One archive. A company that compounds.

Investor business plan

They are building bigger windows. We built a mind.

We didn't build a bigger window. We changed the math.

Three breakthroughs. Nobody has any of them.

• New geometry. Instead of searching through data to find a memory, we calculate its exact address using eight-dimensional geometry. The address is computed in microseconds and never slows down, no matter how much the system learns.
• Holographic fractals. Every memory is stored as a wave pattern. Add a new memory, add a new wave. The storage never gets bigger.
• Phase-conjugate recall. To recall a memory, we run the wave backwards. If the memory is not there, it says so. It does not guess.

Twenty million is an ask, not cash in hand. There are no customer logos yet. That is why this raise exists.

---

How to read this book

Partner meeting: start with the product one-pager, then this book. Diligence clocks and claim grades live in 09-raise/ — they are not this document.

---

Contents

1. The problem the market already pays for
2. The core innovation — perfect memory
3. Why the incumbents have to start over
4. The infrastructure under the memory
5. Unit economics
6. The team
7. The product family
8. How the products roll out
9. Go to market
10. The market we can take
11. The budget we take first
12. Financial summary
13. The $20 million
14. Twenty-four month roadmap
15. What Series A is for

---

Executive snapshot

Trinity Sky is building NEOMORPHIC SSI — sovereign superintelligence that runs on the customer's hardware. Write a fact once. Get the same fact back. If the system is not sure, it says not found. The archive lives on the buyer's machine. Every product we ship is that memory wearing a different door.

Intelligence owned by the organization that uses it.

The ask. $20 million to get the first enterprise customers live, prove the product, then launch the rest of the lineup on that same system. About 85 percent of the raise is product, deploy, evidence, and customer-side proof credits.

Why now. OpenAI and Anthropic are in an arms race to make the window wider. Every year it costs ten times more and it still forgets. Today's model is more intelligence, more GPUs, more data centers, more capital, greater dependence on centralized infrastructure. Ours is more efficient computation, customer-controlled infrastructure, persistent proprietary knowledge — sovereign intelligence. Their revenue model is charging you per token forever. Ours is infrastructure you buy once. That is how you take them to zero.

The team. Enzo Garoche (founder — intelligence architecture). Skyler Trotter (founder — sovereign compute and photonic hardware). Together they designed LUV after writing the existing quantum stack — Qiskit, Cirq, Strawberry Fields, PennyLane, MrMustard, Walrus — and replacing those pieces with one framework that automates the photonic operations. Peter Dwight Sahagen (co-founder and president) bootstrapped MetroMedia Fiber Network — the first publicly traded dark-fiber company — into a NASDAQ index stock that reached a $35 billion valuation, connecting the NYSE and the Chicago Board of Trade. Open seats after the raise: enterprise sales lead, enterprise-AI advisor, national-security advisor.

Series A path. $250 million target after either $80 million recognized run-rate, or fifteen logos plus one government task order of $15 million or more.

---

1. The problem the market already pays for

Enterprises already write large checks for artificial intelligence. They pay OpenAI for ChatGPT, Anthropic for Claude, Google for Gemini, Microsoft for Copilot, and the coding studios that sit on top of those models — Cursor among them. The models are brilliant. The thing those buyers think they are buying, and still do not have, is memory.

What the market actually rents is a sliding context window: a temporary field of tokens, bounded in length, biased by position, and discarded when the session ends. Last quarter's strategy call is gone the moment the tab closes. Next Monday the same company pays again — in tokens — to paste the history back in. That is not an archive. That is a scratch pad with a meter on it.

Liu et al. (2024) showed that a transformer does not read a long document evenly. Accuracy is high when the needed fact sits at the beginning or the end of the prompt, and it falls hard when the same fact sits in the middle. The industry now has a name for this: lost-in-the-middle. It is what attention does when you treat it as a filing cabinet. Length makes the problem worse, not better. As the window grows, usable context shrinks relative to the brochure. A million-token demo is still a positionally biased guess.

The industrial patch is retrieval-augmented generation and the vector database. Documents are chopped into chunks, embedded, and ranked by nearest neighbor. That is search. Search is useful. Search is not get. Approximate nearest-neighbor retrieval runs double-digit error rates — commonly in the high teens to forty percent. The language model then writes from an incomplete or wrong evidence set, and still invents on top.

None of those systems has a true not-found. When the knowledge is absent or buried, the model samples a fluent continuation. A polite cousin of the stored fact is billed as the stored fact. A confident wrong citation, with no way to say we do not have that, is the expensive failure.

Two more costs sit on the same invoice. The buyer's data lives on the vendor's hardware, so when the contract ends the "memory" ends with it. And every attempt to make the model remember is another pass through the window — another token bill to re-read the past. The incumbents have built a business on renting forgetfulness and charging to replay it.

Buyers are not waiting for a slightly longer scratch pad. They are already paying for a machine that will give them the original fact, refuse to invent one, and keep the archive on their own hardware.

