01
Dropped in as received.
Exports keep their vendor banners, their merged headers, their archives and their quirks. Reformatting by hand is where errors enter, so the server reads the file the source system actually emitted.
An answer that resembles the question is not the same thing as an answer that adds up.
A total needs every row. So the index never does the arithmetic.
Anything that has to add up is computed deterministically over complete data, and anything that reaches an outside party comes only from a fact a named person signed off. This is the half of the product that exists because a language model is confident whether or not it is right.
Retrieval returns passages. Nothing more.
TWO PATHS · ONE IS NOT ARITHMETIC
Retrieval-augmented chat works by finding the pieces of text most like your question and handing them to a model. That is a good way to find what a document says. It is a terrible way to work out a total, because the index has no idea whether it returned all of the rows, most of them, or the five that happened to sit near your phrasing.
So the gateway keeps two paths apart on purpose, and the analysis path never touches the index.
Cited passage · from the index
Computed figure · from tables
Estimated figure · never
org{•}tech / gateway / two-pathsFind and cite · compute and prove
The index path Good at finding
QuestionIn plain language
Nearest passagesSimilar, not complete
Cited extractNamed file, named page
A total, from passagesNot a path this product offers
The table path Good at adding up
Raw exportAs received
Canonical tableTyped and validated
Integrity checksFirst, or nothing runs
SQL, every rowQuery returned with the answer
The greyed node is the thing everyone assumes a document assistant is doing when it gives them a number. It is not a route through this server.
What "computed" actually means here.
DETERMINISTIC ANALYSIS · SIX HABITS
Nobody opens a spreadsheet, and nobody retypes a figure into a prompt. Raw exports go in exactly as the source system produced them, and the server does the unglamorous work of turning them into something that can be queried honestly.
02
Profiled before trusted.
The server looks inside archives without unpacking them, finds the real header row underneath the banner, and works out what each column is before anything is loaded.
03
Canonical tables, typed.
One shape per kind of record, with types enforced. A value that cannot be parsed is a failure to report, not a zero to quietly substitute.
04
Integrity checks run first.
Row counts, date coverage, duplicate keys, orphaned references, totals against control figures. If the checks fail, the questions do not get answered. That ordering is the whole point.
05
A governed set of questions.
Not free-form querying. A defined set of questions your organization has agreed on, each answered in SQL over every row of the canonical table, so the same question gives the same answer to everyone who asks it.
06
Each answer says how.
The figure comes back with the query that produced it and enough detail to reproduce it. A number you cannot trace is a number you cannot defend in front of an auditor, a lender or a board.
And what it will not do
The server lists the questions it cannot answer from a given source, with the reason. If a source holds no supplier cost, gross margin is not derivable from it, and the honest output is that sentence rather than a plausible percentage. Absence is reported, never filled.
One question, from raw export to a figure that holds.
WALKTHROUGH · BILLING EXPORTS
A generic operations example, following the sequence the one real build actually runs. An analyst drops a set of raw billing exports into the right folder, exactly as they came out of the billing system. On the next scheduled run the mirror has them. Then, in the assistant they already use, they ask what was billed last year and who the largest customers were.
analysis · billing exports · record R-4417
DETERMINISTIC
profile folder "/exports/billing/R-4417" --as-received
[profile] 9 files · 2 archives read in place · header row found beneath vendor banner
Canonical tables built✓ invoices, credits, customers
integrity --all
Row count against control total✓ matches
Date coverage✓ 12 of 12 periods
Duplicate invoice keys✓ none
questions --available
Billed revenue by customer✓ answerable
Customer concentration✓ answerable
Gross margin✕ not derivable: no supplier cost in this source
answer revenue_by_customer --year 2025
returned: 41,208 rows read · query attached · reproducible
workbook standard-outward --publish
Reconcile against canonical totals✓ exact
Tabs with complete data✓ written
Tabs without data△ headed templates, owner named on each
Counterparty namescoded · key returned as a separate link
no spreadsheet opened · no figure from a model
Illustrative output1 refusal · 1 workbook
Had reconciliation failed·the workbook step would have written nothing at all
Three things in that sequence are worth naming. The refusal came before the answers, because knowing what a source cannot tell you is more useful than a number you will have to withdraw. The workbook reconciled a second time at publish, against the same canonical totals, and it writes nothing at all when that fails. And the gaps were emitted as headed templates naming who must supply each one, so the missing half of the work is a list somebody can act on rather than a silence.
