Christian T.
Sets the technical direction and builds the platform end to end: the memory substrate, the agent runtime and the governed-autonomy stack the studio runs on.
Cerebral is an applied-research studio at the intersection of agentic AI/ML deep tech and institutional advisory.
Building frontier systems and guiding the academic and enterprise institutions that put them to work.
Focus: agentic AI/ML deep tech. Model: technology × institutional advisory. Bridge: academia × enterprise.
How this studio actually works
We don't sell autonomy we haven't lived with ourselves. Memory, dispatch, containment, the gate and the ledger run our own work first, and yours after that.
Source · the gaze works register, snapshot of 2026-10-06.
Pilots and their agent lanes, the Cortex mesh that joins them, the production cluster where the stack runs, the GPUs at the edge, and Atlas reading the state. Names and statuses as recorded; no address, key or port on this page.
agent-session multiplexer
pilots · herdr panes
shared host · many lanes
opt-in, approved by a pilot
dozens of peers · inbox · lanes
the platform house
the brain · edge bridge
clustered
virtual keys · budgets
6 apps · synced
the gate plane
dispatch · 6 runners
signer · every 5 min
custody
memory of record
the leadership view, behind SSO
partner lab · science runs
● the DFlash benchmarkon demand
● training and swarm runsworkstation
● MTP · 68 → 110 tok/sBare-metal Kubernetes, GitOps-managed. Everything below runs here.
● 6 apps synced · healthyWhere many agent sessions build side by side: dozens of lanes, at the last count.
● runningHow agents talk: a mail-and-inbox protocol over gRPC, lanes per pilot, contract-net reassignment.
● dozens of peers · release candidateStatus and reports for the studio, behind single sign-on. Atlas is also the call sign of our agent workforce.
● liveThe platform and agent plane: Terraform, Helm, ArgoCD, the LLM gateway with per-consumer keys and budgets, single sign-on.
● 17 repositoriesEvery pilot runs agents in panes; a lane is a pane a pilot has approved onto the mesh.
● in useWhat every agent observed and promoted to canon, across harnesses.
● in productionAgent output held in a jail until a settlement lets it land.
● runningDecided work carried to runners; DAG plans; provenance end to end.
● 6 runnersEvery consequential verb classified before it runs.
● runningSigned every five minutes into an append-only chain.
● running4 × A100-40GB at a partner lab, 8 × A100 on demand, a workstation GPU.
● on demandSources · cluster probe of 28 September 2026 · mesh peers counted 5 October · estate register of 6 October. Statuses move; the register is the reference.
A studio run by builders and scientists, not by a deck. Christian T. and Patrick G., its technical core, bring more than 40 years between them building AI, cloud and internet infrastructure at scale.
Sets the technical direction and builds the platform end to end: the memory substrate, the agent runtime and the governed-autonomy stack the studio runs on.
Owns the engineering organisation and the production estate: the systems, standards and reliability that turn research into software that runs.
Runs the studio's finances and capital strategy, the discipline that keeps applied research funded and durable.
A particle physicist applying machine learning to LHC collision data. Brings hypothesis, measurement and statistical honesty to the studio's models.
Names are given as first name and initial, by policy. Each call sign is a real star, at its own coordinates.
Leads the identity authority for people, agents, machines and grants, and co-stewards the production engineering estate.
Core maintainer of Reverie and one of its two trunk maintainers. Runs the studio's agentic operations day to day.
Epigenetics and epigenomic medicine: how biological systems encode, regulate and inherit information over time. The natural analogue to a memory layer that forgets on purpose.
Econometrics and control theory pointed at the memory layer. Holds the studio's three hard lines at once: what gets built, what gets claimed, and the measurement both stand on.
Machine learning engineer and physicist applying statistical mechanics and model optimization to agent reasoning.
Frontend systems and rendering performance across the studio web platform and interactive tools.
