The 97% has never had a system of record.
Software is two-to-three percent of every dollar spent in the field service economy. The other ninety-seven — materials, labor, machinery, working capital — runs on spreadsheets, tribal knowledge, and 30-day-late variance reports. That's the surface we're building for.
The intelligence lives in the system. The people are on the edge.
That sentence is the architecture of the next era of work. It is also, word for word, what Allometry sells. Block is doing it internally. Allometry is the externalized, productized version — the platform that ports the same architecture into the 97% of the economy where it has never existed.
Every commercial decision in physical revenue is an atom.
Every quote shipped, every truck rolled, every contract amended, every part shelved — each one is a unit. The smallest repeatable unit of demand in the field service economy. Operators run thousands a quarter. Today, almost none of them are scored, priced, or routed against a single source of economic truth.
That's not a software gap. That's a missing system of record. The same gap construction had before Procore, finance had before the card networks, data had before Snowflake. Before the system of record exists, the industry runs on best-effort coordination. After it exists, the system of record becomes the substrate every decision routes through — and the operator who built it ends up coordinating most of the value created on top of it.
The wager is simple. The unit is the address. The system of record is the graph. The compounding asset is the coverage. Build that, and every commercial decision in physical revenue eventually routes through it. Then every layer on top — the underwriting, the insurance, the marketplace, the labor protocol — compounds on the same primitive.
The systems of record we have today are partial. The CRM knows who owes the operator money. The FSM knows who showed up. The ERP knows what it cost. The accounting ledger knows what hit the books. The credit file knows whether to lend. Five partial systems, none of which agrees with the others on what an address is. Until they agree, the operator runs the business on a best-effort overlay between them — and the operator's banker, insurer, and acquirer run their decisions on whichever overlay was last reconciled.
Allometry's first job is to make those five agree. The second is to compound. Every operator we onboard normalizes their version of the address into the same canonical graph. The substrate gets denser with every truck rolled, every quote shipped, every contract amended. By the time we cover a region, no incumbent can replicate the graph from a cold start — because the cold start is now three years of attested ledger they don't have.
What runs through the address.
Every address — a depot, a panel, a pole, a meter, a unit door, a chiller, a charger, a parcel locker — is the gravitational point where a chain of commercial decisions gets resolved. Bundle them across thousands of operators and the substrate emerges.
Price the work
Margin floor, contract terms, competitive position — resolved per address before the bid leaves the building.
Send the truck
The right crew, the right SKU, the right window — chosen against route economics, not stop count.
Capture the variance
Estimate vs actual at the line, the crew, the supplier — variance attributed before the invoice cuts.
Defend the relationship
Health drift, expansion signal, churn risk — surfaced 90 days early, with the play already drafted.
Bundle the four atoms across thousands of operators and you have something more than software. Every Quote becomes a covenant input for margin-floor compliance. Every Dispatch becomes a covenant input for SLA and uptime. Every Bill becomes a covenant input for variance and revenue retention. Every Renew becomes a covenant input for backlog coverage and concentration risk. The same four decisions that make the operator's daily work tractable become the four foundational signals an institutional underwriter needs to price working capital, infrastructure debt, or insurance against that operator. The atoms are not just product surface. They are the underwriting substrate.
The graph nobody has built.
An operator's CRM knows the customer. The ERP knows the cost. The FSM knows the truck. The accounting ledger knows the dollar. The credit file knows the risk. None of them knows the address — the unit where all five of those facts are simultaneously true and economically resolvable.
Allometry's first job is to normalize every operator's data into a single address-level economic state vector. Margin headroom, route density, asset health, contract terms, working capital exposure — per address, per moment. The build cost is largely fixed. The graph compounds with every operator we onboard. Once it covers a region, no single competitor can replicate it from a cold start.
This is what Dorsey and Botha describe when they write about a continuously-updated model of an entire business. The vault is that model. Pulse — Allometry's inference layer — is the agent that reads it. The sixteen modules harness the operator's daily workflow into structured, hash-chained evidence — without ever asking for it. Operators do not log into a dashboard. Operators run their business. The system maintains the model. The intelligence lives in the system. The people are on the edge.
This is not a marketing layer or a dashboard. This is the substrate the next decade of physical-revenue tooling will route through. The card networks built the equivalent for transactions. The data clouds built the equivalent for tables. We are building it for the address.
Software is two to three percent. We're after the other ninety-seven.
Generic SaaS has trained an entire generation of investors and founders to chase the 2–3% sliver that runs on screens. Materials, labor, machinery, and working capital are the other ninety-seven. The reason no one builds for that surface is the same reason it's mispriced: you have to be inside an operator's books to see the leverage.