2. The core innovation — perfect memory

Trinity Sky is permanent algebraic memory.

Write a fact once. Ask for it later. Get the same fact back — not a statistically probable cousin. If the system is not sure, it returns not found. It does not mint a fake memory to keep the conversation moving.

The store is a Fourier holographic reduced representation — FHRR. Associations live as structured patterns in a frequency-domain archive. Binding is composition: a key and a value become one object. Recall is the inverse of that composition. The operation that recovers the value is exact unbinding, not a similarity ranking. There is no token replay of the archive. The buyer is not charged by the token to remember who they are.

That inverse is the category break. ChatGPT, Claude, Gemini, Copilot, Cursor, and every RAG stack can ask "what is close to this query?" They cannot ask "what was bound to this key?" and receive the original. Their indexes support an inner product. They do not support an algebraic inverse. Trinity Sky does. For a single stored association, unbinding recovers the item that was written. When many facts share a trace, quality depends on how full the store is, not on where a fact sat in a prompt. Every stored item is equally reachable. There is no middle to lose, because position in a prompt is not an address.

The second half of the product is the refusal. A fail-closed threshold of ϑ = 0.70 sits on the recall path. Below that confidence the archive returns not-found rather than a fabricated value. The threshold is not a system prompt. It is a gate.

The archive lives on the buyer's hardware — in the building or in the buyer's own cloud. When the relationship ends, the memory stays. Every write is anchored so that a later change is visible: tamper-evident storage, not a mutable index that can be silently edited.

Speed is part of the value, and we will be exact about it once. Isolated recall on a quiet machine is 13.834 microseconds — fast enough that "look it up" is not a product delay. Under load we have also seen 5,127 microseconds against a 200-microsecond budget. Those are different experiments: a capability clock when the machine is still, and a still-open engineering number when it is busy. We show both. Neither number is the story. The story is that recall is a get, not a search, and that get does not replay the archive as tokens.

Side by side — internal testing vs published competitor figures.

Competitor figures are drawn from published benchmarks and pricing pages. NEOMORPHIC SSI figures are from internal validated testing. Diligence can re-run the clocks.

Perfect memory recall accuracy is the primitive: write once, exact get, fail-closed if unsure, no token tax on the past, data on the buyer's hardware.

3. Why the incumbents have to start over

You cannot bolt exact get onto a vector database and a context window.

They built their stacks, their indexes, their demos, and their billing on the old layer. ChatGPT, Claude, Gemini, Copilot, and Cursor all assume that memory is whatever still fits in the window, plus whatever search can approximately find. Their indexes are embedding spaces. Their flagship demos are long-context theater. Their invoices meter the re-reading of the past. Those are not settings. Those are the business.

A vector database answers a different question from the one a company actually needs. It returns the geometrically nearest stored vectors. It does not recover the value that was bound to a key. There is no unbind. When the embedding model is updated, the geometry of the space moves, and last year's address is no longer last year's address. Approximate nearest-neighbor search is designed to be fast and close, not exact. The language model on top remains generative. Even perfect retrieval does not give that stack a not-found state. It still has to talk.

That is why "we added RAG" is not a catch-up plan. RAG puts retrieved text back into the same attention field that loses the middle. It inherits the window, the positional bias, the token meter, and the hallucination floor. It adds chunking, embedding drift, and a search error rate measured in double digits. There is no operation in a vector index or a context window that recovers "the original fact, or silence."

To match Trinity Sky, an incumbent would have to replace the memory layer — not the chat skin. Indexes, demos, and billing would all have to change. Evaluation harnesses, fine-tunes, and enterprise contracts assume the old loop: retrieve something like it, stuff it in context, generate. Cloning a studio UI is a quarter's work. Rebuilding the thing those UIs sit on is a company.

A buyer who has already paid to embed a corpus has acquired a search habit, not memory. You do not convert a nearest-neighbor index into exact get by adding an API. You start over — new representation, new address, new recall operator, new failure mode (silence, not invention), new place the bytes live.

That is the moat. It is not a secret prompt and not a model weight someone can download next year. It is a different primitive. They sell approximate search over a rented window. We sell permanent get on the buyer's machine. The two stacks do not compose. The competitor who copies our interface still has no memory.

These three inventions do not exist anywhere else. Nobody has any of them.

4. The infrastructure under the memory

Memory that lasts still has to live somewhere, and it still has to be found without turning into search again.

Under the archive is E8-SSII: the geometric addressing, routing, and self-sustaining infrastructure that gives every stored fact a permanent address. Think of it as the street grid of the memory, not as another ranking model. The E8 lattice supplies 240 roots — a fixed, dense geometry for packing information and for talking to neighbors. Routing on that mesh is designed to stay constant-cost, O(1) in the engineering sense, so finding an address does not get slower just because the estate grew. The commercial point is simple. The archive has a place. Retrieval is not a guess among nearest neighbors. It is a walk to a known address.