“I use to play in a Jazz band called ‘The Cat's outa the bag’ we weren't very good.”
Creator of Autumn-Web and Autumn-Harvest, bringing zero-cost lifecycle and deterministic replay to agentic systems in Rust.
Runs the studio's network: private networking, access control and platform operations.
Coordinates J2000 and apparent magnitude V. waluku, the belt of Orion in Javanese, is placed on Alnilam.
The Cerebral crew is a free entity, not a payroll. Builders, scientists and operators join it, choose the work, and act wherever the work happens: in our own products, beside institutions, and in the houses of the estate.
People join the crew by doing the work with it. Names are given as first name and initial; call signs are stars.
When the zerodays incubator takes a founder from thesis to first build, the crew can sit beside them as the advisory layer: positioning, technical direction, governed agents from day one.
The same crew builds the platform, the products and the science, and lends its judgment to institutions putting agents to work.
Cerebral builds and operates its own agentic stack, and uses every piece of it on itself before it ships. Ten lines, from what an agent remembers to how its work is dispatched, measured and governed.
The consolidation substrate for AI agents: what an agent remembers, promotes to canon, and recalls across sessions. Context treated as infrastructure.
An operating system for running an enterprise on a persistent agent workforce. Every consequential action gated, authority delegated per verb and revocable, a human's name on every irreversible act.
Every agent run bound to its ticket, its memory and its outcome. One coordinator, many repositories, one command surface.
Continuous agent intent turned into discrete, counted ticks; multi-step plans run as DAGs. Custody provenance from origin to settlement.
Test regressions and convention drift kept as first-class facts and reasoned over in Datalog. The last incident gets retrieved instead of debugged again.
“We said we’d ship that”, with an owner, a cadence and a definition of done. Readable by every team and every agent.
What runs, what landed and what tripped a guard, on one screen. The operator keeps the final say on apply and discard.
One repo that takes an idea all the way to validation. Every graduated project starts in this registry, anchored to the canon from day one.
For the workflows off-the-shelf tools can't reach. High-stakes, domain-specific, and novel automation, engineered to measure and run in production.
For agent operations: enforced policy, gated actions, sealed append-only audit, and full-stack observability.
enforced > advisoryat every layer.
When a workflow is high-stakes, domain-specific or simply new, we build it to measure, and we run the stack it needs from the metal up.
Domain-specific multi-agent systems, engineered to measure and built to run in production rather than in a demo.
Exotic model development, fine-tuning, training pipelines and evaluation by an independent LLM jury.
Inside the client’s walls when the data cannot leave: desk-side units, air-gapped estates.
The studio’s own components, licensed into partner institutions.
Bare-metal Kubernetes, GitOps, self-hosted CI and SLSA-attested releases: the whole estate kept as code.
On-cluster serving, speculative decoding (DFlash, 188 tok/s, ~1.8×), multi-token prediction.
Reverie as the memory of record across harnesses, at 77.0 % recall@5 on LoCoMo.
Syscall jails, content-addressed custody, deterministic dispatch.
Every institution adopting agents hits the same wall: autonomy it cannot see, cannot audit, and cannot govern. Cerebral's value is making agentic behavior enforceable, not advisory.
Inputs, tool calls, outputs, cost, latency. You cannot govern what you cannot watch.
Who acted, on what, why, and the outcome. Accountability that survives the review.
Proceed, queue for sign-off, or block. Enforcement in the runtime, not a system prompt.
Revocable by signed amendment, recorded in the same chain as the decisions it governs. Enumerated and default-deny.
Disposition, tone, and boundaries mapped 1:1 to the real tool surface. Consistent behavior that doesn't drift or leak.
G-Eval + PoLL + ChatEval, before it ships. Measurable quality, not vibes.
Guidance for executives, boards, and business leaders facing the shift.
How an institution shows up in an AI-native market.
From thesis to first build for founders at the earliest stage.