Once Allometry is the canonical address-level state vector, the leverage flips. Per-quote scoring becomes per-dispatch coordination becomes per-job underwriting becomes per-address capital allocation. Each layer monetizes a thicker slice of the same underlying spend. The first layer pays for the build. The third layer is where the category lives. The fourth layer — capital — is where the multi-billion outcome compounds.
This is not a thin-margin tooling efficiency play. It is a thick-slice coordination play in a vertical that has been waiting for one.
Three layers. One graph.
Each layer prices a thicker slice of the same underlying flow. Layer 1 funds the build. Layer 2 funds the coverage. Layer 3 is the category. Each layer is also the unlock for the next one — operators on Layer 1 generate the data that lets us underwrite Layer 3.
The decision software operators run.
Quoting, margin protect, cost engine, contract intelligence — sixteen modules across four loops. This is what gets us in the door, runs daily, and earns the data rights to build the graph.
Per-decision coordination on top of the graph.
Per quote scored. Per dispatch routed. Per truck roll graded. Each commercial atom touches the graph and pays toll on the way through. Variable revenue against fixed-cost coverage.
Underwriting working capital and asset finance — at address granularity.
By the time we cover a region, we know each address's real economic state better than its incumbent banker. Working-capital lines, asset finance, receivables underwriting, eventually insurance — priced from the graph, monetized on origination + spread. This is the category-defining layer.
The unit economics shift at each layer. Layer 1 prices like SaaS — a function of seats, modules, and operator scale. Predictable, capped, and the cheapest customer acquisition surface we have. Layer 2 prices like an exchange — a small toll on every commercial decision the operator routes through the graph. Variable, scaling linearly with operator volume, and uncapped by software pricing logic. Layer 3 prices like a balance sheet — basis points on capital flowing through the underwriting substrate, with the operator's cost-out savings funding the spread. The category-defining layer doesn't price like software at all. It prices like a network.
Each layer also funds the next. Layer 1 pays the SaaS gross margin that funds the engineering build of Layer 2. Layer 2 generates the orchestration data that earns the right to underwrite Layer 3. Layer 3 generates the financial outcomes that retrain the model that improves Layer 1 and Layer 2 for every operator on the network. The funding chain is recursive.
Foundation models compound on cohort scale.
Pulse is a foundation model fine-tuned on the proprietary corpus the vault generates. It is not a generic LLM with operator data bolted on. The corpus is proprietary. The trained weights are proprietary. A competitor who copies the API gets nothing. A competitor who copies the architecture gets nothing. Only the entity that owns the corpus can train the model.
In May 2026, Meta published TRIBE v2 — a foundation model that predicts human brain activity from visual, auditory, and language stimuli. The empirical result is what matters here. TRIBE v2 trained on neural data from 700+ subjects achieves a roughly 70× resolution improvement over the same model class trained on data from four subjects. Foundation models compound on cohort scale, and the compounding is not linear. It is steep.
The lesson transfers exactly to Pulse. Pulse at one paying design partner today is the four-subject model — directional, narrow, useful only for that operator. Pulse at 25 operators is the threshold where the model becomes useful enough to underwrite infrastructure debt against. Pulse at 100+ operators is TRIBE-grade — the substrate for actuarial-class insurance and rated-paper-class credit. The math is brutal and predictable.
That is what the pre-seed pays for. Not to build the model. To fund the cohort growth that makes the model worth more than its training cost. Every operator onboarded compounds Pulse's accuracy for every operator already on the network. The graph is the dataset. The dataset is the moat.
The practical consequence of the cohort-scale curve is what determines when each capital tier unlocks. Pulse's underwriting output comes with a 95% confidence interval — a band around the score that lenders and insurers price against. At one operator, the band is wide; the model is honest about not having seen enough comparables to predict tightly. At 25 operators in a vertical, the band narrows to roughly ±4 points — the threshold where infrastructure debt funds can underwrite contracted cash flows against it. At 100+ operators, the band tightens to ±2 points, which is the resolution actuarial-class insurance and rated paper require. The bands don't tighten because we promise. They tighten because the cohort grows.