Reconcile before publish
When reconciliation fails, the workbook writes nothing at all.
Not a warning banner. Not a partial file. No override flag, for anybody, including us. A document that goes to an outside party is either reconciled or it does not exist.
- 01Twice, not once. Reconciliation runs when the tables are built and again at the moment of publication, because the underlying export can have changed in between.
- 02No privileged path. There is no administrator switch and no forced write. The absence of an override is the control; a control with an exception is a suggestion.
Three states, and only one of them goes outside.
CONFIRMED · CANDIDATE · MISSING
Beside the documents and the tables sits a third store: the facts your organization has actually agreed on. It exists because the alternative is re-deriving the same value on every request and hoping everyone derives it the same way. Each fact carries a state, a source and, once confirmed, the name of the person who confirmed it.
| State | What it means | May be used for internal work | May reach an outside party |
|---|---|---|---|
| Confirmed | A named person signed it off against their verified identity, with the source cell recorded. | ✓Yes | ✓Yes |
| Candidate | Proposed from a document or an export. Explicitly not authoritative. | △As a proposal | ✕No, enforced in code |
| Missing | Known to be needed, not yet supplied. Appears as a gap with the action that closes it. | ✓As a to-do | ✕Emitted as a named gap |
Outward-facing scaffolds accept confirmed facts only. That is a code path, not a policy document, and not an instruction to the assistant.
A long gap list is the system working.The line we repeat on workshop day
The first time a readiness check runs against a real record, it usually returns a great deal of red. That is the product telling you what you did not know you did not have, and the same computation doubles as the to-do list and as an ask list, written as rows a person can paste into an email with the owner named against each item. Nobody enjoys the first run. Everybody prefers it to finding out later.
Where this claim stops.
THE BOUNDARY · STATED WITH THE CLAIM
Two limits that belong next to the claim
Deterministic analysis is proven for one source export format. In the one real build its output was reconciled exactly against the client's own workbook, and each new source format is adapter work with its own profiling and its own integrity checks, scoped before it is promised. And the prose is not ours. The server returns source-linked evidence: cited passages, computed figures with their queries, scaffolds with filled slots and flagged gaps. Your assistant writes the sentences around them. So we say source-linked evidence and never source-linked answers, and we offer no faithfulness or groundedness guarantee, because the server never sees what the assistant finally wrote. The full list is on what the gateway does not claim.
01Can we ask it anything, or only the governed questions?
Search and document reading are open: ask anything, and you get cited passages. Computation is deliberately not open. A governed set of questions is defined with you, each one implemented, checked and reproducible, so that two people asking the same thing get the same figure. Adding a question is a small piece of work, not a prompt.
02What stops the assistant from doing the arithmetic itself?
Nothing in the assistant, which is exactly why the server hands back the query and the row count alongside the figure. If a number appears without a citation or without a query attached, treat it as the assistant's own and check it. We tell clients that plainly on workshop day: a number with no source is a bug worth reporting.
03Who signs off a confirmed fact?
A person your policy allows, acting under their corporate sign-in, with the decision recorded against their verified identity and the source cell it came from. What the confirmation mechanism does and does not yet cover is set out on the limits page rather than in a footnote.
04Does this replace our finance or reporting system?
No, and it should not be sold as one. It computes over exports from the systems you already run, so that an answer can be produced and defended in the place your people are already working. Where a calculation is the zero-error critical path of your business, keep it in the system of record. That principle has its own page: where AI does not belong.
⎯⎯ Book the strategic assessment ⎯⎯
Your files, inside your control ·
Bring one real export and one question you have to be able to defend. We will tell you whether it can be computed, what it would take, and where it cannot.