Independent research, no budget.
James Sterling Tuttle, working as Alpha Data Omega LLC. No institution, no grant, no investor, no staff. The work so far was done without money, and it is published as it happens — including the parts that failed and the parts later retracted.
James Sterling Tuttle works as Alpha Data Omega LLC. There is no school behind him, no grant, no investor and no staff. The work so far was done without money. It is published as it happens. That includes the parts that failed, and the parts he later took back.
The company is looking for funding to grow as a research company. Not to launch a product. There is no revenue model being promised here.
The company wants money so it can do more research. It is not building a product. It is not promising a way to make money.
A memory that
refuses.
You give it a body of text. It answers out of that text word for word, and when what you type does not match what it holds, it says so instead of producing something plausible. It distinguishes three cases, not two: yours in order, your words out of order, and not yours at all.
You give it a body of text. It answers using only that text, word for word. If what you type is not in that text, it says so. It does not make up something that sounds right. It tells three cases apart, not two. Your words in your order. Your words in a different order. Not your words at all.
A derived
address geometry.
Bytes become
addresses.
Bytes become level by level, compared by rather than by an attention window — which is why recall does not decay from the front of a haystack to the back. One constant is genuinely derived — the chance floor, at exactly 7/16, from two lines of algebra. The rest were chosen and then measured, and the theory page lists which is which rather than claiming otherwise. Choosing a value because it measured better is tuning; saying so costs nothing.
Bytes turn into , one level at a time. Two things are compared by how much their addresses match from the start, not by an attention window. That is why it recalls a passage at the end of a long text as well as one at the front. One constant is genuinely derived — the chance floor, at exactly 7/16, from two lines of algebra. The rest were chosen and then measured. Choosing a value because it measured better is tuning. The theory page says which is which. Saying so costs nothing.
An append-only
log, retractions in.
A log that is only
added to.
Every result is written to a working log that is never edited, only added to. Where a later entry contradicts an earlier one, both stay. The engine's own README carries a correction against its earlier drafts, in the README, rather than a quiet edit.
Every result goes into a working log. Nothing in that log is ever changed. Things are only added. When a later entry disagrees with an earlier one, both stay. The engine’s README carries a correction to its own earlier drafts. The correction sits in the README. It was not a quiet edit.
Read it there, not here.
This page is about the company. The research has its own surfaces, and describing them twice would only put a second version of each claim into the world.
This page is about the company. The research has its own pages. Writing it out twice would put a second version of each claim into the world.
The mechanism end to end, in the order the machine does it, with the limits in the body rather than a footnote.
How it works, step by step, in the order the machine does it. The limits are in the main text, not in a footnote.
The algebra and the exhibits, including the places where the notebook later contradicted itself.
The maths and the worked examples. It also shows the places where the notebook later contradicted itself.
What is unresolved, ranked by how many people reproduced it rather than by how many liked it.
What is still unsettled. It is ranked by how many people got the same result, not by how many liked it.
The self-reported LongMemEval-V2 number with the reproduction bundle and the hashes.
The LongMemEval-V2 number we reported ourselves, with the files and the hashes needed to run it again.
Preregistrations, matched baselines, and the falsification ledger — the predictions that broke, with the dates they broke.
Plans written down before the test, fair comparisons, and the record of broken predictions, with the dates they broke.
It does not generalise.
It only knows what you gave it.
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% depending on how the probes are sampled. After being fed those same books it reads about 92% of the same 828 probes.
It was tested on three books it had never been fed. There were 828 probes. Each probe gave it 400 bytes of context and asked for the exact next word. The settings were the shipped ones. On those books it reads about 2% of them correctly, ranging 1–3% depending on how the probes are sampled. Then it was fed those same books. On the same 828 probes it then reads about 92%.
There is no transfer. Feeding it more material makes more material known; it does not improve the unknown. Anything on this page that sounds like capability should be read against that pair of numbers, and if a use depends on it answering about material it was not shown, it will not work.
There is no transfer. Feeding it more text makes more text known. It does not get better at text it was never given. Read anything on this page that sounds like a capability against that pair of numbers. If your use needs it to answer about text it was not shown, it will not work.