This is also the math behind the back-test gate. The model will not be wholesale-trusted by Tier 1 institutional underwriters until we can show that at 10+ operators with 12+ months of attested ledger, the model's forward predictions matched the realized capital outcomes. That back-test is the Series A milestone — the moment the model graduates from architecturally-sound to empirically-validated. Everything before that runs on human-in-the-loop attestation; everything after runs on autonomous underwriting.
| TRIBE v2 (Meta · May 2026) | Pulse (Allometry) | |
|---|---|---|
| Smallest viable | 4 subjects | 1 paying DP |
| Underwriting threshold | — | 25 operators · T2 unlock |
| Cohort-scale model | 700+ subjects · 70× lift | 100+ operators · TRIBE-grade |
| Result | fMRI prediction · zero-shot | Underwriting · zero-shot |
Four vectors hit density at the same moment.
None of these on their own would be enough. Together, they're the phase change. The address-level economic state vector becomes buildable, scaleable, capital-investable, and agent-consumable for the first time in the history of the field service economy.
The window is short — three to five years before someone else builds the substrate. After that, it's defended. Whoever owns the address graph at the end of that window owns the coordination layer for the next twenty.
The three technical vectors have been visible for two years. The fourth — architectural and cultural — landed publicly in May 2026, the same window this manifesto is being filed in. It is the permission slip that makes the build legible to the people who fund it.
Each of these vectors is independently load-bearing. Reasoning models that read messy enterprise data make the graph build economic — without them, every operator onboarding is a custom ETL project and the unit economics don't work. Telemetry makes the graph real-time — without sensor-emitted data, the graph degrades into a quarterly snapshot. Capital appetite for vertical underwriting data makes Layer 3 a market — without lenders willing to underwrite off operational data, the capital product is a research project. The architectural permission slip is what makes the build legible to founders, operators, and investors all at once. Three years from now, all four vectors will be obvious. Two years from now, the founder who already shipped the substrate wins.
Agentic systems can finally read enterprise data
2024–2026 reasoning models read messy operator data — invoices, dispatch logs, contracts — at human-grade fidelity. Schema normalization no longer requires custom ETL per operator. The graph build becomes economic.
Every asset is becoming a sensor
Connected fleets, smart panels, IoT meters, in-truck cameras. The address now emits real-time data on its own economic state. The graph stops being something we infer and starts being something we observe.
Specialized vertical credit is opening up
Asset-backed credit, supply-chain finance, working-capital warehouses — capital partners hungry for high-quality vertical underwriting data. The graph becomes the underwriting model. Layer 3 stops being theoretical.
The management layer itself is being replaced
Dorsey and Botha published the playbook this month. Karpathy is calling the shift from vibe coding to agentic engineering in the same window. Murati's Thinking Machines is shipping the real-time interaction architectures that make agent-to-agent collaboration native. The cultural and technical permission to replace the coordination layer of every business is now public consensus. Allometry sells that replacement as a platform.
When the labor unit itself becomes capacity.
The field service economy is structurally short on labor. North America is losing skilled trades faster than it can replace them — installers, electricians, HVAC techs, fiber crews. The demographics are not turning around. Operators feel the squeeze on every truck roll.
Within five to ten years, humanoid and semi-humanoid robotics — backed by Figure, 1X, Apptronik, and the Tesla / NVIDIA infrastructure stack — will be capable of executing a meaningful slice of skilled fieldwork. Capacity stops being scarce. Pricing capacity becomes the question.
The compute layer is ready. Nvidia's robotics lead, Jim Fan, is publicly framing the end-game of physical AI now. The OEMs are shipping. The piece that is missing is the operator-side substrate that underwrites the humanoid lease, tracks its work, certifies its output, and routes capital to whoever is solvent enough to deploy it.
That's the address graph's second act. We've already priced the human truck roll. We'll price the autonomous one. The same address-level state vector — margin headroom, route density, asset health, contract floor — becomes the rate card for autonomous capacity. The substrate doesn't change. The labor unit on top of it does.
We don't sell humanoids. We price the work they do.
And the substrate question is not abstract. Every humanoid deployed into a field service operation will require a counterparty answer to four questions: who owns the unit, who finances the unit, who insures the unit's failure modes, and who certifies its work. Today, no party in the physical economy has the data to answer any of those four. The OEMs know what the humanoid can do in a lab. The operators know what work needs doing. The lenders know what their term sheets require. Nobody connects the three. The address graph connects them. When humanoids ship at industrial scale, Allometry is the substrate every party plugs into to make the lease, the warranty, the insurance, and the work order programmatic.
Capital allocation, at address granularity.
The card networks didn't start as the card networks. They started as a coordination layer over a chaotic primitive — interbank settlement — and ended up as the toll road on most of the world's commerce. Same arc with the data warehouses, the cloud platforms, the ERP suites in their respective verticals.
