Deterministic AI research
Every output can be re-derived by hand.
Aleten builds Ackren, a deterministic conversational engine for regulated deployment. No statistical inference anywhere in the pipeline, single-digit millisecond responses, and a derivation log for every answer. Ackren is pre-product: the development log is public, the specifications go out as each one is finished, and the engine is live to try while the product demo is four to six months away.
A binder's back mark. A correct gathering shows an unbroken diagonal, and a wrong one visibly breaks. Six hundred years old, no instrument required, read by eye in a second. Hover or tap the mark to put it out of sequence.
Announcements
Recently announced
Ackren answers in 3.6 milliseconds
Measured end to end on an 80-word reply, against seconds for a hosted model. Latency here is architectural rather than a tuning result, and it does not erode under load. The harness goes out with the figure.
Serena, for emotional interpretation
A second engine, reading affect and stance rather than propositional content, on the same inspectable footing as Ackren. The specification is being written and will go out in full, limits included.
Quire, where domain packs are gathered
A quire is a set of leaves folded together, which is what a pack is. Packs are authored, checked and signed in Quire before they ship. Early access opens with the first design partner.
Development log
What changed, and in which build
Every entry in full, monthly minimum. If the cadence is missed, this page says so rather than going quiet. Corrections are a first-class entry type and superseded figures are struck, never deleted.
Showing all 9 entries
The formal specification is finished at v0.4
It goes out as it stands, with no form in front of it and no email address collected. The maths is the strongest asset here, so it is the first thing to leave the building.
Spec v0.4
Trigram hit rate raised to 0.461 on the held-out set
Up from 0.449. Pruning still removes 55.1% of trigram mass, so a 16 MB build remains a bigram model with trigram exceptions.
Build 7f2a9c31
Majority voting withdrawn from the annotation regime
It only works if errors are independent, and on ambiguous items they are not. Replaced with triple annotation and expert adjudication.
Regime v2
The AI Act dates after the Digital Omnibus
Annex III high-risk deferred to 2 December 2027. Article 50 transparency stays on 2 August 2026, and it applies to us. Determinism buys nothing there.
Reviewed 11 Jul 2026
Derivation log schema settled, versioned and open
The record format fixed before anyone asked for it, under a permissive licence, with a changelog and an invitation to implement it.
Schema v1.0.0
Silent lookup errors, counted rather than estimated
Roughly 6,100 in a billion queries at the current index width. They are bought off with bits, and the bits cost most of a 16 MB budget.
Spec v0.4 §7
The verification programme, and what it pays
Four hundred hours at fifteen pounds, paid per verified hour rather than per item, so the incentive is not to rush the unsure cases.
Programme v1
Octavo fits in 16 MB, with a weaker guarantee
One false accept in 4,096, against one in 16.7 million on Quarto. Twelve-bit fingerprints are all the budget allows, and that is said with the name.
Build 4c81e0a2
The energy comparison has been withdrawn
It rested on a batch size nobody has published, so it was not a figure we could defend. Struck until it can be recomputed. Latency is unaffected.
2 to 6 ×10⁵ · withdrawn, not replaced
Claims v1.1
Mean fanout measured at 2.60
One to three reads per lookup, which is what makes a memory-mapped artefact viable.
Build 9d1c40b7
Speech act layers one and two land in the core
They come with the engine as a mathematical consequence, not as a product decision.
Spec v0.3
IEC 62304 gap analysis, first pass
Deterministic software has a certification path. Most of the gap is documentation we have not written.
Reviewed 8 May 2026
The SRAM figure was quoted without its qualifier
"Runs in 50 KB" does not survive one technical reader. The full form is now the only form.
Claims v1.0
The repair path was built before the response path
A system that is honest before it is capable can be made capable. The other order rarely recovers.
Build 2b7fd014
Contract terms published before hiring anyone
IP assignment and confidentiality shown before the first batch, not after it.
Programme v0.9
Thirty-six thousand qualia relations, and what they miss
The count is the easy part. Coverage is the number that decides whether the engine can answer you.
KB v0.6
First end-to-end turn on Quarto hardware
One question, one answer, one derivation record, on a board with an SD card and no operating system.
