Alpha Omega Labs · Seed · Europe 2026
Autonomous work needs an evidence layer.
Cortex governs what an agent may do, and records it while it happens. omegaXiv is the first extension on top — research, turned into a result someone else can re-run.
01 · The problem
Agents already act. Nothing records why they were allowed to.
- An agent acts across tools. An API call, a file written, a deployment, a purchase. Each one competent, each one in a different system.
- The permission was ambient. A key in an environment variable, a human saying yes in a chat, a role granted last quarter to somebody else.
- The record is assembled afterwards. By hand, from logs that rotated, in a document written by the person who most wants the answer to be yes.
Six months later
“Show us how this result was produced.”
The authority, the environment and the decisions are already gone.
The gap is not another model. It is controlled execution and a record that survives review.
02 · Why now
Capability, capital and sovereignty arrived first. The control layer did not.
The binding constraint moved. It is no longer what a model can do — it is whether an institution can put it into production and defend the output afterwards.
03 · What exists
We did not write a slide about infrastructure. We built the parts.
- 01Cortex · governed runtime
- 02omegaXiv · research extension
- 03quarks · run engine · alpha
- 04Cortex Control · control plane
- 05Cortex Code · coding agent
- 06Engram · organisational memory
- 07Gauntlet · agent benchmark
- 08Engineering Harness · internal
- 09Dentate · RLVR training
- 10Sema · language and runtime
- 11Spiral · architecture research
- –Sema as the Cortex runtime · planned, not built
omegaXiv seats on Cortex because that is what an extension is, and the one dashed outline is drawn because it is not in the set.
04 · Cortex · the platform
Cortex makes authority visible at the moment an agent acts.
Decide before, not audit after
Capability policy separates what an agent can do from what it may reach, evaluated at the call.
Execute inside a boundary
Isolated runtimes, bounded service bindings, deployment and rollback.
Record while it happens
Approvals, denials, versions and artifacts land in an append-only trail, correlated to the run.
05 · omegaXiv · the first extension
One extension proves the platform sells more than itself.
The loop
Question sources governed run manuscript, code, data and the context to run it again.
Why research first: it is the only market where the buyer is contractually required to produce evidence — and an accepted evidence format transfers to every other regulated workflow we sell into later.
Capacity is a spend decision, not an engineering risk. The architecture scales horizontally and already carries real research runs; how many tenants it serves is set by infrastructure budget — which is what the round funds.
06 · What is proven, what is not
There is real machinery. There is not yet commercial proof. We priced that in.
| Asset | State | What the evidence supports today | Boundary |
|---|---|---|---|
| Cortex platform | Operating | Policy, approval, audit, runtime and deployment paths, running the company’s entire estate | no external scale proof |
| omegaXiv v3 | Deployed | Running through Cortex with telemetry and public health routes; architecture scales horizontally | capacity not load-tested externally |
| quarks | Alpha | The run engine behind omegaXiv, exercised at working scale on our own estate | alpha testing |
| Cortex × omegaXiv mediation | Partial | Extension runs on the platform; full policy mediation of every research run is not yet evidenced end to end | first round milestone |
| Customers · auditor · ISO | Absent | No paying customer for Cortex or omegaXiv, no external benchmark rerun, no assessor opinion | what the round buys |
The premise: finance the transition from internally exercised technology to externally accepted evidence and repeatable revenue — not the invention of the technology, which already runs.
07 · Who we sell to
Three named B2B segments. Every account countable, every buyer already holds the deadline.
01 · Research evidence
Principal Investigator on an ERC, FWF or Horizon grant · Head of Research Data Management · HPC centre director.
Holds a mandatory data-management-plan deadline and a result that must survive project handoff.
33 accounts · €25k first engagement · direct award to €140k needs no tender.
02 · Product security
Product Security Officer · PSIRT lead · Head of Product Compliance, who signs the declaration of conformity.
Holds the Cyber Resilience Act Article 14 reporting duty and must show an established PSIRT process.
571 accounts · €31.5k first engagement · under the €50k signing ceiling on purpose.
03 · Operator evidence
Head of OT Security, who signs the self-declaration personally · CISO · Head of Third-Party Risk.
