
Decisions left open because the design, procurement, or physical testing is not finished.
Huey V4 — active physical build
V4 is the current embodiment. Huey Core is no longer the active machine name; where that term survives on the site it is historical proof-body lineage.
Rolling audio base
Salvaged wooden/audio base on four 3-inch casters, with three speaker housings and four speakers total. The marine-grade amplifier/stereo has already been tested successfully.
Compute midsection
A modified Thermaltake Mozart chassis sits on its left side and backwards relative to its original orientation, secured through the wooden interface platform with three bolts.
i9-12900K + ASUS Z790
The V4 primary compute baseline is an Intel Core i9-12900K on an ASUS TUF Gaming Z790-Plus WiFi motherboard.
Hybrid WowWee head
The V4 head is being rebuilt from two 2005 WowWee animatronic monkey heads. It is not yet mounted; Arduino-class low-level control is the present direction.
Brain, Body, Farm
Cognition and orchestration
Local model access, PyHuey/operator interaction, memory/state coordination, evidence, logging, and governance execution.
Embodiment and control
Sensing, actuation, audio, animatronics, power, cooling, environmental interaction, physical safety, and recoverable low-level control.
Expandable resources
Accelerator, model, storage, and support capacity used by the Huey node—including the planned four-V100 compute architecture.
Compute and storage direction
4 × V100 32GB / intended 128GB
One NVIDIA Tesla V100 32GB is currently on hand and three still need to be sourced. The intended unified 128GB resource remains a target that requires real topology and runtime validation.
Five-device PCIe path
A separate approximately 16GB general-purpose GPU is also targeted. Extenders/risers do not create PCIe lanes; the exact Z790 electrical topology and peer-to-peer behaviour still need proof.
Dual 1000W + air bring-up
Dual 1000W power remains the direction, but the final models/load split are provisional. Conventional directed air cooling comes first; the existing external liquid heat-exchange room remains a paused future path.
Optane root + mirrored bulk
Target root/OS is 4 × Intel Optane M10 16GB in RAID 10 (~32GB usable). `/home` is unresolved. Bulk data targets 2 × 10TB WD HDDs in RAID 1 after a safe backup/migration out of the current My Book Duo enclosure.
Cognition and internal architecture
PyHuey + local model service
PyHuey is the primary operator GUI/runtime/orchestration surface. Python 3.13 is the primary target; 3.12 remains a compatibility sidecar and 3.14 is under evaluation. Ollama is the preferred first service path, not a locked backend.
gpt-oss-120b — the walnut
The model is an encapsulated reasoning engine inside Huey’s governed framework, not unrestricted authority. gpt-oss-120b is the first large-model acceptance target; a V100-compatible multi-GPU runtime still has to be demonstrated.
Pebbles
Sandboxed jobs/roles with bounded prompts, context, permissions, parameters, triggers, and approved actions. The target is 128 logical roles per District—not 128 simultaneous LLM instances.
Districts + provisional Atom
The first Atom is a four-District grouping, with a working one-V100-per-District concept. Four Districts imply 512 logical role slots; actual concurrency and scheduling remain unresolved.
HIMS
Huey Internal Messaging System is the planned authenticated/encrypted routing layer between sandboxed components. Addressing, key management, persistence, routing, audit, and degraded behaviour remain unresolved.
Bifurcation
Exact, augmented, and hybrid bifurcation remain structured growth directions, not completed replication/scheduler machinery.
Governance working model
The V220 working model uses a President, a three-member Supreme Court, and a Senate. In the first four-District Atom, one elevated Pebble from each District is intended to fill one of the four special offices: one President and three Justices.
Each elevated Pebble keeps its ordinary job and receives extra compute/time for governance. Current research concepts include one vote per Pebble, District proposals and polling, multiple voting rounds where useful, and a deliberately difficult constitutional-amendment path. The present working threshold is 3 of 4 Districts, with two-thirds internal approval inside each approving District.
The Founding Father remains an accepted bootstrap concept for initial population/setup and limited founding tie-breaking. Governance timing, elections, terms, review procedures, and other mechanics remain research. These are not claims of implemented constitutional software.
Major engineering gates still open
- Map and benchmark the exact five-device PCIe topology; risers do not create lanes.
- Design directed airflow for passive V100s before sustained accelerator load.
- Validate the final dual-1000W PSU models, connectors, transient load, and load split.
- Back up the My Book Duo before any direct-SATA migration; enclosure RAID/encryption may affect accessibility.
- Test gpt-oss-120b on an actually compatible V100 multi-GPU runtime rather than assuming VRAM capacity is enough.
- Measure Pebble/District scheduling before treating 512 logical roles as inference concurrency.