Build 5e9a1c33
The repair surface replaces the answer
Not a banner above it, not an amber underline. When something is flagged there is no answer to decorate.
Spec v0.2
The names are printer's sheet folds: folio is one fold, quarto two, octavo three. Format names the hardware floor, not the competence, and any pack runs on any format subject only to flash. This is not a capability ladder.
Folio
One fold · LinuxThe full artefact at gigabyte scale, memory mapped, where storage is not the binding constraint. Fingerprints at full width.
Quarto
Two folds · PSRAMRoughly 200 MB against a card. The middle floor, and the one most embedded targets can actually meet without redesigning the board.
Octavo
Three folds · 16 MB XIPSixteen megabytes, executed in place, no card at all. Octavo carries a weaker guarantee rather than merely less content: twelve-bit fingerprints are all the budget allows.
Domain packs
GatheredCompetence ships as a pack: authored, listed and checkable. Conversation is the base engine rather than a pack, and authorship is a research problem rather than a forthcoming one.
Run it yourself
The engine is live at ackren.com
It grounds what you tell it, asks its own follow-up questions, and prints the derivation log for every turn, including the typed flag on the turns it refuses. It knows nothing yet and says so rather than guess. The playground needs no account and the conversation never leaves your browser.
Research
Publish the mathematics
Every document is dated, versioned and citable, and every figure carries the basis it was derived from: measured, computed, modelled or cited. A number without its basis is not a claim, and it does not appear on this site.
- Grammar and parsing
- Qualia relations
- Determinism
- Verification
- Derivation logging
- Regulated deployment
- Jul 2026 Formal specification of the Ackren engine v0.4
- Jul 2026 The eight stages: pipeline specification, with typed flags v0.7
- Jul 2026 Derivation log record, an open and versioned schema v1.0.0
- Jun 2026 Failure accounting: eleven rows, two of them unsound Note
- May 2026 The 16 MB fit is vocabulary limited, not n-gram limited Note
These are not posted yet, so this is a list rather than a set of links. Each one goes up at a stable URL as it is finished, with nothing behind a form. Ask for any of them now and you will get the current version: research@aleten.com.
Limits
What this does not solve
Every known failure mode is published before anyone else finds it. Four of them are below, with the detection condition stated alongside each.
The full accounting runs to eleven rows, two of which are unsound. It goes up with the specification. Each of the four below has its own page: the problem, the impact, the honest reading and the roadmap.
Novel bridging
Detected, never resolved. Coverage is strictly less than one no matter how much is authored, and no amount of authoring closes the gap.
Non-conventional implicature
Detection is achievable. Resolution is not. This one is open for everyone, including transformer models, which do not flag it when they fail.
Long-form coherence
Nothing in the model holds an argument together across paragraphs. Authorship is a research problem, not a forthcoming feature.
Cross-build determinism
Determinism holds within a build, identified by artefact hash. It does not hold across builds, and we do not claim that it does.
Work with us
We are hiring, and the terms are published
The verification programme pays fifteen pounds an hour, per verified hour rather than per item, so the incentive is not to rush the unsure cases. The quality regime is public before you apply: five per cent seeded gold items, twenty per cent triple annotated, and Fleiss' kappa reported.
Annotator, verification programme
Verify the lexicon and qualia entries the engine derives from. Binary judgements, learnable in an afternoon, adjudicated rather than voted on.
£15 / verified hourContractRemote
Grammar author
Author rules and frames for a scoped regulated domain. Centralised by design: this is the one part of the knowledge base that is not crowd-verified.
PermanentLondon or remoteLinguistics
Embedded systems engineer
Make a 16 MB execute-in-place artefact behave on hardware that has no filesystem, no MMU and no second chance.
PermanentRemoteC, no RTOS assumed
Regulatory and standards lead
Own the AI Act position and the standards path: IEC 62304, DO-178C, ISO 26262, EN 50128. Includes owning the review cadence on a page people rely on.
PermanentEU or UKNamed owner
Research engineer, parsing
Work on the eight stages, the flag algebra and the repair path. The repair path is built before the response path, and it stays that way.
PermanentRemoteFormal methods
No application form and no tracking system. Every role above has its own page with the terms on it. Write to work@aleten.com saying which one you want and a person will reply.