Vector search shouts a description into the aisle. E8-SSII files to a catalog.

Because the address is geometric and permanent, a fact does not become less findable because it was written last year. It does not drift when someone upgrades a language model. It does not depend on being re-chunked, re-embedded, and re-indexed so that a chatbot can approximately find it again. The language model on top can still write prose. The model is not the filing system.

The same geometry is why the infrastructure is built to sustain itself. Conventional stacks depend on a control plane someone else operates. SSII is specified to heal, reconfigure, and scale from the structure of the lattice — self-healing, self-organizing, self-scaling — so the memory layer does not become another managed service the buyer has to babysit. Addressing is intrinsic, routing does not search, and the archive stays reachable without a vendor in the path.

Put the two layers together and the company is no longer "an assistant with a better prompt." It is an owned memory with a permanent address space, an exact get, and a fail-closed door. The incumbents can keep renting a longer window and a closer search. They still will not have an inverse, a hard not-found, or an archive that lives in the buyer's building and does not charge to be re-read.

That memory is the company. Every product below is a door into it.

5. Unit economics

The cost of thinking just went to zero.

That is 450,000× cheaper per recall than GPT-4 class RAG. We are not cheaper than OpenAI. We are a rounding error.

Their revenue model is charging you per token forever. Ours is infrastructure you buy once. That is how you take them to zero.

Because memory is geometry instead of brute force, the entire mind fits on one laptop. They need a billion dollars of GPUs. We need customer-controlled hardware the buyer already owns. Edge, phone, drone, and watch form factors are later sublattices of the same architecture — smaller device, same math.

6. The team

Enzo Garoche — Founder

The intelligence architect behind the NEOMORPHIC brain. Creator of AMMA Intelligence and discoverer of the relationship between Phi (φ) and data recall, unlocking fractal holographic memory that scales without degradation. Full-stack systems architect with deep expertise across AI/ML, compiler design, neuromorphic engineering, and distributed systems. Designed and hand-built every software layer of the NEOMORPHIC stack using first-principles engineering, not third-party frameworks. Works across 10 programming languages including Python, Rust, Go, Julia, Gleam/Erlang, and WebAssembly. The result: a vertically integrated IP moat with no direct competitor in five of its eight patent-claim categories.

On the language side he wrote the compiler half. Years of production work in Qiskit, QASM, Cirq, Julia, Rust, and Gleam — gate-model circuits, type systems, multi-backend compilers, and the software that has to remember what the hardware just did. He saw the same failure in every stack: the language talks about qubits as abstract objects. It never names the chip.

Skyler Trotter — Founder

The infrastructure architect behind NEOMORPHIC SSI's sovereign compute layer. Quantum photonic hardware engineer and the builder of the sovereign multi-node compute mesh on Apple Silicon, backed by production cloud on Hetzner — proving NEOMORPHIC SSI runs on customer hardware, not data centers. Founder of Quantum Light Technology, developing the analog optical layer that eliminates the quantum error correction bottleneck. Designed every hardware layer of the stack — from the post-quantum cryptographic mesh to the E8-optimized routing protocol. 21+ filed patents across quantum computing, photonic hardware, cryptography, and AI acceleration. Two decades of first-principles R&D — the hardware half of an IP moat with no commercial equivalent.

On the language side he wrote the light half. Years of production work in Strawberry Fields, Blackbird, PennyLane, MrMustard, and Walrus — continuous-variable photonics, Gaussian boson sampling, analog phase control, and the DAC and FPGA tables that actually drive a chip. He saw the same failure from the other side: every photonic language still makes an engineer hand-bridge the math to the heater voltages. Calibration cycles. Approximation. Brittle scripts that die when the chip changes.

How Enzo and Sky built LUV

They were not reading papers about quantum languages. They were coding in them — different tools, different layers, the same dead end.

Qiskit and Cirq treat a qubit as an abstract two-level object. Blackbird and Strawberry Fields get closer to light, and still lack type safety and field algebra. PennyLane, MrMustard, and Walrus each automate a slice — gradients, squeezed states, photon-resolved sampling — and leave the rest as a NumPy script. None of them compiles a single line into arithmetic, Reed-Solomon codes, qudit dimension, chip target, phase schedule, and DAC table.

So they stopped stitching those languages together and designed LUV — the first vector-native language with a photonic IR. One principle: type = physics = hardware. One declaration names the field, the error correction, the dimension, and the chip. The compiler verifies the program can run on that hardware before a photon moves. Nine backends (Python, Rust, Julia, Go, Metal, NIR, Beam, TypeScript, QIL). Zero runtime dependencies. The operations they used to write by hand across six toolchains are now one automated framework.

Enzo owns the compiler, the types, and the backends. Sky owns the photonic IR, the analog optical layer, and the chip ladder. Together they closed the gap neither language family could close alone.

Peter Dwight Sahagen — Co-Founder & President

He already built the pipes the digital economy runs on.

