// qmetrika_labs · company_paradigm · thesis_document
A company model where the unit of capital is neither the labor hour nor the physical asset, but the token: the inference compute that, governed and captured, accumulates as proprietary, defensible, and compounding knowledge. Applied to a deep vertical — bioinformatics and AI — where specialization turns today's spending into tomorrow's moat.
// 01_thesis
The marginal cost of producing a text, an analysis, or a decision has dropped to near zero. When cognitive work becomes cheap and abundant, the scarce asset is no longer the work itself — it becomes what that work leaves behind.
For two centuries, a company built capital on two things: labor (person-hours) and physical assets (machines, buildings, inventory). The balance sheet measured the tangible; the income statement measured how much work was needed to produce revenue. Generative AI breaks that framework at its foundation: cognitive work — writing, classifying, reasoning, deciding — is now produced in tokens, at a cost trending toward zero and with a practically infinite supply.
This forces the question of where value lives. If anyone can rent the same foundation model and generate the same tokens, the raw token is not capital: it is a commodity. Capital appears when the token stops being ephemeral — a response that is consumed and forgotten — and becomes something that accumulates: a captured human decision, an evaluation data point, a fine-tuned model that does tomorrow for free what cost dearly today.
This document describes that architecture. Its founding motto — inherited from QMetrika — is literal, not metaphorical: energy that transforms into knowledge. In QMetrika's science, the free energy (ΔG) of an RNA molecule is profiled to extract biological knowledge. In the company, computational energy (tokens) is spent to extract organizational knowledge. It is the same equation on two layers — and that coincidence is not wordplay, it is the axis of the entire model.
// 02_token_classes
The first accounting error of the AI-company is treating all tokens as the same input. There are two classes, with opposite economics — and the company's health is measured by the ratio between them.
From a third-party frontier model (Opus, GPT-4o, Gemini). Powerful but expensive, and you don't own it: you rent it per call. Pure OPEX — each use vanishes from the balance sheet the moment it is consumed.
Its proper role: doing the hard bootstrap work, the reasoning you haven't yet learned to distill. Necessary, but it is cost, not asset.
From a model running on your own hardware (Ollama/MLX, an NPU/iGPU, a Linux node) — at marginal cost ≈ zero once amortized. It is the CAPEX counterpart: upfront investment that yields nearly free tokens forever.
Its role: absorbing all the routine and specialized work you already know how to do. Every task that migrates from Class A to Class B is margin freed up.
| Dimension | Foundation token (rented) | Local token (owned) |
|---|---|---|
| Marginal cost | High · per million tokens, in € | ≈ 0 after amortizing hardware |
| Accounting nature | OPEX — consumed and gone | Amortized CAPEX — yields in perpetuity |
| Ownership | You don't own it; you depend on a vendor | Yours; runs offline, no quota, no data leakage |
| Optimal use | Hard reasoning, bootstrap, novel cases | Routine, high volume, already-distilled tasks |
| Autonomy | Subject to rate limits, pricing, and vendor policy | Total — more local, more sovereignty |
| Sensitive data | Leaves your boundary (GDPR, IP) | Never leaves your infrastructure |
The Capital Q company does not optimize for "spending less on AI." It optimizes a single ratio: proportion of work served by local tokens versus foundation tokens. The more it shifts toward local, the greater the autonomy, the lower the cost, and — this is the key — the more proprietary knowledge accumulated. The goal is not to turn off the expensive tokens: it is to convert their output into local capability before the spending evaporates.
This distinction is already operational in QMetrika's Agent Board. Its work-in-progress (WIP) limit is not measured in number of agents, but in cost, and multidimensionally: { agents, € total, € on foundation models }. An agent running a local model enters with almost no restriction; one using Opus draws from the foundation budget and, when exhausted, drops to a dedicated lane — ⚡ No budget — where you can see why it was throttled. Cost is not a footnote: it is the physics of the system.
// 03_durable_capital
If the raw token is a commodity, where does capital accumulate? In three deposits that fill as a byproduct of daily work — but only if the company is built to capture them. Without capture, the same work generates zero capital.
