Dweve

ABMN working papers

Read six July 2026 ABMN working papers covering the model-family contract, NER, classification, reranking, lexical retrieval, and one generalisation test.

Six July 2026 ABMN working papers. Each record links authors, abstract and topics; the draft is released on request.

Adaptive Binary Memory Networks

Category
Model family
Authors
Dweve
Draft date

A model family that compiles learning into one artifact: byte-identical on every architecture, with no seeding, and replayable wherever it runs.

Abstract: Adaptive Binary Memory Networks (ABMN) ask a narrow question: can a learned model be fully and truly deterministic on every architecture, with no seeding involved, and does that property survive a change of task? The family answers with a contract rather than a tolerance. ABMNs do not compress or quantize a trained network, they compile it. One build produces the artifact, that artifact is byte identical wherever it is built, and it replays the same decisions on every host that opens it. What is deployed is the model itself, not a floating point checkpoint identified within a numerical error bound. This paper states the contract, the evidence a build has to carry and the limits of what it covers, then holds it across sequence labelling, classification, retrieval reranking and lexical retrieval, and carries it to a structurally different logical learner.

Topics: Deterministic machine learning, Reproducible builds, Cross-architecture portability, Compiled models, Deployment contracts

Read the summary

ABMN-NER

Category
Sequence labelling
Authors
Dweve
Draft date

Deterministic named entity recognition, compiled rather than compressed, and byte-identical on every architecture.

Abstract: ABMN-NER asks whether the ABMN build contract holds for sequence labelling, where every decision depends on the tokens around it and a small difference propagates along the sentence. The terms are the ones the family paper sets out. Nothing is compressed or quantized: learning is compiled into one deployed artifact, that artifact is byte identical on every architecture with no seeding involved, and decoding is exact rather than approximate, so the same text replays the same labelled sequence on any host that opens it. Figures are reported per dataset and per host. On CoNLL-2003, on an Intel i9, the artifact reaches 91.24% F1 and tags 606,000 tokens per second. Further datasets and benchmarks are reported in the paper itself.

Topics: Named entity recognition, Sequence labelling, CoNLL-2003, Deterministic inference, CPU throughput

Read the summary

ABMN-Classification

Category
Text classification
Authors
Dweve
Draft date

Supervised text classification compiled into one artifact: byte-identical across ARM and x86, with no seeding and nothing quantized.

Abstract: ABMN-Classification asks what the build contract costs on ordinary supervised text classification, and whether it survives a change of instruction set. Learning is compiled rather than compressed or quantized: the deployed artifact is the model, one build produces it with no seeding involved, and it replays the same predictions wherever it is opened. Independent builds of all eleven tasks produced byte identical artifacts and identical prediction streams across ARM and x86. Measurements are per dataset and per host. On DBpedia-14, on an Intel i9, the artifact reaches 98.89% accuracy and 98.89% macro F1; on IMDb, on the same host, it sustains 90.75 million tokens per second at 1.38 microjoules per token. The full set of datasets and benchmarks appears in the paper.

Topics: Text classification, MASSIVE, DBpedia-14, Cross-architecture reproducibility, Energy per token

Read the summary

ABMN-Ranking

Category
Information retrieval
Authors
Dweve
Draft date

Deterministic candidate reranking compiled into one artifact, with no dense model in the hot path and the same ordering on every host.

Abstract: ABMN-Ranking asks whether the build contract reaches a stage that is normally served by a dense model: reranking a candidate list. It does not compress or quantize such a model, it replaces the stage with a compiled artifact, so no dense vector or transformer inference runs in the hot path and scoring is exact integer work. One build produces the artifact, no seeding is involved, and the same query returns the same ordering on every architecture that opens it. Figures are per dataset and per host. On SciFact, on an Intel i9, the artifact reaches 0.809 nDCG@10 at 14,059 queries per second, which is 593 microseconds per query. Additional datasets and benchmarks are set out in the paper.

Topics: Retrieval reranking, Search relevance, SciFact, Query latency, Deterministic ranking

Read the summary

ABMN-Relevance

Category
Lexical retrieval
Authors
Dweve
Draft date

Corpus-adaptive lexical retrieval compiled into one artifact, deterministic and identical across architectures without seeding.

Abstract: ABMN-Relevance asks whether the build contract reaches the retrieval stage itself, where the corpus rather than a training run decides what the deployed model contains. The index is compiled into one artifact rather than compressed or quantized, and retrieval over it is deterministic: no seeding is involved, and the same artifact replays the same ranking on every host and every architecture it is opened on. Against a tantivy BM25 baseline over the same collection it returns better ranked results and serves them faster, from a smaller index and a lower resident memory footprint. The measured numbers are held back until the paper publishes. What is claimed here is the contract, and that a lexical retrieval index can be held to it.

Topics: Lexical retrieval, Search indexing, MS MARCO, Index footprint, Replayable retrieval

Read the summary

ABMN-Tsetlin

Category
Logical learning
Authors
Dweve
Draft date

Reproducible binary model compilation carried beyond associative classification to a structurally different logical learner.