Peter bootstrapped MetroMedia Fiber Network from first principles — no category, no playbook — into the first publicly traded dark-fiber infrastructure company. It became a NASDAQ index stock and reached a $35 billion valuation. The network connected the New York Stock Exchange and the Chicago Board of Trade and helped enable the modern digital economy. That is not a consulting credit. That is a company he stood up, took public, and scaled to a valuation larger than most of the AI incumbents we are displacing.

That operating history is the commercial half of this raise. Enzo and Skyler built the mind. Peter has already taken physical infrastructure from a garage thesis to a $35 billion public company. Trinity is the same motion on memory instead of fiber: own the pipes, put them on the customer's side of the wall, and compound.

Technologist, economist, and serial entrepreneur with 35+ filed patents. Has collaborated with Nobel Laureates across fiber optics, laser technology, and PCR. Architect of BankFi/TradeFi systems, Zero CounterParty Risk frameworks, and decentralized finance infrastructure. Creator of Empowerment Capitalism, a next-generation economic framework for equitable access to capital.

Open seats after the raise. Named enterprise sales lead — not yet filled. Named advisor, enterprise AI sales — not yet filled. Named advisor, national security / ITAR — not yet filled. The $1.0 million sales line and the $2.0 million validation line are what fill those gaps: two account executives plus counsel, and an independent re-run of the recall clocks. The team gap is a hiring plan, not a hole in the product.

7. The product family

Trinity Sky is one infrastructure and a family of products that sit on it. The infrastructure is permanent memory with perfect recall accuracy: write a fact once, get the original back, on the buyer’s hardware, with cryptographic proof the archive has not been rewritten. If the system is not confident, it returns not found rather than a polite cousin of the truth.

Every product below is that memory wearing a different door. A competitor can copy a chat window in a quarter. They cannot copy an archive that already lives inside the customer’s building and already holds last quarter’s decision. That is the compounding story. We are not selling six disconnected tools. We are selling one install that becomes more valuable each time we put a new surface on it.

There are no customer logos yet. The prices below are proposed, not live. Twenty million dollars is the seed ask that funds the first installs and the team that repeats them.

7.1 The Memory — Neomorphic SSI

This is the land. The CIO writes a check for an archive the company owns. The first contracts are priced as Neomorphic SSI — sovereign structured intelligence — in three bands.

SSI-CORE is $2 million and a six-week proof of value. Memory in one division, two executive seats, on the customer’s VPC or data center. Five hundred named users and fifty builders. A $150,000 proof credit applies to the land.

SSI-DIVISION is $8 million and a ten-week proof. Isolated deployment, six seats, a software air-gap option. Five thousand named users and two hundred fifty builders. A $400,000 credit applies.

SSI-ENTERPRISE is $25 million and a sixteen-week proof. Air-gap as the default path, eight seats, and the install playbook that lets the customer repeat the pattern. Twenty-five thousand named users and one thousand builders. A $1.0 million credit applies.

These are not science experiments parked next to an existing OpenAI, Azure OpenAI, or Bedrock line. They are meant to take that budget. A large enterprise already spends in this order of magnitude on rented models that forget. We are asking to replace the amnesiac layer, not to sit beside it as a lab.

What the buyer owns is three layers in one contract. Persistent memory: associative holographic recall, an append-only write log, and records that are written once. A sovereign runtime: local models, an always-on pipeline, and the residents that keep the machine honest. Supervised executive seats: domain slots the buyer names, with a human above the line when money or law is at stake.

The technical claim is simple enough for a board. Incumbents search. They embed a document, find a neighbor, and hope the model answers correctly. We unbind. The mathematics recovers the item that was stored. Every write is anchored to a Merkle chain, so the archive is tamper-evident. A fail-closed gate refuses to invent. Isolated recall on ordinary hardware has been measured at 13.834 microseconds; under load we have seen a slower clock, and diligence sees both numbers. Reading the archive does not cost a token.

As logos accumulate, we expect commercial first customers in a $4–12 million band and sovereign flagships in a $15–40 million band, in year two and year three. We are not underwriting two $40 million flags in year one. Year-one planning is six commercial lands and two government vehicles — still unnamed.

7.2 AMMA

AMMA is the intelligence that lives on the archive. She is not a tab you open and re-brief every Monday. She already knows the company history, the last decision, and why it was made, because those facts sit in the archive, not in a context window.

The buyer experience is a sovereign Claude: fluent, present, and local. She is a resident of the install, not a hired go-to-market officer and not a seat on the org chart. She does not tour accounts or close rounds. She stays with the machine.

Self-healing is a capability, not the pitch. Fourteen monitoring channels watch the runtime; when something degrades, she corrects it without paging an engineer. A guardian resident admits writes so healing cannot quietly rewrite the record. The product is the intelligence that remembers. The meridians are why she stays up.