Every time a human approves or denies an agent's action at the governance gate, they produce a label: a tagged decision about what was correct. That judgment is AI's most expensive input — and here it is generated for free, as a residue of work.
→ Preferences for fine-tuning. Rules to harden. Expert intuition, made data.
Every tool call and every decision is recorded in a hash-chained ledger: immutable, verifiable, impossible to alter without breaking the chain. It is not just compliance — it is a clean, traceable dataset of everything the organization has done.
→ Foundation for evaluations, reproducibility, regulatory defense, and demonstrable trust.
With accumulated judgment and audit, you fine-tune local models that internalize the specialized work. Each cycle shifts load from Class A to Class B. Knowledge stops living in an external vendor and moves to weights you own.
→ Proprietary capability, composable, that lowers cost and raises autonomy simultaneously.
This is the conceptual inversion that defines the paradigm. In the old AI economy, the asset was the model. In Capital Q, the model is replaceable and the asset is the substrate that trains and evaluates it: the accumulated record of correct human decisions in your specific domain. A better foundation model only makes your capital yield more, because it starts from better raw material. It never decapitalizes you.
// 04_balance_sheet
If this is capital, it needs a ledger. At QMetrika, that ledger is not a separate spreadsheet — it is the same operational board read through a different lens. One system, two views: the Agent Board governs action; the Capital Board reads that same action as accumulating capital.
Live Kanban: each agent is a card advancing through columns. You see what each one touches, its model, its tokens, and its cost in real time. Actions with side effects don't execute on their own: they land in Needs Input and wait for human approval with an operator token.
The decision about what is allowed is code — deterministic allow/ask/deny policy — not model judgment. Control lives in the tool, not in the prompt.
The same substrate — audit, cost ledger, approvals — read as capital that accumulates. Cost by vendor (foundation vs local). Decisions by gate outcome. Token usage local vs foundation. And at the top, capital captured per week: human judgment labels + audit entries for evaluations.
Where the operational view asks "what is happening now?", the accumulation view asks "how much moat have I built this week?".
// 05_accumulation_spiral
The company's engine is a Design–Build–Test–Learn (DBTL) cycle. Each turn converts today's foundation spending into tomorrow's local capability, leaving behind a fuller capital deposit. The spiral gains momentum: each turn costs less and yields more.
A normal services business spends the same on task 1,000 as on task 1. The Capital Q company does not: task 1,000 runs nearly free on local because the previous 999 left the judgment to fine-tune it. Marginal cost decays while the moat grows — the two curves go in opposite directions, and that scissor is the business model.
The Learn leg is not a metaphor: it is a concrete pipeline that converts Deposits 1 and 2 (captured human judgment + verifiable audit) into Deposit 3 (fine-tuned local models). Traceability is not a separate requirement from fine-tuning — it is what makes the fine-tuning dataset trustworthy.
The data emitted by the gate already has the shape of a training example — context → proposed action → verdict — so there is no need to fabricate labels: the operation produces them as residue. That data is exploited in three ways:
The (context → approved action) pairs become the "gold" behavior: the local model learns to propose what the human would have approved. The simplest approach — ideal at startup.
KTO fits almost perfectly: it works with loose binary signal — approve / deny — which is literally what the gate emits. DPO uses approved-vs-denied pairs when you have them.
You aggregate decisions to train a model that scores future actions. It becomes a pre-filter: it raises the bar of what the gate lets through without asking.
The hash chain gives the pipeline four things a normal fine-tuning setup lacks: provenance (proof that every label came from a real human decision), immutability (no one poisons the set silently), lineage (you can reconstruct which decisions trained which model version — what the EU AI Act and SaMD dossiers require), and deduplication by hash (you know which decisions have been "spent" in prior training runs).
As the local model improves, fewer cases reach the gate — and fewer new labels are generated. The cycle can suffocate itself. The cure is active learning: deliberately routing a fraction of cases to human review, to keep the judgment flow alive and detect drift. Capture must never drop to zero entirely.
// 06_vertical_wedge
Capital Q does not compound horizontally. A generic assistant competes with the foundation model on its own turf — and loses. It compounds in a narrow and deep vertical, where the generic token is worth little and specialized capital is worth much.