Holds NIS2, the Austrian self-declaration and a continuous control-evidence obligation.
120 accounts · €55k first engagement · sector direct award to €150k.
If we ran a campaign tomorrow we would contact those three role titles inside the 115-account executable subset, where 15 of the 20 Tier-1 routes are personal introductions. Not B2C, and not horizontal SaaS.
08 · Business model and market
Services buy the learning. The platform licence captures the value.
80 Austria · 28 Germany · 7 elsewhere. Derived bottom-up from 226 sources. None of it counts as pipeline until a first engagement is signed.
| First-engagement frame | Accounts | Price | Value |
|---|---|---|---|
| Research evidence | 33 | €25k | €0.82m |
| Product security | 571 | €31.5k | €17.99m |
| Operator evidence | 120 | €55k | €6.60m |
The venture test is the recurring one: 100 contracts × €300k = €30m ARR.
09 · Defensibility
Others generate, orchestrate or govern. We keep the whole path.
| Alternative | Durable execution | Capability policy | Sovereign deploy | Research UX | Rerun package |
|---|---|---|---|---|---|
| Cortex today | |||||
| Cortex + omegaXiv, target | |||||
| Temporal + an agent framework | build it | build it | build it | ||
| Palantir AIP, governance suites | |||||
| Elicit, Edison, Sakana | varies | varies | varies | varies | |
| Consultancy and documents | manual | manual | manual |
Hard to retrofit
Authority and evidence must be coupled to execution. A report written afterwards cannot recreate the decision boundary.
Specific to the organisation
Accepted schemas, policies, connectors and rerunnable packages accumulate as operational memory that does not transfer.
Where we lose
If buyers happily assemble best-of-breed tools and accept a hand-written evidence trail, the integrated product carries no premium.
10 · Team
Four founders, four risks already retired somewhere else.
Marius-Constantin Dinu
Research and product
Author of SymbolicAI with Sepp Hochreiter — CoLLAs 2024, 1.7k stars, cited publicly by Gary Marcus and by IBM and Microsoft AI leadership. PhD under Hochreiter. Serial entrepreneur.
Florian Zeba
Enterprise systems and scale
Principal enterprise architect for data and AI. Platform scaling hands-on and a legal background — the two things a sovereign-deployment contract actually needs.
Philipp Kreisel
Industrial scale
Has already exited. Co-founded Kreisel Electric and carried it through scale-up to the John Deere majority acquisition — operator experience that cannot be hired late.
Andreas Gruber
Assurance and channel
Managing director at TRUSTIFAI, working with TÜV Austria. Turns AI governance into inspection work certifiers accept, with connections across the certification landscape.
11 · The round
Seed · Europe · 2026
Scale the platform, reach the buyers, build the network.
Scale the product
Multi-tenant capacity, and one repeatable deployment across hosted, private and on-premises — the constraint here is infrastructure spend, not architecture.
Reach the customer base
Run the campaign against the three named segments, starting with the 115-account executable subset and its 15 warm Tier-1 routes.
Build the distribution network
Certification and channel partners who already sell to these buyers — and the investor network that opens the accounts we cannot reach alone.
Thank you for reading.
The site playing this record is a Cortex app.
So is this deck.
- 01Cortex · the estate
- 02graviton-theory · this record
- 03investor-deck · this deck
- 04sema-docs · language docs
- 05dentate · SSO-gated
One plate, one deploy path, exposure per app
cortex deploy --manifest .cortex/<app>.app.yml
Ghost in the Frame · Graviton Theory, 2026 · Marius‑Constantin Dinu
A0 · Appendix · diligence plates
Show the work.
Nine plates follow. The appendix exists so that every number in the twelve core slides resolves to a plate here, or to a source link in its own footer.
