How Analect works
Analect is ETT's platform for running AI as an accountable operating cost. It forecasts what each AI workload should cost and which model should do it, and it certifies AI outputs claim by claim against the documents they came from.
Underneath the two services are five workflows. Two of them are what the services run on; the other three come into play in a pilot. Customers and pilot teams sign in to the ones their engagement uses, and nothing stored in one organisation’s account is visible from another’s.
Five tools over the same material
Analect Forecast
Prices every AI workload across the model catalogue, from your bills or a plain description, with the reasoning behind each pick and the catalogue's date on every figure.
Part of the AI Spend ForecastAnalect Evidence File
Breaks an AI output into its claims and checks each one against its source with two independent models, quoting the passage every verdict rests on.
Part of the Accuracy Evidence FileAnalect Explorer
Decomposes documents into knowledge fragments, each tied to the sentence it came from, so you can see what a document actually asserts.
Access arranged as part of a pilotAnalect Proof
Asks the same questions of a document and its fragments, and reports agreement, correctness and completeness separately, with the tokens each route used.
Access arranged as part of a pilotAnalect Store
Weighs every form your material could be kept in, in actual bytes, and models keeping your sources against replacing them.
Access arranged as part of a pilotA certificate is only worth what somebody else can verify
Every Certificate of Accuracy carries a reference and the SHA-256 digests of the exact source and output it was issued against. Change a word of either and the digest no longer matches.
Anyone holding a certificate, a client, an auditor or an insurer, can check on the public verification page that we issued it, that it has not been withdrawn, and that the file in front of them is the one it covers. Nothing is uploaded to do it: the file is digested in their browser, and the registry it is checked against holds digests and dates, never the documents.
Read a document once, not once per question
A language model has no index into a document. To answer a question it reads everything, so every question pays for every token, and the same document is paid for again on the next question. Analect's knowledge engine reads a document once and decomposes it, one way, into knowledge fragments: small structured statements of what the document asserts, each tied to the sentence it came from. A question can then retrieve the fragments that hold its answer rather than the pages around them.
Status. The engine is in pilot. We will publish its figures when they hold beyond our own sample, with the method beside them, and not before. Until then the way to find out what it does on your material is a pilot that compares it with your current retrieval and reports the questions where it did no better.
What we keep, and what we never do
Your bills stay in your browser
The AI Spend Forecast reads your billing files in your browser and never uploads them. We keep the forecast built from them, and you can leave the cost figures out.
Your documents are not kept
The Accuracy Evidence File never keeps your source documents or the outputs themselves. We keep the certificates: the claims, and the quotes each verdict rests on.
Named models, named terms
Checks run on models from Anthropic and OpenAI, through Vercel's AI Gateway, under their commercial API terms. Our sub-processor list is available on request.
Each step has to earn the next
- Step 1
A fixed-price service
Three weeks on your own bills or outputs. You keep the report and the working file whatever you decide next.
- Step 2
A pilot
Where the service shows something worth going further on, a fixed-fee pilot puts the workflows in front of your team on an agreed slice of your estate.
- Step 3
The platform
Where the pilot earns it, Analect in your environment, with routing, governance and the evidence trail set up to run.
About the platform
Is Analect a RAG tool?
Its knowledge engine solves a related problem differently. Retrieval sends less of a document; the engine changes what a retrievable unit is. Fragments are structured statements tied to their source sentences rather than prose chunks. Whether that is better for you is a question for a measured comparison against your existing retrieval, which is what a pilot runs.
Do we have to delete our source documents?
No, and nothing about Analect suggests you should. Fragments are an index into your knowledge, not a replacement for the record. Each fragment links back to the sentence it came from, which only works while the source exists.
Which workloads does the knowledge engine not help with?
Four of them. Summarise-everything work, because at whole-document level a fragment set is larger than the source. Corpora that change constantly, because decomposition is paid again on every change. Documents you question once, because there is nothing to amortise against. And small corpora, where reading everything is already cheap.
Who can sign in to Analect?
Customers and pilot teams, with accounts set up by ETT. Each account sees only the workflows it has been granted, and nothing stored in one organisation's account is visible from another's.
See it on worked material, then on yours
Thirty minutes with the team that runs Analect: the AI Spend Forecast and the Accuracy Evidence File on worked examples, and what either would look like on your bills or your outputs.