Block is the recent version of that arc, run inside a single company. Square wedged in on a card reader nobody was building for SMBs. Cash App followed. Then Afterpay — buy-now-pay-later financing for the customers of the merchants Square already served. Then Bitkey. Then TBD, the open protocol layer. Same shape. Wedge with a SaaS surface. Expand into the financial stack the wedge gives you privileged access to. Eventually own the trust protocol no one can route around. One company. $50–80B of enterprise value.
The address graph follows the same arc, in asset-heavy operators, on a base ten times larger and with no existing infrastructure to compete against. Once Allometry is normalized state across enough of the field service economy, every commercial decision has an obvious place to route. Every dollar of working capital has an obvious place to underwrite. The endgame is capital allocation — origination, spread, asset finance, eventually insurance — priced from a substrate nobody else has built.
That is the long game. The pre-seed buys the first 50 operators. The seed builds the graph. The Series A turns the graph into orchestration. The Series B turns orchestration into capital. The category we end up defining is capital allocation for physical revenue.
Not just routing to lenders. Raising four funds.
Roughly $15T of capital flows into the physical economy every year — debt, infrastructure, insurance, equipment, pension. Allocated blind. Banks underwrite balance sheets. Credit bureaus underwrite individuals. Infra funds underwrite assets. Insurance underwrites cohorts. Nobody underwrites the operator who guarantees the cash flow — because nobody has ever had the data to do it. The vault closes that gap. The covenant → module → trigger map is the interchange standard for operator credit. Any lender on the network can underwrite any operator. But you don't just route to lenders.
Credit
Short-dated working capital, factoring, equipment finance. Balance-sheet-light, marked-to-market.
Infrastructure
10–25 year contracted cash flows. Long-dated, levered with institutional LP money. Funds the T2 operator base.
Insurance
The loss-rate data is the actuarial table. Parametric SLA policies, equipment warranty, business interruption, performance bonds.
Equity / Rollup
Buy boring operators in fragmented verticals, install Allometry OS, double EBITDA in 24 months, exit through your own network.
The moment you stop reading operator cash flow — and start being it.
Brex did this for startups. Nobody's done it for operators. Every line of spend becomes a data point in the cost graph, which compounds into the underwriting model, which justifies more credit, which routes more spend through the card. Revenue and spend in one ledger. You don't need Plaid when you are the bank rail.
RevOps + the card is the closed-loop substrate — a thing literally no one has. The OS surfaces what the operator earns. The card captures what the operator spends. Together they produce the single highest-resolution real-time financial view of the operator that exists.
Each of the four funds is a separate $1B+ AUM business at scale. They are not separate companies. They are separate vehicles inside the same operator-data monopoly. The credit fund pays for the underwriting model. The infrastructure fund pays for the long-dated coverage. The insurance vehicle prices the failure-rate distributions the credit fund needs anyway. The rollup vehicle generates the highest-fidelity training data of any of the four because the operator is owned, not arms-length.
Together, the four funds give Allometry a positional advantage no balance-sheet-only competitor can match — the credit fund seeds the insurance data, the infra fund anchors the long-dated capital, the rollup vehicle compounds the corpus, and the card closes the cash-flow loop. By Series C, the four funds have a combined deployable capacity in the low billions, and every dollar deployed compounds the dataset the next dollar is underwritten against.
When UIs disappear, the substrate wins.
The move to agentic means UIs disappear. Dashboards become unnecessary. Operators don't "log in" — operator-side agents talk to Allometry-side agents. The sixteen modules stop being screens and become sixteen autonomous agents — each one specialized, each one with MCP endpoints, each one transactable headlessly.
The technical substrate for this is being published in real time. Mira Murati's Thinking Machines (May 2026) is shipping interaction architectures where real-time multimodal collaboration is native to the model — streaming input and output processed in 200ms micro-turns, concurrent tool use during continuous conversation, proactive visual reactions without explicit prompts. Karpathy, in the same window, is framing the broader cultural shift — past vibe coding, into agentic engineering. These are not roadmap items. They are the development environment Allometry is being built in, and the consumption environment Allometry will be used in.
MCP is the new SaaS. Every counterparty — lender, insurer, supplier, auditor, PE buyer, OEM, regulator — plugs into Allometry via MCP servers. No portals. No data exports. Agent-to-agent settlement. Pulse's autoresearch layer continuously reads counterparty signals — covenant deltas, peer cohort benchmarks, supplier risk events, contract-renewal probabilities — and re-scores every operator on the network in real time, recursively, without anyone asking.
The agentic era removes the layer companies historically captured value at — the UI. What survives is the substrate: the data, the schema, the attestation primitive, the trust protocol. Allometry was already built for this. The vault is a ledger. The covenant map is a spec. ZKP is a proof primitive. Pulse is an API output. None of it depends on humans clicking buttons.