A symmetry can constrain but cannot locate. Every theorem that follows from the four-group alone holds equally for a function with zeros off the line, so the route taken here could not have closed it. Roughly fifteen exact restatements of RH were produced along the way and recorded as restatements; an audit of Entries 004–011 killed five of its own routes with the verdicts written down — circular, false, restatement, vacuous.
A symmetry can constrain but cannot locate. Every theorem that follows from the four-group alone is just as true for a function with zeros off the line. So the route taken here could never have closed it. Roughly fifteen exact restatements of RH came out of the work, and each was written down as a restatement. An audit of Entries 004–011 killed five of its own routes and wrote down why: circular, false, restatement, vacuous.
RH remains open. The working log's terminal status line says so at the top.
RH remains open. The working log says so on its top line.
Every figure with what it measured, and its N.
Every figure, with what it measured and how many tries.
A number without its configuration is not a number, so each row names its material and its sample size. Everything in this table can be re-run from the engine and the linked above — with one exception, kept out of the table and stated separately below, because it cannot. Nothing here is an accepted third-party result.
A number means nothing without the settings it was made under. So each row names the text it was run on and how many tries it got. Every row in this table can be run again from the engine and the linked above. One row cannot, so it was kept out of the table and is stated below on its own. Nothing here has been checked by anyone outside.
| What was measured | Result | On what, and how many |
|---|---|---|
| A four-line exact-substring index, on the same probes | identical | 764 of 828, the same figure the engine gets, per book to the unit — because a verbatim probe is verbatim in the corpus. The recall numbers are not what distinguishes this system; the refusal behaviour is |
| The engine | 437 lines · 288 not comment or blank | No dependencies, no model, no training, no weights |
| Recall of passages it holds, by depth in the haystack — recall meaning it spoke rather than refused, which includes the middle verdict | 1.00 at all 10 depths | 300/300 probes, ten bins of 30 · 5.94 MB fed · 1,077,603 states · window 11 · nine pre-1919 Gutenberg books fed, one held out |
| Of those same passages, how many came back as mine, in order rather than my words, not my order | 299 of 300 | The same run. The two figures are different questions: the first asks whether it spoke, the second asks which of the two speaking verdicts it gave. One passage of the 300 was held as “my words, not my order” |
| Refusal of random junk | 1.000 | 150/150 probes, same run |
| False-claim rate on a book never fed | 0.003 | 1 error in 300 probes; the one error was inspected and is an address collision — two different byte strings on the same sequence — not a fabrication |
| Next-word reading, material never fed | about 2% | 828 probes · three held-out books · 400 bytes of context · exact match · shipped defaults |
| The same probes, after being fed those books | about 92% | The same 828 probes, same configuration |
| The Riemann Hypothesis route | negative | Five routes killed in the audit of Entries 004–011, verdicts recorded; RH open |
The one figure that is not in that table
A LongMemEval-V2 Small run scored 56.76% — 256 of 451 questions, official public scorer, 0.187 s average query latency, self-reported and independently unverified.
A LongMemEval-V2 Small run scored 56.76%. That is 256 of 451 questions. The official public scorer was used. A query took 0.187 s on average. We reported this ourselves, and nobody outside has checked it.
It sits outside the table on purpose. That number was not produced by the engine this page documents and cannot be reproduced from anything linked here: the benchmark file in the repository states plainly that LongMemEval measures things this system does not do, and that running it would produce numbers that mean nothing. The figure came from a different, non-public system. It is reported because it exists and because omitting it would be worse, not because a reader can check it.
It sits outside the table on purpose. The engine this page describes did not produce that number, and you cannot get it again from anything linked here. The benchmark file in the repository says plainly that LongMemEval measures things this system does not do, and that running it would give numbers that mean nothing. The number came from a different system that is not public. It is here because it exists, and leaving it out would be worse. It is not here for you to check.
Two of these are limits rather than achievements, and they sit in the same table on purpose. A results table that only carries the wins is indistinguishable from one that got lucky.
