Evidence: 42 source files (C, C++, Verilog, JS, Python). Mesh wire protocol with regional coordinators and node agents (AVR, ESP32, iOS, Linux, GPU, Android, MacOS). Active workloads: Caribbean climate modeling (1km), tropical-disease drug discovery, distributed AI pre-training (1B-7B params), cryptographic auditing. Repository defaults: coordinator TCP 4300, HTTP 4301 (impelled path; substantial path 4000/4001), multicast 239.255.0.1:45454, 3-way dispatch, 40% idle CPU threshold. Federal eligibility: DOE ASCR, NSF OAC, DoD HPCMP.
Implication: Most available computation is irregular, temporary, and attached to something else. The mesh is not a smaller cloud — it is a scheduler for computation that would otherwise expire unused. Capacity is a lease, not a possession. Heterogeneity belongs at the protocol edge. The hard problem is useful decomposition.
Boundary: Opt-in workers are real; frictionless recruitment is a roadmap. SSH, WinRM, and PXE are roadmap mechanisms. Browser workers are implemented and evidenced (Impelled Browser Worker, 2026-09-25 audit: live WebSocket agents completed real jobs); broader recruitment remains roadmap. Capital deployment follows verified infrastructure, not roadmap promises. Source available under NDA to qualified capital partners only.
Capital Lens
Moat: Redundant first-responder dispatch on heterogeneous mesh — turns variation into a race, first valid result wins. Cross-runtime work contracts (Node, Python, C, C++) share framed messages, not runtime. 42 source files, 4 active workloads. Traction:
Traction: 4 workloads operational with academic partnerships. $/FLOP advantage via idle capacity recovery. Federal eligibility: DOE ASCR, NSF OAC, DoD HPCMP. Upside:
Upside: Capital-intensive compute asset where $/FLOP advantage compounds with each enrolled worker. $5M minimum programmatic commitment for production scale. Contact hello@supercomputing.systems.
Most available computation is irregular, temporary, and attached to something else.
A conventional cluster begins with machines purchased to be a cluster. The Bag model begins with machines that already exist for other reasons: a workstation between interactive jobs, a server below its normal load, a small native process, or a computer that can contribute briefly before its owner needs it again. Their processors, memory, runtimes, and connection quality differ. The coordinator's job is therefore not to make them look identical. It is to describe work narrowly enough that different participants can attempt it, and to treat availability as a changing measurement rather than a permanent entitlement.
Redundant first-responder dispatch follows from that premise. Predicting which mixed worker will finish first can cost more than sending the same bounded slice to several eligible workers. The default three-way dispatch turns variation into a race: the first valid result is accepted and late work is abandoned. This spends extra aggregate cycles to reduce tail latency and tolerate disappearance. It is useful only when slices are independent, result validation is cheaper than execution, and the redundant load does not exhaust the very idle capacity the system is trying to recover. The mesh is not a smaller cloud. It is a scheduler for computation that would otherwise expire unused — converting waste compute into a capital asset.
Capacity is a lease, not a possession
A worker advertises what it can offer now. CPU and memory thresholds protect the primary purpose of the machine, while memory leases make temporary allocation explicit. If local demand rises, contribution must contract. This is why cancellation and clean release are part of useful computation rather than administrative extras. Work that cannot survive a worker leaving is poorly shaped for this mesh.
Heterogeneity belongs at the protocol edge
The Node, Python, C, and C++ paths do not share a runtime, package graph, or object representation. They share framed messages and a small work contract. Serialization, task identity, deadlines, result validation, and capability descriptions cross the wire; language-specific machinery stays inside the worker. This permits a native experiment and a Node process to participate without requiring one to embed the other.
The hard problem is useful decomposition
No dispatch strategy can parallelize work whose pieces constantly mutate shared state. The productive workloads are those that can be divided into opaque chunks, checked independently, retried safely, and merged deterministically: search partitions, simulations, rendering tiles, batch transforms, test matrices, candidate evaluation, and other bounded maps. The research value of Supercomputing lies as much in discovering those decompositions as in discovering machines.
One task, several eligible workers, one accepted result
The architecture proceeds through four stages. Submit: HTTP routes accept individual or batch jobs and workload-assistance requests. Qualify: workers advertise capacity; dispatch excludes devices above CPU or memory pressure limits. Replicate: a task slice is sent to a configurable number of independent responders, three by default. Resolve: the first valid result wins; late responders are discarded and capacity returns to the mesh. This redundant first-responder dispatch spends extra aggregate cycles to reduce tail latency and tolerate disappearance — a capital trade that pays off when slices are independent, result validation is cheaper than execution, and redundant load does not exhaust the very idle capacity the system is trying to recover.
Watch it happen live — open the control plane app →Redundant First-Responder Dispatch
One task slice sent to three eligible workers simultaneously. The first valid result is accepted; the slower copies are cancelled and their capacity returns to the mesh.