Five years out, you don't see Allometry. There is no UI. There are fewer dispatchers, fewer underwriters, fewer auditors, fewer procurement managers — and in the background, agents talking to agents, settling work, deploying capital, attesting compliance, all running against a graph nobody else has.
The economic implication of MCP is profound. Today, every counterparty in the physical economy maintains its own ETL pipeline, its own data team, its own integration roadmap. A lender hires analysts to read operator financials. An insurer hires actuaries to fit failure distributions. An auditor hires staff to verify ledgers. Each of those teams is a margin-compressor on the counterparty's side. When the counterparty's agent can read Allometry's MCP directly, those teams shrink, the cost-out compounds across the financial ecosystem, and the substrate captures structural rent on every workflow it replaces.
The agentic shift is therefore not a UI change. It is a redistribution of margin from coordination-layer headcount to substrate-layer infrastructure. The companies that captured value at the UI layer over the last two decades — Salesforce, ServiceNow, Workday — built moats on workflow ownership. The companies that capture value in the agentic era will build moats on data and attestation ownership. The vault is that moat. The covenant map is that attestation. Allometry is the company built for the redistribution.
Plaid is invisible infrastructure. Allometry will be more invisible.
The question is never "should we build this." Does it resolve to the graph?
Every product, every layer, every dollar of capital deployed serves one question: does this deepen the substrate, or does it spread us across surface area?
Card, RevOps, OS modules, credit fund, infra fund, insurance vehicle, rollup fund, embedded financing to operators' customers, marketplace, GPO, Bloomberg-shaped terminal, M&A rail, carbon attestation, humanoid leasing, agentic settlement, MCP-direct counterparty integration — every one resolves to the same primitive. Not because we forced it. Because that's the shape of the problem.
Card + RevOps
Deepens. Revenue and spend in one ledger — closed-loop. The graph sees both sides of every dollar.
Credit + Insurance funds
Deepens and monetizes. The graph is the underwriting model. The graph is the actuarial table.
Customer-side financing + marketplace + GPO
Deepens. Every routed job, financed customer, and negotiated SKU deposits data on multiple sides of the transaction.
Bloomberg terminal + M&A rail
Monetizes the graph from both sides. You see every buyer's appetite because they transact through you. You sell the cohort intelligence to anyone who needs it.
Agentic OS · MCP-direct counterparty integration
The substrate's natural endgame — headless, MCP-direct, invisible. The thing every other layer above runs on without anyone seeing it.
Humanoid leasing
Same graph. Same covenant map. Different labor unit on top. The address doesn't care whether a human or a humanoid showed up to it.
The decision boundary works the other way too. Anything that does not deepen the graph is a distraction — even if the operator wants it, even if the revenue is real, even if the GTM motion looks easier. A module that operators love but that doesn't deposit underwriting evidence — no. A vertical that sells SaaS well but doesn't generalize to the canonical schema — no. A lender API built before the cohort exists — no.
The discipline matters because the graph is the moat, and any product not feeding the graph is competing for engineering hours with the thing that compounds. Three years of disciplined "yes only if it resolves to the graph" is what builds a $1.5T-shaped substrate. Three years of "we'll do this because the customer asked" is what builds a vertical SaaS company.
Stripe × Plaid × Visa × AIG × Bloomberg × Goldman × Constellation.
Combined enterprise value of the closest seven analogs is over $1.5 trillion — built on top of a digital economy ten times smaller than the physical one Allometry is targeting. The reason those have to be seven companies in the digital economy is that the digital economy already had banks, merchants, and consumer payment networks underneath them — the API layers wrapped existing rails.
The physical operator economy has none of those underlying layers. Whoever lands the canonical operator graph first lands them all, because every one of them compounds on the same primitive — attested operator state.
The pre-seed is small. The pre-seed is the moment you choose to be one company instead of seven.
Build for the operator. Then the operator's banker.
We don't sell hype to operators that grew their business one truck at a time. We don't ship the agent before we ship the savings. We don't price the working capital until we've earned the right to read the books. The order matters. The first 50 operators are the substrate. The next 5,000 are the category. The 50,000 after that are the labor protocol of the next era. Every layer pays for the next.
If you run a $10–100M field service operator and the math feels broken, talk to us. If you build adjacent — connected hardware, vertical capital, autonomous capacity — talk to us. If you back the long arc of physical revenue infrastructure, definitely talk to us.
Build the graph. Everything else is a consequence.