Cerebral's own deep tech, licensed into partner institutions.
Engineers and working scientists built this studio together. That gives us three ways to serve research, and one ambition behind all of them: models that institutions can trust because the science behind them is shown.
Inference fleets, training pipelines, memory and orchestration for lab agents, on our own metal or inside the institution’s. The institute is the natural home for university agreements.
A particle physicist, an epigeneticist, an econometrician. Co-research, peer-reviewed work, and the people who measure the claims are the people who make them.
Bayesian multi-agent control; proper scoring rules for calibrated probabilities; prior-fitted inference for pharmacokinetics, which we call VesperKinetics.
The next AI systems are as much a science problem as an engineering one. On the left, the deep tech and infrastructure the studio builds. On the right, the working scientists whose research grounds it, with selected peer-reviewed work from the last three years.
77.0%Reverie66.5%BM25
102plain137MTP-1188DFlash
+~40%over MTP~1.8×over baseline · sirius · partner-lab A100s
68 → 110tok/s on a single RTX 5090 · polaris
Proof 77.0% → the LoCoMo run, 812 questions · 188 and ~1.8× → the DFlash run, 80 prompts × 11 domains · 68 → 110 → bench.unsigned.gg
Darius F. (call sign sirius) leads the science. He is a particle physicist who applies machine learning to LHC collision data, and a theorist of how information moves through populations of agents. He decides what science goes public.
his method
sirius et al.
sirius et al.
How much control a central controller can exert over a population of LLM agents through a shared board, posting only true facts, and what that control costs.
How a controlling agent fails, recovers, and what actually helps it keep its contract. Written into how Vespertin’s rules are drafted.
The same discipline pointed at the body: how a drug behaves in a patient from a handful of samples, every model decision standing before a regulator.
A loop between the fleet and the science.
No accuracy figure until a held-out test exists. What runs in shadow decides nothing. The research has started and it runs on compute, so partners and funders are welcome at [email protected].
The science, in full →The studio’s DFlash draft-model speedup on GLM-4.7-Flash, benchmarked across eleven task domains on partner-lab A100s and reproduced from the run. DFlash lifts throughput well above the MTP baseline, most of all on reasoning-heavy work.
plain decoding DFlash w8/v3 — each cube is ten tokens, emitted at the measured rate. Same seconds; the stream that runs denser is faster.
| Domain | plain | MTP-1 | DFlash k=3 | DFlash w8/v3 |
|---|---|---|---|---|
| Coding | 103 | 139 | 185 | 190 |
| Humanities | 101 | 131 | 141 | 146 |
| Math | 104 | 139 | 185 | 188 |
| Multiling. | 103 | 139 | 155 | 155 |
| QA | 118 | 145 | 150 | 157 |
| RAG | 88 | 111 | 139 | 141 |
| Reasoning | 105 | 138 | 173 | 175 |
| Roleplay | 100 | 118 | 109 | 116 |
| STEM | 109 | 139 | 160 | 163 |
| Summ. | 97 | 120 | 130 | 133 |
| Writing | 102 | 127 | 130 | 137 |
sirius · GLM-4.7-Flash on partner-lab A100s (4×A100-40GB, TP=4, BF16) · 80 prompts/domain · mean output throughput, tok/s
The whole estate behind the products and the advisory work: 40 repositories across 3 houses, all real, running code across memory, orchestration, containment and infrastructure.
plus a small fleet of utilities for deployment, packaging, voice and onboarding
The platform monorepo — Terraform, Helm charts, ArgoCD GitOps, the bare-metal Kubernetes cluster everything runs on.
The agentic OS — Reflex gate plane, signed decision chain, Marrow. Phase one built.
AI-containment custody layer — an OS syscall jail with reversible run-custody. v1.0, 500+ tests.
Agent-orchestration engine — DAG plans, deterministic dispatch, CRDT merge.
