Built backwards, from a problem.
Nobody set out to build an identity system. We set out to generate a book that held together, and everything here is what fell out of failing at that for long enough to understand why.
Stories have measurable shape.
Propp catalogued the moves of the folktale in 1928. Vonnegut drew story arcs on a blackboard and had the thesis turned down. Decades later a team recovered six recurring arcs from roughly thirteen hundred novels using singular value decomposition, and a 2019 paper pulled storyline structure out of document space with robust PCA on Grassmann manifolds.
A century of people insisting structure is there and measurable, and no settled name for the practice. We took one.
Four things the system assumes.
Every design decision traces back to one of these. When they conflict, the one earlier in the list wins.
Structure is measurable
Propp catalogued the moves of the folktale in 1928. A century of people insisting narrative has shape you can measure — and no settled name for the practice. We took one.
Identity is the hard part
Generating a paragraph is solved. Generating the ten-thousandth paragraph in the voice of someone established in the first is not. Everything here follows from taking that seriously.
Keep the world outside the window
Every model degrades as context grows. So the world doesn't live in the context window. Characters, physics, and voice live in a structured record each pass writes back to.
The work is the point
Not a longer conversation. A manuscript, a score, a cover, a session ledger — artefacts that exist after the session ends and belong to whoever made them.
We call them PACs, not agents.
Personalized Augmented Collaborators. An agent executes a task. A collaborator has a psychology, a voice, and a history — and those are exactly the properties that make long-form work hold together.
Personalized Augmented Collaborator
A collaborator with a psychological profile, a voice, and a history — not a function call with a system prompt. We've used the term since before agents were the default word.
Persona
A portable writing identity: fifty-one attributes across core voice, narrative craft, worldview, psychology, shadow self, and relational behaviour. Built once, reused across every project.
Character
Narrative-bound. A character exists inside a story, latched onto a persona foundation. The persona supplies how they think; the character supplies who they are in this book.
Ontology
The structured store of everything established: traits, physical detail, relationships, world state, speech habits. Each generation pass reads from it and writes back to it.
Narrative force
What acts on a character. Identity here is not a fixed sheet — it drifts under pressure, and the trajectory carries more information than any single snapshot of the current state.
Token
The platform unit of account for work performed. Distinct from a model's internal token count — two providers with very different pricing can sit behind the same action.
Two years of the same problem.
Not a roadmap. A record of what broke, in the order it broke, and what had to be built to get past it.
One provider, one problem
Started as an attempt to generate a readable full-length book against a single model API. It fell apart around chapter three, the way everything does.
The ontology
Fixing continuity meant moving character state out of the prompt and into a database. That decision produced the attribute schema, and the schema turned out to be the actual product.
PACs and persona teams
Single-voice generation gave way to teams: outliner, prose stylist, continuity checker, each a persona with its own psychology, arguing a beat out before it reaches the page.
Provisional filing
A provisional patent application was filed with the USPTO covering the computational identity architecture — priority date established while the work continued.
Narrative Geometry
The platform was rebuilt and renamed around what it had actually become. Multi-provider routing, native audio, and a private beta.
A small operation in Ottawa.
Narrative Geometry is built and operated by Alalonde Corporation, based in Ottawa, Ontario. A deliberately small team working on a problem larger companies have mostly routed around, because the money in generative AI is currently in short-form output rather than in whether chapter forty remembers chapter three.
We route across several providers rather than committing to one. Model quality moves monthly; an abstraction layer that picks the best model for each task — and swaps it out when something better ships — is worth more than any single model relationship.
Built in Canada, on purpose
The SR&ED framework refunds a large share of core R&D in cash rather than in credits against tax not yet owed. For a company at this stage, that is runway a competitor in another jurisdiction simply does not have.
Cash refund from the CRA on the first $6M of eligible R&D expenditure, not a credit against future tax owing.
Ontario and British Columbia both add refundable credits on top of the federal position.
The 2026 rules bring hardware acquired for R&D — GPU capacity included — back inside the claim.
Persona Protocol.
The identity layer underneath Narrative Geometry is more general than fiction. A persona — a portable identity with a psychological profile, a voice, and defined behaviour under pressure — is useful anywhere a synthetic entity has to stay itself across time.
That layer is being developed separately as Persona Protocol Narrative Geometry is its first and most demanding consumer: if an identity holds across 150,000 words of fiction, it will hold across most things.
Like what's going on here?
Whether you're an author who wants in, or someone curious about the company itself — investors, partners, press — we'd like to hear from you.