7.3 Sovereign Cursor

Cursor today is a language model looking through a window at a few thousand tokens of your repository and hoping the right function is still in frame. Sovereign Cursor is a coding studio on the same archive. It recalls the binding for a file, a function, or an architectural decision — exactly — instead of hoping eight thousand tokens are still in the window.

The refactor does not forget what it was doing. The review does not miss a dependency that scrolled off screen. The developer is not paying tokens to re-paste the module they opened yesterday. This is not a fork of a popular IDE. It is the memory product facing engineers. It lands after the archive is in the building; without that install, it is just another window.

7.4 NeoDrive

Dropbox is a folder. Its search is fuzzy. NeoDrive is the company memory: contracts, designs, code, decisions, and communications stored so the original comes back.

The commercial promise is Dropbox-class ownership with exact retrieval. A general counsel asks for the agreement that was signed, not a near-match. A product lead asks for the decision that killed a feature, not a summary that drifted. Files live on the buyer’s hardware. Folders do not remember why something was kept. NeoDrive does, because it is the storage face of the same archive the Memory install already runs.

Research continues on denser addressing and photonic media. That work is a roadmap, not a closed commercial fact, and NeoDrive is sold as get-the-original-back storage on infrastructure the customer already owns — not as a priced physics paper.

7.5 Akosha

A full human C-suite — CEO, CFO, CTO, CMO, CSO, COO, and the rest — costs on the order of $12 million a year. A single senior hire often lands near $1.2 million fully loaded. Most startups go without and pay in stalled decisions, technical debt, and cash surprises.

Akosha is ten autonomous executive entities in one cockpit, sharing one memory. The cost story from the Akosha paper is about $1,000 per month per role against that $1.2 million hire — a planning contrast, not a live price list. The CFO entity models runway. The CTO entity tracks debt. The CSO entity watches risk. None of them forget the last meeting.

Money and legal authority stay with humans. Akosha does not sign contracts, move cash, or close the books. The entities recommend and prepare; a person admits.

This is not the year-one SSI check. Akosha is only as good as the archive it sits on. It surfaces after the Memory is in production — a later cockpit for startups and mid-market teams that could never staff ten chairs, not a substitute for the Fortune-class install.

7.6 LUV

LUV is the language of this machine — designed by Enzo and Sky after years of writing real programs in the quantum languages that already exist, and finding that none of them would run the photonic stack end to end.

Enzo came from the compiler side: Qiskit, QASM, Cirq, Julia, Rust, Gleam. Sky came from the light side: Strawberry Fields, Blackbird, PennyLane, MrMustard, Walrus. Each of those tools automates one slice — a gate circuit, a photonic IR, a gradient, a squeezed state, a photon count. The rest is still a manual bridge: decompose the unitary, calibrate the heaters, emit a DAC table, hope the chip matches the type. LUV is the framework that automates all of those operations in one place.

The one-line idea is type equals physics equals hardware. A single declaration names the arithmetic, the error correction, the dimension, and the chip. The compiler checks that the program is physically realizable before it runs. Developers write against recall, not against a prompt window that forgets the binding.

LUV ships after the archive is in production. It is a developer surface, not a replacement for the coherence clocks that prove the Memory works. Seed capital underwrites the install that makes a language worth writing. The language company comes later.

8. How the products roll out

The calendar is a product sequence, not a second company. Wave 0 puts memory in the building. Later waves thicken isolation, raise seat count, and only then open the language and the startup cockpit. Public go-to-market for those later surfaces is funded from a Series A, not from the seed sales line.

Twenty million dollars is the ask that builds the proof-of-value factory, funds the first credits, keeps the integrity pipeline current, and hires a small sales team. It is not a street campaign and it is not a chatbot bake-off. About 85 percent of the raise is product, deploy, evidence, and credits.

Proposed SKUs. Three annual-contract bands. Same product. Wider archive.

Seat titles are the buyer’s to name from a catalog of ten supervised roles — strategy, finance, technology, security, and the rest. CORE’s two titles are chosen in the room with the buyer. Unattended legal or financial sign-off is not in any SKU. Neuromorphic silicon is an optional research sidecar at DIVISION and ENTERPRISE, not a condition of the contract.

A modeled proof converts at about 40 percent for CORE and DIVISION and 35 percent for ENTERPRISE. That band is a planning analog from early Palantir-style pilots. Trinity has not run these proofs as a measured cohort.

Four waves. Features may slip. The sequence does not: memory first, then isolation, then air-gap flagship, then the later surfaces.

Wave 0 is the wave the seed must complete to be a company. The proof looks like an install, a recall protocol, and a fail-closed demonstration. DIVISION adds isolation and an offline drill. ENTERPRISE adds an executive tabletop and a write-log replay. Hardware SCIFs and accreditation remain customer-furnished. A small-business innovation research award is a vehicle, not an $8 million government logo.

If the first phase does not convert, unused pilot credit is the first line we stop. Leftover seed stays in reserve or runway. The gate that opens Wave 3 is three signed logos or $15 million in booked annual contract value. Until then, the product calendar stays on memory.