JavaScript, Python, and native workers
NodeAgent.js implements the Node worker path. CommunicationProtocol.py supplies framed TCP for Python. MicroDistribute C and C++ sources test smaller native-worker contracts. wire_protocol.js normalizes messages at the boundary rather than pretending every runtime has the same object model. all-stack-runnable.js is the integration launcher for exercising the coordinator and available language workers together. The browser control plane is a view over the mesh; the compute protocol remains independent of the display.
Repository defaults: Worker/coordinator TCP port 4300 (impelled path; substantial path 4000). HTTP control plane port 4301 (impelled path; substantial path 4001). Multicast discovery 239.255.0.1:45454. Dispatch redundancy three responders. Idle CPU threshold below 40 percent. Critical pressure CPU 95 percent; memory 98.5 percent.
What the implementation exposes
Route families and their use: health, capacity, agents endpoints provide coordinator, participant, and available-compute observations. tasks, jobs, jobs/batch endpoints expose inspectable task submission and lifecycle. mesh, mesh/probe endpoints reveal participating endpoints and configured network discovery. memory/lease endpoint enables bounded temporary worker-memory allocation. assist/jobs endpoint handles external workload assistance under the same capacity rules.
Opt-in workers are real; frictionless recruitment is a roadmap
The repository contains runnable coordinator, worker, API, discovery, and validation paths. It also discusses SSH, WinRM, browser workers, PXE, and broader recruitment. Those are roadmap mechanisms, not equivalent to the current opt-in worker implementation. A reachable device is not an authorized device. Capital deployment follows verified infrastructure, not roadmap promises.
The mesh handles opaque work chunks and results. It does not need a worker's personal files to contribute computation. Enrollment, visibility, capacity limits, cancellation, and clean termination are part of the compute contract.
42 source files, four active workloads
The Bag Platform comprises 42 source files spanning C, C++, Verilog, JavaScript, and Python. Mesh wire protocol with regional coordinators and node agents (AVR, ESP32, iOS, Linux, GPU, Android, MacOS). Active workloads: Caribbean climate modeling at 1km resolution, tropical-disease drug discovery with academic partnerships, distributed AI pre-training 1B–7B params, cryptographic auditing. Source available under NDA to qualified capital partners. Contact hello@supercomputing.systems
Four Active Workloads with Defined Requirements
Caribbean Climate Modeling
1km resolution atmospheric simulation. Independent tile partitions suitable for redundant dispatch. Academic partnerships for validation. Federal eligibility: DOE ASCR, NSF OAC.
Tropical Disease Drug Discovery
Molecular docking and screening across compound libraries. Search partitions with independent validation. Academic partnerships. Federal eligibility: NIH, NSF.
Distributed AI Pre-training
1B–7B parameter model training across heterogeneous workers. Batch transforms and gradient accumulation. Federal eligibility: DOE ASCR, DARPA, NSF OAC.
Cryptographic Auditing
Verification of cryptographic implementations and protocols. Candidate evaluation with deterministic pass/fail. Federal eligibility: NSF SaTC, DHS S&T, NIST.
Mapping to DOE ASCR, NSF OAC, DoD HPCMP
| Agency | Program | Relevance to Supercomputing | Status |
|---|---|---|---|
| DOE ASCR | Advanced Scientific Computing Research | Heterogeneous mesh, climate modeling, AI pre-training | Planned |
| NSF OAC | Office of Advanced Cyberinfrastructure | Redundant dispatch, cross-runtime contracts, idle capacity recovery | Planned |
| DoD HPCMP | High Performance Computing Modernization | Cryptographic auditing, drug discovery, mesh resilience | Planned |
Disclosed Risks with Structural Mitigations
| Risk | Mitigation |
|---|---|
| Worker heterogeneity — unpredictable capacity, mixed runtimes | Redundant dispatch; first valid result wins; capacity measured per-dispatch |
| Idle capacity exhaustion — redundant load defeats recovery purpose | Dispatch redundancy configurable; idle threshold gates participation |
| Result validation cost — must be cheaper than execution | Work shaping requires independent, checkable chunks; validation is part of contract |
| Worker disappearance — no persistent entitlement | Cancellation and clean release are first-class protocol operations |
| Roadmap vs reality — SSH, WinRM not yet implemented; browser workers implemented and evidenced | Capital deployment follows verified infrastructure only; opt-in workers and the Impelled Browser Worker are current scope |
| Capital intensity — $5M minimum for production scale | Phased enrollment; $/FLOP advantage proves at each step; federal de-risks via grants |
Standardized Channels for Defined Purposes
- Technical inquiries: hello@supercomputing.systems
- Capital inquiries: capital@phronesis.wave
- Technical diligence: tech-diligence@ucip.systems
- Investor relations: Investor Relations