Licensing

Free to use. Not free to sell.

Noncommercial use costs nothing and asks nothing. Commercial use needs a separate licence. And the material a field is fed is not covered by either — that carve-out is deliberate, and it is the part most people miss.

If you are not making money from it, using it costs nothing and asks nothing. If you are making money from it, you need a separate licence. Neither licence covers the material you feed a field. That gap is on purpose, and it is the part most people miss.

Requesting a commercial licence →

The software

PolyForm Noncommercial 1.0.0

Used verbatim, not paraphrased and not modified. A lawyer-drafted instrument that people have already read and relied on is stronger than anything clearer we could write.

It is used word for word. It is not reworded and it is not changed. A licence written by lawyers, which people have already read and relied on, is stronger than anything clearer we could write.

It is source-available, not open source. The Open Source Initiative's definition forbids barring commercial use, so calling it open source would be wrong — and the word matters to people who rely on the label.

It is source-available, not open source. The Open Source Initiative’s definition does not allow a licence to block commercial use. So calling this open source would be wrong. That word matters to the people who rely on it.

The corpus

Not licensed by that at all

PolyForm licenses software. It says nothing about the material in Δ/, and that silence is not a grant.

PolyForm licenses software. It says nothing about the material in Δ/. Saying nothing is not the same as granting it.

belong to whoever made them. Someone may want to run the engine on their own material without acquiring rights in anybody's writing, and someone reading a story acquires no rights in the code. Both directions hold.

belong to whoever made them. You may want to run the engine on your own material. Doing that gives you no rights in anybody else’s writing. And reading a story gives you no rights in the code. Both directions hold.

What you may do without asking

Noncommercial use is free, and that is not a trial.

If you are not making money from it, it is free. And it is not a trial.

There is no key, no sign-up, no expiry, and nothing to renew. Read the source, run it, modify it, fork it, teach with it, write about it, use it in your own noncommercial work.

There is no key, no sign-up, no end date and nothing to renew. Read the source. Run it. Change it. Copy it. Teach with it. Write about it. Use it in your own work, as long as you are not making money from it.

UseUnder the free licence
Reading the source, running it locallyyes
Personal projects, hobby workyes
Teaching, coursework, a thesisyes
Modifying and redistributing, licence intactyes
Publishing what you found, including criticismyes
Anything inside a business, or that earns moneyneeds a licence
Serving it to customers, hosted or embeddedneeds a licence
Training, fine-tuning, evaluating or benchmarking a commercial systemneeds a licence

That last row is the one that catches people, and it is worth saying plainly rather than leaving in the small print: under PolyForm, evaluating the software for a company is itself commercial use. If you want to benchmark it against something you run at work, ask first. The answer to that particular question is very likely yes and costs nothing — see below.

That last row is the one that catches people. It is worth saying out loud rather than leaving it in the small print. Under PolyForm, testing the software for a company is itself commercial use. If you want to measure it against something you run at work, ask first. The answer to that question is very likely yes, and it costs nothing. See below.

Evaluating it

Ask, and the answer is yes.

A strict reading of the noncommercial term means an engineer at a company cannot benchmark this over a weekend, and a business cannot assess it before buying — because assessing it is commercial use. That is a licence blocking the step that leads to the licence, and it is a real cost of the instrument, not a misreading of it.

Read strictly, that term means an engineer at a company cannot test this over a weekend. A business cannot look at it before buying, because looking at it is commercial use. So the licence blocks the step that leads to the licence. That is a real cost of this licence, not a misreading of it.

So evaluation is granted by permission rather than standing open. Write to Admin@alphadataomega.com, say what you want to measure, and you will be told yes. It is one email, not a negotiation, and there is no fee.

So testing is allowed by asking, rather than being open to everyone. Write to Admin@alphadataomega.com, say what you want to measure, and you will be told yes. It is one email, not a negotiation. There is no fee.

What permission covers

Internal evaluation and benchmarking, including comparison against systems you operate commercially. Evaluation only — no production use, no redistribution. If you publish numbers, publish the method with them so somebody can check the work.

Testing and measuring inside your own organisation. That includes comparing it against systems you run for money. Testing only. No use in a live product, and no passing it on. If you publish numbers, publish the method with them so somebody can check the work.

Revocable at any time. Permission is a permission, not a grant that runs with the software, and it does not become a production licence by continued use.

It can be withdrawn at any time. Permission is permission. It is not attached to the software, and it does not turn into a licence for live use just because you keep using it.

Enforcement

What this is actually for.