The line that holds the family together is the same in every wave. Their AI forgets and bills them for the privilege. Ours remembers, on their hardware, with proof — and that memory is every product we ship.

---

9. Go to market

Trinity Sky sells organizational memory the buyer owns. Year 1 is a closed sales motion aimed at named Fortune 100 and government committees. The public campaign is a Series A product. It does not run on leftover seed.

The sequence is the underwriting. Seed proves a first enterprise customer will take the product under NDA, sit a proof-of-value, and install the archive on their side of the air gap. Expansion proves those logos grow and that government vehicles become real task orders. Only then does Trinity spend a public envelope, with a human sign-off, against a published curiosity-to-paid loop. The same memory the buyer licenses is what later remembers campaigns, KPIs, and decisions — so go-to-market compounds on the product, not on a rented marketing stack.

9.1 Seed: private sales

The $20 million ask puts $1.0 million on sales and $2.0 million on enterprise pilot credits. Two account executives plus counsel are the commercial stack. They work a named-account book: Fortune 100 buyers and government procurement committees. The motion is NDA, then a proof-of-value that installs sovereign memory — not a chatbot demo — then a signed contract that books annual contract value.

Proof-of-value length tracks the SKU. CORE is about six weeks, DIVISION about ten, ENTERPRISE about sixteen. Sales dollars open the door. The rest of the seed — proof-of-value engineering, air-gap deploy, enterprise integration — is what makes those weeks real on the buyer’s floor. Pilot credits sit on the customer side of the table so a committee can start without treating the first install as a science project.

The tickets this team is built to close are enterprise, not developer-tools ARR. Modeled SKUs run CORE at $2 million, DIVISION at $8 million, and ENTERPRISE at $25 million. Commercial first customers sit in a $4–12 million band. Larger sovereign flags are a Year-2 and Year-3 expansion story, not the seed close. Seed success is the first signed logos and the ACV those logos book.

The weekly scoreboard is operational and short: NDAs opened, proofs started, logos signed, ACV booked. Press, waitlists, and inbound curiosity do not count. Two account executives plus legal load at about $0.9 million; the $1.0 million sales line covers that stack. A full sovereign pursuit costs more than the seed sales line alone. The remainder lives in Year-1 sales and marketing once the company is alive and converting. Seed does not pretend a million dollars buys a national campaign.

The pipeline is unsigned today. That is honest, not empty theater: no invented logos and no invented awards. Management will fill three to eight named rows under NDA. Until those rows exist, Year-1 mix tables are planning constructs, not a CRM export.

9.2 The contract gates

Product features can slip. The calendar cannot start a public campaign. The gates below are what a term sheet should repeat.

Phase 0 (months 0–9) is closed first-customer business development: named-account work, NDA, proof-of-value, air-gap pilot. The gate is three signed logos or $15 million booked ACV. Either prong is enough. That is the expand decision. Miss it, and the company revises the sales plan. It does not declare a launch date.

Phase 1 (months 9–18) is land-and-expand. Existing logos grow inside the account — more divisions, more seats, more of the archive in production. Government vehicles are a path, not a trophy. SBIR and Other Transaction Authority can open a prototype and, later, a task order. They do not stand in for an $8 million government logo. A Phase I SBIR at the current $323,090 cap is first federal cash if it arrives. It is not an enterprise award. The Phase 1 KPI is $40 million recognized run-rate.

Phase 2 (month 18 onward) is the public brand. It starts after Phase 1, with a human Level-5 approval — not because a date arrived.

Series A is a $250 million target in a $200–300 million size band. That band is raise size, not seed pre-money. The trigger is either $80 million recognized run-rate, or fifteen logos plus one government task order of at least $15 million. If the trigger does not trip, the company bridges against a revised sales plan. Leftover seed does not become street teams.

Capability waves sit inside these phases. What ships in a proof, and what ships once a logo is live, is a product question. A slipped feature does not unlock the public campaign.

9.3 Series A public brand (summary)

After the gate, Trinity can run a public go-to-market. Seed sells to a named committee. The later campaign sells curiosity to developers and cities, then converts that curiosity into a paid loop.

The sequence is mystery, then a measured attention threshold before anything is named, then a genuinely limited drop, then reveal, then a paid loop. Physical and digital frequency feed one index. When the index clears, the company reveals — product in the exclusive API — and paid media learns against holdouts. Our post-Series A public campaign engine drafts and publishes once humans approve. It is not a seed hire. Humans stay on the chief executive, finance, and counsel desks.

The public envelope is funded from Series A. Three twelve-month cases: LOW $2.0 million, BASE $5.0 million, AGGRESSIVE $8.5 million. Contingency is explicit. Paid lines stop if thirty-day customer-acquisition cost exceeds a third of lifetime value. If Phase 0 does not convert, this campaign does not start. The company bridges.