A1 · Maturity ledger
Product state and evidence grade are two different columns.
| Asset | State | Evidence grade | Investor-safe wording |
|---|---|---|---|
| Cortex platform | Operating | Internal measured | Runs the company estate; customer repeatability unproven |
| omegaXiv v3 | Deployed | Deployed, not load-tested | v3 runs through Cortex; the 18 Aug audit found v3 on staging and graded it a prototype |
| quarks | Alpha | Internal runs at scale | The run engine behind omegaXiv, on our own estate; no external run |
| Engram | Operating | Verified in production | Organisational memory, exercised in the 14 Aug 2026 production pass |
| Gauntlet | Operating | Recorded runs | Three suites with dated run records; no third-party rerun |
| Engineering Harness | Internal | Used daily | The harness that produced the estate; internal tooling, not a product |
| Dentate | Research prototype | Internal diagnostics | Working training pipeline; a research asset with no commercial surface |
| Sema | Research prototype | Narrow live use | 2 of 59 feature families shipping; its release contract says not ship-ready |
| Spiral | Research prototype | Internal experiments | The architecture Dentate trains; a research direction, not a capability |
Nothing in the core deck claims a grade above the one recorded here. Where the audit and the core deck disagree, this table governs.
A2 · Architecture
The product is an authority graph coupled to an evidence graph.
| Authority graph · Cortex | What it names |
|---|---|
| Identity | human, agent, runtime, organisation |
| Capability | tool, file, service, infrastructure |
| Decision | allow, deny, require approval |
| Execution | isolated runtime, bounded service binding |
| Record | audit, trace, artifact, rollback |
| Evidence graph · omegaXiv | What it names |
|---|---|
| Question | claim, constraints, acceptance criteria |
| Source | paper, data, citation, lineage |
| Run | code, environment, parameters, result |
| Review | challenge, rerun, compare, approve |
| Package | paper, code, data, evidence |
The unproven join
Cortex mediates our own estate and omegaXiv v3 is deployed through it.
Full policy mediation of every research run is not yet evidenced end to end. Closing that join — a signed run package carrying its policy trace — is the round’s first product milestone, not a claim made here.
A3 · 36-month plan
Capital releases proof in sequence. Hiring follows evidence.
0–9 months · close the proof path
- Cortex mediates omegaXiv end to end
- 3 paid design partners
- External rerun and assessor format
- Founder and IP governance complete
9–12 months · prove conversion
- 5 paid design partners
- First annual licence conversion
- Deployment inside 8 weeks
- Security certification programme
12–24 months · productize distribution
- 12 contracted organisations
- 2 signed channel partners
- ISO 27001 target
- Gross margin ≥65%
24–36 months · earn the Series A
- 30 organisations target
- €5m+ ARR target
- Gross margin ≥75%
- 2 repeatable evidence templates
Kill / narrow gate: fewer than 3 paid design partners by month 9. We narrow the segment and stop all scale hiring.
No-scale gate: delivery exceeds 2 FTE-months, or an assessor rejects the evidence form. We fix delivery before any channel work is funded.
A4 · Commercial hypotheses
The sales plan is a sequence of tests, not a list of logos.
| Hypothesis | Test | Pass gate | Failure response |
|---|---|---|---|
| Research buys a finished outcome | 15 founder-led interviews, 5 proposals | 3 paid partners by month 9 | narrow the segment, stop scale hiring |
| Outcome converts to infrastructure | annual offer at engagement midpoint | ≥50% pilot to licence | reprice, or reject the platform thesis |
| Delivery productizes | time and cost every deployment | ≤8 weeks, ≥65% margin by M24 | fix deployment before any channel |
| Evidence survives scrutiny | external rerun plus assessor review | accepted form, bounded exceptions | change the evidence contract |
| Channel reduces CAC | two co-sell experiments | signed distribution, one sourced deal | stay direct, do not count channel |
A5 · Operating model
€6m funds the proof path without requiring revenue to stay alive.
| Operating target | Y1 | Y2 | Y3 |
|---|---|---|---|
| Operating release | €1.30m | €1.62m | €2.06m |
| End headcount | 7 | 12 | 18 |
| Contracted organisations | 5 | 12 | 30 |
| ARR target | €0.3m | €1.5m | €5m+ |
| Gross margin target | 45% | 65% | 75% |
| Primary evidence gate | paid proof | conversion | distribution |
€6.00m - €0.24m - €0.78m = €4.98m
Gross raise, less a 4% financing allowance and a 13% reserve held back until the evidence gates are met.
No-revenue operating plan
The €4.98m release is budgeted with no customer receipts. Revenue extends the reserve rather than funding the plan. Organisations, ARR and margin are milestone targets: no customer evidence exists for any of them today.
