AI that keeps thinking
when half the team goes dark.
Picture six computers working together on one conversation — like teammates sharing the load. Now picture three of them getting unplugged, jammed, or destroyed mid-sentence. With Mycelium, the conversation just keeps going. No pause. No restart. No "something went wrong." That's the breakthrough — and you can watch it happen in three seconds, just below.
Three things you can verify.
Substrate properties, not marketing claims. Each one is observable in the public companion repo or in a 30-minute walkthrough.
Survives worker loss.
Substitute-on-failure dispatch at the per-token MoE expert dimension. When half the mesh becomes unreachable mid-token, the coordinator substitutes a zero vector for the missing expert and the dense MLP keeps producing. Output coherence degrades proportionally to loss, not catastrophically.
OpenAI-compatible surface.
Any client that speaks /v1/chat/completions works without code changes. The mesh is an implementation detail your operators never see — they get the same interface their tooling already targets.
Tamper-evident audit chain.
Every chat HMAC-chained into a canonical-JSON ledger, RFC 3161 trusted-timestamp anchored. Replay package available under evaluation NDA. A 3PAO can verify the chain end-to-end.
What survives when half the mesh dies.
Distributed inference frameworks don't document what happens when an inference worker becomes unreachable mid-token. Most options are catastrophic: hang, error, drop the request. Mycelium's coordinator substitutes a zero vector for the unreachable worker's experts and the dense MLP path keeps producing tokens. Output coherence degrades proportionally to loss.
Day-27 reduction-to-practice recorded 2026-05-08 with HMAC-chained tamper-evident evidence. Not in any chaos-engineering catalog. Not in MITRE's experiment-process methodology. Not in NIST SP 800-204.
Distributed adaptive AI mesh network
Predictive expert prefetching · cross-model style normalization · behavioral proof-of-inference verification
- › Substitute-on-failure dispatch on the per-token MoE expert dimension
- › HMAC-chained tamper-evident audit ledger over canonical-JSON
- › Behavioral fingerprinting + Merkle execution traces + economic reputation
- › Auditable replay package available under evaluation NDA
Same screen. Three seconds apart.
On the left: six computers, all green, working together on one answer. On the right: three are dead — and the answer is still coming through. Nobody hit a button. Nothing was rerouted. The system simply kept going. Substitute-on-failure dispatch at the MoE expert dimension — not in any chaos-engineering catalog, not in MITRE, not in NIST SP 800-204.
The user typing the question never noticed anything happened.
Filmed live · 2026-05-08 · HMAC-chained audit log · Replay package available under evaluation NDA
R2I-compliant by design.
"If you do not expose your interfaces and your documentation, you will not be able to join the ecosystem."
— Secretary of the Army Dan Driscoll · R2I hackathon announcement, 2026-05-05
- Exposed APIs — OpenAI-compatible /v1/chat/completions, /healthz, /mesh/status, /audit/verify, /v1/models
- Published documentation — OpenAPI 3.1.0, operator runbook, threat model, NIST 800-53 mapping, STIG-lite
- Vendor-neutral — any OpenAI-compatible client connects without code changes
- Modular Open Systems Architecture — coordinator + workers cleanly separated; pluggable fallback model
- Cryptographically auditable — every chat HMAC-chained into a tamper-evident ledger
- No vendor lock-in — cosign-signed binaries (Linux + Windows + macOS, RPM + DEB + .exe + .dmg)
Auto-tiered to whatever you've got.
Mycelium detects local hardware on first run and assigns a tier. The mesh aggregates tier capability live; available models grow as peers join.
| Tier | Hardware | Capacity | Role |
|---|---|---|---|
| Micro | Raspberry Pi · mobile | 0.6B – 3B | Edge routing, simple queries |
| Edge | Laptop (8GB) · Jetson | 3B – 8B | Token gen, classification |
| Standard | Desktop GPU (RTX 3060+) | 8B – 14B | General inference |
| Power | Workstation (24GB+ VRAM) | 14B – 70B | Complex reasoning · expert hosting |
| Fortress | Multi-GPU server | 70B+ | Heavy inference · training · aggregation |
Operator UX on any laptop. No GPU.
The evaluator stack is a brand-coherent replica of the operator chat + ops dashboard with synthetic-but-interactive worker state. Kill workers from the dashboard and watch the headline rewrite itself. ~50 MB image. Zero GPU. Zero model checkpoint.
For the production stack — six GPU workers running real Gemma-4-26B-A4B FP8 with substitute-on-failure dispatch — see docs/INSTALL.md.
$ git clone https://github.com/thornveil-ai/mycelium-overview.git
$ cd mycelium
$ docker compose -f docker-compose.evaluator.yaml up
# Then open
→ http://localhost:8094/ # operator chat
→ http://localhost:8094/dashboard # ops dashboardToday vs. roadmap.
A pitch lives or dies on candor. Here's what's shipped and what's still ahead.
v1.0.0 tagged
todayCosign-signed binaries on GitHub. SBOMs. Reproducible build.
Operator UI + ops dashboard
todayBrand-coherent web UIs. 5-minute evaluator stack via Docker.
Substitute-on-failure dispatch
todayReduction-to-practice recorded 2026-05-08 with HMAC-chained evidence. Replayable.
NIST 800-53 controls mapping
todayAuthor's interpretation. 41 of 44 in-scope controls met (2 partial, 1 N/A). Pending 3PAO review.
Physical ruggedized-tablet validation
roadmapDemonstrated on RTX PRO 6000 with 6 simulated worker processes; physical bench scheduled when partner hardware lands.
NIST FIPS 140-3 certification
roadmapFIPS-cipher-list build option exists; module validation not yet pursued.
Build the AI tier of your stack on Mycelium.
Standards seat. Licensing terms. Live demo. Whatever shape the conversation needs to take, the door is open. Direct contracts open December 10, 2026.