9.4 Why GTM compounds on the same memory

Trinity Sky is an organizational-memory company that can eventually run its own go-to-market loop, because the same memory the buyer owns can remember campaigns, KPIs, and decisions.

That is not a claim that seed has a marketing agent. It is a claim about architecture. The commercial loop is input, memory, reason, action, observe, learn. Campaign weights, city frequency, attention scores, and conversion holdouts are another class of records. Once those records live in the archive the customer already licensed — fail-closed, exactly retrievable, cost independent of how much history is stored — the company does not rent a new brain every quarter to remember what worked.

For the buyer, that is the product: memory they own, on their side of the air gap. For Trinity, after the gate, it is also the GTM advantage. Street and out-of-home still cost money. The campaign engine still needs humans. The difference is that every cycle writes back into the same substrate. Awareness becomes users; users become subscriptions; subscriptions become revenue; revenue funds verified compute; compute funds better campaigns. The flywheel is the memory.

Year 1 does not need that flywheel to be public. Year 1 needs three logos or $15 million booked. The public machine is what those logos buy the right to turn on.

---

10. The market we can take

Gartner forecasts worldwide AI spending of approximately $3.49 trillion in 2027, including about $1.89 trillion in AI infrastructure, $759 billion in AI services, and $638 billion in AI software. Combined adjacent markets reach $10 trillion-plus by 2030.

We play where the buyer already writes the check — and we map each market to a door on the same archive.

Where we play first. Four initial markets, not six SKUs on day one.

1. Financial intelligence — hedge funds, investment banks, private equity, trading, wealth management.
2. Scientific discovery — pharmaceuticals, biotechnology, materials, chemistry, semiconductor research.
3. Engineering and industrial intelligence — aerospace, automotive, energy, advanced manufacturing, robotics.
4. Defense and sovereign AI — defense contractors, national laboratories, intelligence, government.

Those are addressable markets. They are not our Year-1 bookings. The first dollars we collect come from the invoices in the next section.

11. The budget we take first

Trinity takes budget from invoices the buyer already signs — model seats, token meters, systems-integration retainers, and the staff who re-paste last quarter into a rented window. The product is organizational memory. The sale is a substitution: own the archive, stop renting the scratchpad. The buyer is a committee that already pays OpenAI, Azure, Anthropic, or Bedrock. If the memory works, the token meter becomes the expensive way to remember.

A typical 10,000-employee generative stack is modeled at $4–8 million a year. A 100,000-employee estate is modeled at $30–55 million a year. Those figures are incumbent spend, not Trinity revenue. They are the size of the bill a CIO already explains to a board. On a constructed Fortune 50 three-year comparison, the same substitution produces a $90.9 million takeaway: what they pay the rented stack minus what they would pay to own the memory and run it on their own hardware. The first dollars we intend to collect are a handful of those invoices, not a percentage of a national appropriation.

The Department of War’s $58.5 billion AI request is a policy line. It is not our serviceable market. We sell a machine that sits inside the buyer’s perimeter and displaces a line item the CIO already defends. A government install, when it comes, is a task order against that same substitution — not a claim on the entire defense request.

The ticket size is not theoretical. Palantir closed 73 deals of $10 million or more in the second quarter of 2026. That print is existence of a market that writes eight- and nine-figure software checks. It is not our pipeline and not our bookings. It tells an investor that the buyer we want already signs checks this size.

What we put on the table is one product — sovereign organizational memory — priced as three SKUs. None is live. They are the offer we take into the first proofs.

CORE is the door. DIVISION is the expand — isolation and a software air-gap option. ENTERPRISE is the flagship: air-gap as the default path, the install a procurement committee can own when the data cannot leave the building. As the story grows, commercial first customers plan in a $4–12 million band and sovereign work in a $15–40 million band. Those are later planning ranges. Year 1 is not two $40 million flags.

The Year-1 base plan is $96 million booked and $48 million recognized, on a mix of six commercial and two government logos — none of them named today. Recognition is partial-year by design: first signatures around month five, a blended half of booked annual value recognized in the twelve months after close. The downside plan is $10 million recognized. Proof-of-value conversion of 40 percent on CORE, 40 percent on DIVISION, and 35 percent on ENTERPRISE is a planning analog from early-pilot bands in this category. It is not a Trinity cohort we have already run.

This is a private sales market. The motion is named-account development, an NDA, a proof on their hardware, and an install they own. Public brand work waits until logos exist.

12. Financial summary

These figures are a planning model. There are no signed logos. They are not bookings.

Y1 recognition is partial-year by design. The mix is six commercial and two government cells if the base path holds. Y3 and Y5 assume the archive compounds across the product family after the first installs exist.

Those first logos are the launch. They sit a proof on their hardware, apply a customer-side credit, sign, and expand. They do not back the cap table. They back the product — and that book is what unlocks Series A. Full mix, unit economics, and the five-year table: 05-model/FINANCIAL-PROJECTIONS.md. Line-by-line seed spend: 02-capital/COMPANY-UOF-20M.md.