A6 · Capital cases
Choose the raise by how many risks must be retired in parallel.
- Cortex × omegaXiv integration
- DACH design partners only
- Security programme
- Defer US entry and broad connectors
Risk: product, trust and sales proofs compete for the same small team.
- All proof tracks funded in parallel
- 18 end headcount
- 2 channel experiments
- €0.78m reserve (13%) preserved
Balance: enough time to earn repeatability without institutionalising pre-PMF burn.
- Parallel US entry
- 28 end headcount
- Domain connector programme
- Larger assurance budget
Only rational after 3 paid partners, or with a strategic lead underwriting distribution.
A7 · Category financing
The category is funded, and crowded. Our capital advantage is staying software-only.
| Company | Financing signal | What it does, and does not, prove |
|---|---|---|
| Temporal | $300m Series D | durable agent workflows matter — and raise “why not just assemble it?” |
| Lila Sciences | $550m total | AI science factories are funded; wet labs are capital-heavy |
| Sakana AI | $200m Series B | AI scientist plus sovereign AI attracts capital; a broad competitor |
| Edison Scientific | $70m seed | a research-agent demand thesis; not verified enterprise revenue |
| Elicit | $22m Series A | a paid research workflow exists; seat pricing is not our ACV proof |
Position: the missing layer — governed execution plus reusable evidence — not another AI scientist.
Capital discipline: connect to external engines and labs; never own wet-lab capex.
Financing amounts validate category activity. They are not used as comparables and they validate nothing about our demand, valuation or product-market fit.
A8 · Risks and diligence gates
The fastest route to trust is to make the red flags inspectable now.
| Risk today | Why it matters | Financed proof or close gate |
|---|---|---|
| No product revenue | nothing has been sold as Cortex or omegaXiv; no willingness-to-pay proof | 3 paid partners by month 9; no scale hiring before that |
| omegaXiv capacity | the 10,000-run figure is our own architectural assertion, never load-tested externally | an external load test published with its method |
| Mediation not evidenced | the Cortex thesis is not yet proven end to end on omegaXiv runs | a signed end-to-end run package with its policy trace |
| No ISO or auditor opinion | procurement and evidence-acceptance risk | independent rerun, assessor opinion, ISO 27001 programme |
| Founder governance | titles, commitments, IP, vesting and the cap table are unverified | founder-signed governance pack before circulation |
| Unsigned channels | venture-scale distribution is still a hypothesis | count only signed distribution plus one sourced deal |
| Delivery work vs product | cortex.a2olabs.com publicly lists Prisma and GEMBA Network as built and managed on Cortex; this deck counts no revenue from them | founders confirm commercial terms, then restate the customer line before circulation |
Every row is a condition we hold ourselves to, and every one is checkable against the internal ledgers named below.
A9 · Source ledger
Every number in the core deck, with its origin.
External · market, problem and category
- Gartner · proportional agent governance, 26 May 2026 · gartner.com
- McKinsey · State of AI 2025 survey · mckinsey.com
- PLOS Biology · reproducibility survey, 5 Nov 2024 · pmc.ncbi.nlm.nih.gov
- NSF · $400m national AI cloud-lab network, 22 Jul 2026 · nsf.gov
- European Commission · AI Factories, 10 Oct 2025 · digital-strategy.ec.europa.eu
Company and founders
- Portfolio states and thesis · a2olabs.com
- Cortex · public product surface · cortex.a2olabs.com
- omegaXiv · product page · a2olabs.com/work/omegaxiv
- Public repositories · github.com/Xpitfire
- Capability audit, 18 Aug 2026 · analysis/capabilities.md, internal
- SymbolicAI · arXiv:2402.00854
- Kreisel Electric · official milestones
- TRUSTIFAI · AI inspection with TÜV Austria · trustifai.at
Before external circulation: confirm founder titles and time commitments, portrait rights, IP assignments, vesting and the cap table. This draft is founder-prepared and has not been legally reviewed.
A10 · Product surfaces
What the parts actually look like.
Cortex Control · runtime dashboard
Cortex Control · diagnostic events
omegaXiv · publication workspace
Dentate · training run, research prototype