The licence exists to keep the work attributable and to make commercial use a conversation. It does not exist to catch people. Noncommercial products are fine — that is what the licence already says, and it is meant plainly rather than as a technicality waiting to be argued.

The licence exists so the work stays credited, and so commercial use starts a conversation. It does not exist to catch people out. Products that do not make money are fine. That is what the licence already says, and it is meant plainly. It is not a technicality waiting to be argued.

There are two things worth enforcing against, and they are the same two either way:

Two things are worth acting on. They are the same two either way:

  • Using it to break the immutable laws of the system. They are published, content-addressed and unedited: denying choice, enforcing outcomes, fabricating certainty, optimising for dependency, rewriting trace, crowning authority, manufacturing motion.
  • Using it to break the fixed laws of the system. Those laws are published, named by their own bytes, and never edited. They rule out: taking away choice, forcing outcomes, faking certainty, building things people cannot leave, rewriting the record, crowning one authority, and making motion for its own sake.
  • Gaming the universe it runs in. Passing off a modified engine as the canonical one, forging provenance, or presenting a result the record does not support.
  • Cheating the system it runs in. Passing off a changed engine as the official one, faking where something came from, or showing a result the record does not support.

Short of those two, nobody is looking for you.

Read the laws before building on this. They are the part that is not negotiable, and unlike the licence they were not chosen for legal reasons.

Read the laws before you build on this. They are the part that cannot be argued. Unlike the licence, they were not chosen for legal reasons.

Commercial licensing

What a commercial licence covers, and what it cannot.

A commercial licence is granted by Alpha Data Omega LLC and covers the software: production use, hosting it for others, embedding it in a product, and training, fine-tuning, evaluating or benchmarking commercial systems with it.

Alpha Data Omega LLC grants the commercial licence. It covers the software: using it in a live product, hosting it for other people, building it into a product, and training, tuning, testing or measuring commercial systems with it.

It cannot grant you the corpora

No software licence can. The material under Δ/ belongs to the people who wrote it, and a licence from us conveys nothing in it — including the story in Δ/V·001, which is reserved to its author with no licence to reproduce, adapt, distribute, or use for training, fine-tuning, evaluation or benchmarking.

No software licence can. The material under Δ/ belongs to the people who wrote it, and a licence from us gives you nothing in it. That includes the story in Δ/V·001. It is kept by its author, with no licence to copy it, adapt it, hand it on, or use it for training, tuning, evaluation or benchmarking.

If your use needs particular material, you license that material from whoever holds it. Every registry entry carries its rights statement and its attribution string for exactly this reason. A contribution that states nothing grants nothing.

If you need particular material, you license it from whoever holds it. Every registry entry carries its rights statement and its credit line for exactly this reason. A contribution that states nothing grants nothing.

What it does not promise

  • It is not a warranty. PolyForm's disclaimer applies, and a commercial licence does not turn research software into a supported product unless the agreement says so in terms.
  • It is not a warranty. PolyForm’s disclaimer still applies. A commercial licence does not turn research software into a supported product, unless the agreement says so in terms.
  • It does not make the system generalise. Measured on three books it has never been fed — 828 probes, 400 bytes of context, exact next-word match, shipped defaults — it reads about 2% of them correctly, ranging 1–3% with sampling, against about 92% of the same probes after being fed those books. Feeding it more makes more material known; it does not improve the unknown. No licence changes that. If your use depends on it answering about things it was not shown, it will not work and you should not buy it.
  • It does not make the system read material it was never given. This was measured on three books it has never been fed. There were 828 probes. Each one gave it 400 bytes of context and asked for the next word, matched exactly, on the settings that ship. It reads about 2% of them correctly, ranging 1–3% with sampling. On the same probes, after being fed those books, it reads about 92% correctly. Feeding it more makes more material known. It does not improve the unknown. There is no transfer. No licence changes that. If your use depends on it answering about things it was not shown, it will not work and you should not buy it.
  • It does not transfer any trademark or the right to present your build as the canonical one.
  • It does not hand over any trademark. It does not give you the right to present your build as the canonical one.
Requesting a licence

Three things decide the terms.

There is no price list and no form to fill in badly. Terms are set per agreement, and they turn on the same three questions every time — what you are building, roughly what scale, and whether you need any specific corpus.

There is no price list and no form to fill in badly. Terms are set for each agreement. They turn on the same three questions every time: what you are building, roughly what scale, and whether you need any particular corpus.

That last one matters more than people expect. The software licence never conveys rights in anybody's material, so if your use depends on a particular corpus, that is a separate conversation with whoever holds it, and it is better had at the start than after the build.

