Mycelic is building an enterprise intelligence platform that connects evidence across teams and systems while the raw data stays where it lives.
One architecture, the Mycelial Fabric: every system keeps its own evidence and exchanges only typed artifacts. Each film shows it in a different kind of system. None is a separate product.
Each film is a simulation rendered from code, not a recording of any system and not a measurement. The label on each track gives its status.
When one site fails silently, probes from three vantage points that cannot get through all stop at the same place, and the intersection of those failed paths locates the cause.
A timing error enters through the feeder every line shares, three lines fall out of phase, and readings from one station on each are traced upstream to the feeder before the lines restart in dependency order.
When one robot leaves its assigned path, its lineage shows the route it took and the robots it exchanged with, a boundary isolates it, and the rest of the fleet routes around the boundary.
Can an ontology that spans separate infrastructure systems expose autonomous agents as they cross between them, so that routing and independent verification identify the paths they took and a minimum cut isolates them?
No Mycelic capability for government exists, and none is deployed.
Take a network outage. The routing team sees one part, the change log holds another, the path probes a third. No one may hand over the raw data, so the parts never meet, and one early guess is copied between dashboards until it looks like agreement. Mycelic lets a team investigate across those boundaries. Each owner keeps its records. Only bounded questions, findings and corrections move.
Every claim below carries a type. A solid label is a measured Finding. An outlined label is open research.
Persistent memory of entities, time and evidence across a long-running investigation.
FindingF-03 · 83.2% on LoCoMo, local modelsIndependent sources compared with each other, not one model's opinion repeated.
FindingF-01 · 19 → 69 of 90 supported answers, one outage FindingF-02 · 0.778 capability survival, syntheticWeak local models upgraded through changes that are independently gated and reversible.
EditCandidateEC-01 · in research, no result claimedFirst application: telecom network operations (TraceNOC, a research prototype). Planned next: robotics. The same architecture in four settings.
A hierarchy of evidence, not a hierarchy of authority.
Every scope, from one person to the planet, is a container. A container owns its evidence, answers bounded questions, and never ships raw records upward.
Membership has a declared type and can overlap: a team spans cities, an enterprise crosses borders. Where something sits tells the system where to look and whom to ask. It is never proof that a parent may read a child's data.
At the apex of the design sits the world orchestrator kernel. It would route questions and hold findings with their lineage, and never the raw records.
Six typed artifacts move between containers. Raw records never do.
Questions, corrections and threats travel the same links between containers. Every link has a cost, and crossing it takes time. Protecting a network of autonomous agents means the first two arrive and the third is stopped. That is three problems on one graph.
Each link has a cost: delay, compute, exposure. Links that policy forbids are removed first. Dijkstra's algorithm then returns the cheapest path through what remains. It is exact, because no cost is negative.
d(s, t) = min Σ w(e), over authorized paths from s to tTriangulation is not consensus. A claim copied ten thousand times still rests on one root. Lineage records every derivation, so support is counted in independent roots, a withdrawal reaches everything derived from it, and a finding with too few roots is withheld.
κ(k) = min |F|, where every derivation of k meets FWhen a container is compromised, Tesseract computes the minimum cut: the fewest links whose removal separates what the threat has reached from what must stay clean. By the max-flow min-cut theorem, that number equals the count of link-disjoint paths still open to the threat.
min cut(S, P) = max flow(S, P)All three move in time. A threat reaches a container after the length of its shortest path from the source. A correction starts later and has to arrive first. The later a compromise is found, the more has been reached and the further out the cut must fall.
The graph below computes this on a small example. Each link's cost is also its delay. A threat uses every link; a question uses only the links policy allows. Click a container to make it the source.
Measured so far. Lineage-aware repair, in a synthetic benchmark where minimum-cut placement is the oracle upper bound of 0.889 (F-02). Withholding an answer when its evidence is missing, in the outage benchmark (F-01). Not built. Cost-aware routing across the ontology, Tesseract containment and the time model are specified here and not deployed. The graph is computed in your browser with Dijkstra's algorithm and a max-flow minimum cut; it is an illustration, not a recording of a running system.
Dijkstra, 1959 · Ford and Fulkerson, 1956 · Menger, 1927
Three separate studies. Benchmarks, not customer deployments.
We rebuilt the 2020 CenturyLink outage from public RIPE RIS and RIPE Atlas data: 180 questions, 90 answerable and 90 where the right answer is to withhold. Adding a deterministic evidence planner raised the number of correct, evidence-supported answers on the answerable questions.
Adding more agent roles on top of the planner gave no statistically established advantage and cost about three times the tokens. TraceNOC, our first telecom application, is a research prototype.
Benchmark. One historical incident. 90 answerable questions.
We ran repair strategies through a deterministic simulator (nine failure interventions, 30 seeds) and measured mean capability survival: the share of capability still standing afterwards, from 0 to 1.
Lineage-aware repair moved 5,448 bytes on average against 9,877 for full replication. The benchmark's artifacts regenerate byte for byte.
NeuralGraph is our open-source memory: one Rust binary, one SQLite file, local models. On LoCoMo under Mem0's published protocol, a fully local stack (9B extractor, 35B reader and judge) answered 1,281 of 1,540 questions correctly, as scored by the local judge.
The local judge replaces the protocol's GPT-4o judge. The routing result is on a fixed 282-question set with the same judge.
NeuralGraph on GitHub · MIT licenceBenchmark. Local models.
From agents using a world model to a world model that identifies and repairs failing agents.
The design: model weights live in shared memory attached over CXL (Compute Express Link), and an edit is a sparse, versioned overlay on the shared base weights.
No Mycelic weight-editing result is claimed yet.
MATH-500, after editing 131,072 weights. 0.016% of its weights.
In November 2026 MIT CSAIL researchers reported editing 131,072 weights of a 0.8B-parameter open model, with no gradient descent, and moving its MATH-500 score from 7.9% to 48.1%.
It is the authors' reported result and it is not peer reviewed. The edit removed a reasoning-loop failure. It did not teach new mathematics. What it suggests is that a small, targeted edit to a model's weights can remove a specific failure. Project Manta asks who may authorize such a change, on what evidence, and how it is undone.
MIT CSAIL, “Sparks of Weight-Space Programming”, November 2026.Design partners and research collaborators: write to [email protected].