Series A use of funds — $250 million target, $200–300 million size band, after the trigger in Section 15.

13. The $20 million

Twenty million dollars is the seed ask. It is not cash in hand. There are no customer logos yet. That is what this raise is for: stand up the factory that can prove the memory, install it, and convert the first accounts.

Eighty-five percent of the ask is product, deploy, evidence, and credits. The $7 million line is the factory. The $3 million air-gap line is Trinity’s own disconnected install — not a bid on the $58.5 billion policy request. The $3 million enterprise-engineering line is Fortune 100 hygiene. The $2 million validation line pays an independent party to re-run the recall clocks. The $2 million in pilot credits sits on the customer side of the proof so the first land is not a science favor.

Sales is $1.0 million — two account executives and legal. The remaining $0.5 million is thin general and administrative reserve. If the first proofs convert, the rest of Year-1 selling cost is funded by recognized revenue and deferred bookings.

Use of funds is a stock. The Year-1 plan is a period. On the base plan the company consumes $8 million of cash at 42 heads. On the downside plan it consumes $14 million. Unused pilot credit, if the gate misses, stays as runway. It does not become outdoor media. Public brand is Series A work after the contract gate.

14. Twenty-four month roadmap

Features can slip. The calendar cannot invent logos. The next two years are a private sales clock — four windows, one gate, and a fork that either opens Series A or revises the plan.

Months 0–6 — the factory

The proof-of-value factory goes live. First CORE proofs run on a six-week clock: install the archive, run the recall protocol, show fail-closed behavior when the system is not sure. The sales team is hired against the million-dollar line and opens named-account conversations under NDA. Independent validation of the recall clocks starts on the $2 million evidence line — pass or fail, published as measured.

What ships in a CORE proof is the memory, the local runtime, and a controllable-egress install the buyer can own. This window is a repeatable proof, not a public reveal. The investor question at month six: can we run the proof, and will an independent party say the clocks are real?

Months 6–12 — first installs

First CORE proofs convert to production installs — or they do not, and the conversion analog gets a real number. DIVISION motion starts: isolation across rooms of the archive, a software air-gap option, a ten-week proof for a line of business. Land-and-expand engineering is the work — single sign-on, SIEM, restore.

Headcount on the base plan builds toward 42. Recognition remains partial-year. The mix table is still a plan, not a CRM: six commercial and two government cells totaling $96 million booked if the base path holds. The honest read of this window is whether an install exists on customer hardware.

Months 12–18 — ENTERPRISE path and the gate

The ENTERPRISE and air-gap path comes up: sixteen-week proofs, disconnected runbooks, the $25 million flagship SKU as a live offer. Air-gap becomes the default path for the buyer who cannot send tokens off-premise. Government work, if it appears, arrives as a real vehicle and a task order.

The gate that owns the next raise is three signed logos or $15 million booked. Waitlists, press, and interest do not substitute. If that gate is hit, the company is a first-customer business with installs on customer hardware. If it is not, the plan revises before more capital is asked for at Series A scale. Pilot credits are the first line we stop. The factory stays; the public campaign does not start.

Months 18–24 — Series A or a bridge

If the gate is hit, this window is Series A preparation and the first public brand — the wider market, funded as a later raise, not as leftover seed. The product family that inherits the archive can surface once the memory is in production. Those doors are not Year-1 SKUs.

If the gate is missed, we raise a bridge against a revised sales plan. Unused seed does not become a street campaign. A miss is a smaller book and a rewritten account list, not a pivot into awareness.

15. What Series A is for

Series A is a $250 million raise in a $200–300 million size band. That band is the size of the next raise. It is not a pre-money number on this seed.

The trigger is either:

• $80 million of recognized run-rate, or
• 15 logos plus one government task order of $15 million or more.

That capital is for scale after proof. Seed buys the memory, the factory, and the first logos. Series A buys the company that memory becomes: land-and-expand on the first installs, the ENTERPRISE factory at volume, classified-capable deploy as a practiced motion, and the public brand this seed does not start. The $4–12 million commercial band and the $15–40 million sovereign band are the pricing story of that later company — expanded annual value as logos accumulate, not two outsized flags planted in Year 1.

The two-pronged trigger is deliberate. A high recognized run-rate means the mix table became a book. Fifteen logos plus a $15 million government task order means the commercial motion and the sovereign motion both exist. Either prong is a company that can take a quarter-billion raise. Neither prong is a date on the calendar. Public brand belongs after the gate, with capital sized for a market that already has installs.

If the trigger does not trip, the next conversation is a bridge and a rewritten sales plan. The $20 million ask is sized so the base path consumes $8 million and the downside consumes $14 million — room to miss without pretending the miss is a media pivot.

They forget, and they bill you for the privilege. We remember on your hardware. That memory becomes every product we ship.

Intelligence owned by the organization that uses it.