The choice of vertical at QMetrika is not arbitrary: it is where the metaphor becomes literal. The scientific engine is a stacking free energy (ΔG nearest-neighbor) profile of RNA. The company already lived on transforming energy into knowledge — thermodynamic energy into biological knowledge. Capital Q is that same equation raised one layer.
Science layer: thermodynamic energy (stacking ΔG) is spent and biological knowledge is extracted — variant classification, mRNA optimization. Company layer: computational energy (tokens) is spent and organizational knowledge is extracted — labels, evaluations, proprietary models. The company is fractal with respect to its own science.
And the vertical doesn't just give narrative coherence — it gives compounding defensibility:
P202630522 (OEPM, 2026) covers the mRNA thermodynamic fingerprint method — both variant classification and codon optimization. A foundation token cannot reproduce a patented method. The PCT extension (April 2027) internationalizes the moat.
Real production logs, clinical pilots, experimental validation. Expert judgment on specific variants. This data is not in any model's training corpus — only your operation produces it.
A biophysical profile orthogonal to what competitors see (CADD, REVEL, AlphaMissense operate at the protein level; the ΔG engine sees mRNA thermodynamics). It doesn't compete: it complements.
In a domain where being wrong has clinical cost, being able to prove that you're right — with a verifiable audit chain and a proprietary evaluation bank — is itself the product.
Start narrow and deep. One agent excellent at variant classification for a handful of genes is worth more than a mediocre one that tries all of genomics. Depth is what makes the specialized local token outperform the generic foundation token — and only then do you widen the wedge.
// 07_stack
The paradigm materializes in three pieces that have already been built and published. Each covers a phase of the cycle. Together, they are the Capital Q company.
| Framework | Phase | What it does | Role in capital |
|---|---|---|---|
| Agent Canvas Model | Design | 9 blocks that force the expensive decisions before coding: segment, value proposition, channels, autonomy, KPIs, resources, capabilities, integrations, costs and risks. | Prevents burning tokens on poorly planned agents. |
| Agent Board | Govern | Supervision board + deterministic gate (allow/ask/deny) + WIP measured in cost + hash-chain audit. LLM-agnostic. | Turns every action into a label and an audit entry. |
| Capital Board | Capitalize | Second lens on the same substrate: reads cost, decisions, and audit as accumulated capital. | Makes visible how much Capital Q has been built. |
| ΔG Engine · Bio-AI | Defend | Nearest-neighbor free energy profile for classifying variants and optimizing mRNA. Patent + four publications. | The moat. Where specialized capital defensibly outperforms the generic token. |
// 08_principles
Eight rules that distinguish a Capital Q company from one that merely "uses AI." These are not aspirations — they are architectural decisions that either get made at build time, or don't get made at all.
Treat every foundation spend as an investment that must leave a deposit filled (label, evaluation, model). If a token is consumed without capturing anything, it was pure cost.
The ratio of local to foundation work is the health metric. Each turn of the spiral should shift it toward local. More local = more margin, more autonomy.
The decision about what gets executed is deterministic code at the gate, not LLM judgment. The effect is unreachable without crossing the gate, regardless of which model invokes it.
Every approval is a golden data point. Design so that expert judgment is recorded, labeled, and reusable. The label is AI's most expensive input; here it is a byproduct of work.
Verifiable chain of everything agents do. Not just compliance: it is the evaluation dataset and the demonstrable trust that no generic can rent.
The vertical wedge is where capital compounds. One agent excellent at a specific case beats a mediocre one that tries everything. Widen only when depth yields returns.
Models expire; captured judgment and your domain's evaluations do not. A better foundation model only makes your capital yield more — it never decapitalizes you.
The Canvas before the code. Ambiguity about who the agent is for and what autonomy it has is paid for in burned tokens. Resolve it first, on paper, where it's cheap.
A Capital Q company is a machine for converting compute into proprietary knowledge, governed so that every action leaves capital behind, sharpened on a vertical where that capital becomes defensible.
At QMetrika, that vertical is RNA thermodynamics — but the model is general. Energy that transforms into knowledge, on two layers, compounding turn after turn.
The token is not a cost to minimize. It is the raw material of a new kind of capital — one that compounds when governed, and evaporates when not.