Two rows in that table are limits, not wins. They sit there on purpose. A table that shows only the wins looks exactly like a table that got lucky.
What money would be for, and what it would not.
Funding would be to grow Alpha Data Omega as a research company — one person doing this full time rather than around everything else, and the specific unrun measurements getting run. Concretely, that is four things:
Money would let Alpha Data Omega grow as a research company. One person could do this full time instead of in the gaps around everything else. The measurements that have not been run would get run. That is four things:
- The scale that has not been measured. Nine books is not PG-19. PG-19 is roughly 28,000 books and about 11 GB. The reading would move from 11 to roughly 18 at that scale, and that prediction is currently unmeasured because the compute and storage to run it do not exist here.
- The scale that has not been measured. Nine books is not PG-19. PG-19 is roughly 28,000 books and about 11 GB. At that size the reading should move from 11 to roughly 18. Nobody has measured that yet. The computers and storage needed to run it do not exist here.
- Independent verification. The benchmark result is self-reported. It should be re-run by someone with no stake in it, which costs money that a self-reported result does not.
- Independent verification. We reported the benchmark result ourselves. Someone with nothing to gain from it should run it again. That costs money. Reporting it ourselves did not.
- Time. The research is currently done without a budget, and that is the binding constraint on how much of it happens.
- Time. The research is done with no budget. That is what limits how much of it gets done.
- Keeping the record public. The log, the corpora and the content-addressed artifacts have to stay hosted and pinned for the reproduction claims to mean anything a year from now.
- Keeping the record public. The log, the texts and the stored files have to stay online and pinned. If they go away, the promise that anyone can run this again means nothing a year from now.
What it is not for
Not a launch, not headcount, not a go-to-market. There are no users, no revenue and no traction to report, and none is being forecast here. The only commercial surface that exists today is the licence: the software is PolyForm Noncommercial 1.0.0 — free to use, modify and share for any noncommercial purpose, with commercial use requiring a separate licence from Alpha Data Omega LLC. It is source-available rather than open source, because the OSI definition forbids restricting fields of endeavour and that is the honest word. The corpora are covered by neither. See ado.earth/license. That is the whole of the commercial position.
Not a launch. Not hiring. Not a sales push. There are no users, no money coming in and no growth to report, and none is being predicted here. The only business thing that exists today is the licence. The software is PolyForm Noncommercial 1.0.0. You may use it, change it and share it for free for any purpose that is not commercial. Commercial use needs a separate licence from Alpha Data Omega LLC. It is source-available, not open source. The open source rules do not allow limits on what field you use it in, so open source would be the wrong word. The texts are covered by neither. See ado.earth/license. That is the whole business position.
The amount, the term, and the structure are James's to state, and are not written here. Background and history likewise: nothing has been composed on his behalf, and this space stays empty until he fills it.
How much, for how long, and on what terms are James’s to say. They are not written here. The same goes for his background and his history. Nobody has written anything for him. This space stays empty until he fills it.
One address.
Funding, licensing, evaluation permission, or a correction to anything above — Admin@alphadataomega.com. It is read by the company, which is one person, so a reply is a reply and not a ticket.
Write about funding, licensing, permission to test it, or a correction to anything above. The address is Admin@alphadataomega.com. One person reads it. A reply comes from that person, not from a ticket system.
If you have found something on this page that the record does not support, that is the most useful message you can send.
If you find something on this page that the record does not back up, say so. That is the most useful message you can send.
Terms
- address
- A short string of symbols that says where a thing sits in the field. Two things are alike when their addresses start the same way.
- prefix agreement
- How many symbols two addresses share from the start. The more they share, the closer the two things are.
- window
- The small run of bytes just before the spot being read. At that moment the machine looks at that run and nothing else.
- corpus
- The body of text you feed it. That text is the only thing it can answer from.
- ledger
- The record of what the machine was fed, written down as addresses. It is only added to. Nothing in it is edited later.
- trit
- The first symbol of a byte’s address. It is what the index is sorted by. The name sounds like three, but it has four possible values.
- holdout
- Books the machine is never fed. They are used to test it on writing it does not have.