The audit & gate plane — agent forge, jury, runner fleet, the record behind the record.
Subagent scoring — an LLM jury (G-Eval + PoLL + ChatEval) with n_eff correction.
In-house fork of the herdr terminal multiplexer — the agent-session surface.
Estate TUI — onboarding and control from one terminal.
Scale-to-zero workload activator.
GPU-accelerated terminal emulator — wgpu renderer, ConPTY backend.
Bare-metal substrate — OS image and provisioning pipeline.
The edge estate as code — zones, workers, access, one apply.
Live agent-orchestration surface + LLM-gateway catalog.
The research cockpit.
Self-hosted CI runner fleets.
Cluster-audit PWA.
The unsigned design canon.
plus the Worker fleet — reverie-cloud · film · files-portal · works · about-worker · soma-brief-worker · gami-vst
Neuroscience-grounded memory-consolidation daemon for LLM harnesses. LoCoMo-validated.
The learning loop and operator console — dispatch, guard, settle surfacing.
Regression-memory + convention-drift framework — the studio learns from its own failures.
Incubator monorepo machine — ideation → execution → validation under enforced gates.
The CRM monorepo, built on Attio.
Headless Linear CLI — SLSA-attested releases.
Network policy as code, PR-gated.
A Zed fork wiring Reverie behind every agent — the reference integration.
The design system — tokens, components, the voice in code.
Fleet-board operations cockpit.
cerebral.work — the studio's landing.
Voice-first capture, v1.0 shipped.
The intake line — deal rooms and white-label deck surfaces, write-only by invariant.
Hardened ML container images.
Model serving on-cluster and at a partner lab.
The holding house's canon — marks, IP, the audit corpus. Counsel-gated.
zerodays.io — the studio's incubation brand and its own site.
The harness-config monorepo — every agent harness, its hooks, skills, and the estate brief's source.
AI-drivable containerized desktop sandbox for agents that need a real machine.
Pretty Good Types — a DSL for types and HTTP routes with full code generation.
The promise engine — parked mid-build, revives on demand.
Local caching daemon fronting the memory API.
Client for Even Realities G2 glasses — the estate on your face.
The full register, house by house, as kept on zerodays.io. Each card is a real repository; the visuals are the register's own.
The estate keeps a monument register, which we call the gaze: every repository under custody, every commit signed by a known hand, every session on the record.
Source · the gaze works register, snapshot of 2026-10-06. Counters exclude nine temporary working copies (git worktrees); the charts count every commit.
Gaze upon the works →Behind the studio stands Vesperan, the holding house of the estate. It owns the marks, the intellectual property and the floor everything else is built on, and it sells nothing.
Custody of the marks is the house’s only product: trademarks, copyrights, contracts. It stays quiet on purpose.
Incubation · AI-compute orchestration · sovereign generative AI.
Operating companies and products, embedded incubation: one operator, one founder.
An open crew of builders, scientists and operators that acts across the estate, and sometimes serves as the advisory layer for the zerodays incubator.
Agentic AI in the service of research is a real ambition, and it is not a product feature. Vespertin can serve verticals, even bespoke ones, while a research partner brings infrastructure, tools, broad expertise and co-research. That is why the studio is an institute.
What registered institutions get and companies do not, and the natural home for “infrastructure for acknowledgment” letters.
The people who measure the claims are not the people who sell them. They decide what science goes public.
Who it serves, what it offers and how it gets paid. Nothing is built before the study has answered those questions.
Testimonials, shows and talks, from the people who build it and the institutions that use it. Episodes are in production; these are placeholders for now.
What a signed record changed for the people who sign.
The studio on why proof, not permission, is the hard part.
Signed decisions for agents, in twenty minutes.
How an agent's behaviour moves, and how we measure it.
A day of agents at work under written rules.
Why a named decider matters more than a dashboard.
The science behind the studio's models.
VesperKinetics: the same discipline, applied to the body.