That last one matters more than people expect. The software licence never gives you rights in anybody’s material. So if your use depends on a particular corpus, that is a separate conversation with whoever holds it. It is better had at the start than after the build.

Send this

Copy the block below, fill it in, and send it. It exists so a request arrives usable rather than as a paragraph that needs three rounds of questions.

Copy the block below, fill it in, and send it. It exists so a request arrives ready to act on, rather than as a paragraph that needs three rounds of questions.

Organisation
Contact name and role

What you are building
  one or two sentences, plainly

Deployment
  internal only / hosted for customers / embedded in a shipped product

Scale
  rough corpus size, rough number of users or queries

Corpora
  your own material only, or do you need a specific entry from the registry
  if so, which

Timeline
  when you need an answer by

Anything else we should know

Where it goes

Send it to Admin@alphadataomega.com, with the template in the body. One address, read by the company, no ticketing system in front of it.

Send it to Admin@alphadataomega.com, with the template in the body. It is one address, read by the company. There is no ticketing system in front of it.

What happens next

  • You get a straight answer on whether a licence is available for what you described, and roughly what shape it takes.
  • You get a straight answer on whether a licence is available for what you described, and roughly what it would look like.
  • If you are evaluating rather than deploying, you do not need any of this — the evaluation grant above already covers it, and nobody needs to reply for that to be true.
  • If you are testing it rather than putting it to use, you do not need any of this. The permission above already covers it, and nobody has to reply for that to be true.
  • If what you want depends on the system generalising to material it has not been fed, expect to be told no. That is not a negotiating position. On three held-out books, 828 probes at 400 bytes of context, it reads about 2% correctly — 1–3% depending on sampling — against about 92% of the same probes after being fed those books.
  • If what you want depends on the system reading material it has not been fed, expect to be told no. That is not a bargaining position. On three held-out books, 828 probes at 400 bytes of context, it reads about 2% correctly, and 1–3% depending on sampling. On the same probes, after being fed those books, it reads about 92% correctly.

This page is a summary and not an offer. No licence exists until there is a signed agreement, and nothing here creates one.

This page is a summary. It is not an offer. No licence exists until an agreement is signed. Nothing here creates one.

Contributing a corpus

You keep what you contribute.

Admission to the registry grants hosting, serving and pinning. Nothing else. No copyright is transferred, no exclusivity is taken, and the software licence does not extend over your material.

Being admitted to the registry lets us host your material, serve it and pin it. Nothing else. No copyright changes hands. No exclusive right is taken. The software licence does not reach over your material.

Before anything is admitted, an automated licence scan reads the material for rights signals — an explicit reservation, a NonCommercial or NoDerivatives term, ShareAlike, a layered licence where an ancient text is public domain but its modern editorial layer is not. Some sources carry one of these and some carry none. Run over the four texts that ship — node scan.js, 2026-08-18 — Pride and Prejudice returns a copyright notice with a year, nine times over, and Moby-Dick returns one public-domain assertion; Frankenstein and the story itself return CLEAR, no signal at all. So the scan finds things a filename never shows, and it finds nothing at all about half the time.

Before anything is admitted, a program reads the material for signs of who holds the rights. It looks for a stated reservation, a NonCommercial or NoDerivatives term, ShareAlike, or a layered licence. A layered licence is one where an ancient text is public domain but a modern editor’s work on top of it is not. Some sources carry one of these and some carry none. It was run over the four texts that ship on 2026-08-18. Pride and Prejudice came back with a copyright notice and a year, nine times over. Moby-Dick came back with one claim that it is public domain. Frankenstein and the story itself came back CLEAR, with no signal at all. So the scan finds things you could never see from a filename, and about half the time it finds nothing.

The scan reporting CLEAR is not permission. A bare text file has no header to find, which tells you nothing about whether you may use it. Provenance has to come from outside the file, which is what the citation block is for.

The scan saying CLEAR is not permission. A bare text file has no header to find. That tells you nothing about whether you may use it. Where a text came from has to come from outside the file. That is what the citation block is for.

Where a licence requires credit, the attribution string is stored on the entry and displayed wherever the corpus is used, not buried in a metadata file. Somebody transcribed and proofread that text, usually unpaid, and the credit line is the whole of what they asked in return.

Where a licence requires credit, the credit line is stored on the entry and shown wherever the corpus is used. It is not buried in a metadata file. Somebody typed that text out and checked it, usually for no pay, and the credit line is the whole of what they asked in return.

Terms

corpus
The body of text you feed it. That text is the only thing it can answer from.
holdout
Books the machine is never fed. They are used to test it on writing it does not have.