Abstract: ABMN-Tsetlin asks whether the build contract is a property of one model family or of the approach behind it. It takes a structurally different logical learner, trained by a different rule, and holds it to the same terms. The trained Tsetlin machine is compiled into one deployed artifact rather than compressed or quantized. Across five image and text tasks that artifact came out byte identical on every architecture, with no seeding involved, and replays the same decisions wherever it is opened. Results are recorded per dataset and per host. On MNIST, on an Intel i9, it reaches 97.61% accuracy from a 40.4 KiB artifact. The paper carries the remaining datasets and benchmark tables.

Topics: Tsetlin machines, Logical learning, MNIST, Image classification, Per-host evidence

Read the summary

These are working drafts. The repository does not establish a journal, conference, or peer-review outcome.

Choose the audience that matches your question

The page contains three selectable readings of the same subject.

For consumers

Working papers with measured results and readable summaries, each draft released on request. Every number stays tied to its task, dataset, configuration, and environment.

For businesses

All six entries are July 2026 working papers. The repository records the source and the claim boundaries; it does not establish submission, peer review, or publication.

For engineers

Papers that describe how the stack is built, not just what it achieves. Read them to understand why a subsystem was designed the way it was. Each entry stays labelled as a working paper.

The ABMN papers explain one approach to building compact, repeatable AI systems. Start with the family paper, then follow the task that interests you. These are July 2026 working drafts, not claims of peer-reviewed publication. They exist to show that full determinism is achievable and holds across different kinds of task, not to chase top scores. The repositories publish with the papers.

Six ABMN working papers set out the family contract, task-specific memory structures, measured results, and stated limitations. The register separates draft evidence from any claim of external acceptance. The purpose is to demonstrate that full determinism is achievable and generalises across task types, rather than to claim state-of-the-art results. The repositories publish with the papers.

The six July 2026 drafts describe the ABMN family contract, four task-owned instantiations and one generalisation test that applies the same build contract to a logical learner. Each record carries the abstract, the section outline, the measurements and the stated limitations. The set exists to prove that full determinism is achievable and generalises across task types, without disclosing the proprietary work behind it. The repositories publish with the papers.

Each card gives you the question, the subject, and a summary of the paper. The titles stay exactly as they appear in the source papers. There are six of them, all dated July 2026. The complete draft is with external reviewers, so ask for it through the contact form and we release it once we approve the request.

Use this shelf as the review register. It contains the complete current ABMN working set and does not mix in product brochures or unsupported sector papers. Six titles carry one draft date and two named authors, so a review can start from the register rather than from a search. The drafts themselves are released on request.

The family paper defines the shared contract. Four companion papers instantiate it for NER, classification, reranking, and lexical retrieval, and a sixth asks whether the same build contract holds for a structurally different learner. All six carry the same draft date. The summary stays public; the draft is released on request.

One family paper, four task papers, one generalisation test

All six source documents carry this date

You do not need to read all six. The family paper provides the map, the task papers show how the same pattern changes for names, labels and search, and the last one asks whether the pattern survives a different kind of learner. The order below runs from the contract to the tasks and then to retrieval, but it is a suggestion rather than a ranking, and any single paper stands on its own. If you only read one, read the family paper, because the rest assume the shape it describes.

Review the family contract before comparing task results. It prevents a benchmark from being mistaken for a system-wide guarantee. The sequence runs from the family contract to the task papers, then to the two retrieval papers, and last to the generalisation test. It is editorial guidance rather than a maturity ranking, and any of the six can be read on its own once the contract is understood.

Read the family definition first, then compare how each task owns its memory, its build boundary and its evidence ledger. The sequence runs contract, task papers, retrieval papers, generalisation test. It orders the reading, not the maturity, and each paper states the scope its own measurements cover. The family paper is the only one the others depend on, so it is the one to read first if you read only one.

Definitions, build phases, and four memory organisations

Bridge reranking and corpus-adaptive indexing

The sequence is editorial guidance, not a maturity ranking.

See how the six papers fit together before choosing a task.

Follow a compact system that finds names and places with exact structured decoding.

See how a text classifier can become a reproducible binary build.

Inspect search reranking without a dense model in the query path.

Inspect lexical retrieval with deterministic scoring.

Establish the shared model-family contract and its evidence boundaries.

Review reported quality, footprint, throughput, and the NER limitations together.

Review the cross-architecture artifact and prediction identity claim.

Review the measured bridge-memory reranking path on SciFact and NF-Corpus.

Review the corpus-adaptive serving path and its MS MARCO measurements.

Define learn, strip, execute, and the four task-owned memory organisations.

Trace sequence labelling from one build to exact decoding.

Trace text classification to byte-identical artifacts across ARM and x86.

Trace typed sparse atoms into compiled bridge tables and integer reranking.

Trace lexical retrieval from the corpus build to deterministic top-k results.

The papers share a basic move: learning prepares compact task memory, build-only state is removed, and a fixed procedure uses the result. Four subjects cover the family paper and the task papers, and each of those belongs to exactly one subject. The generalisation test sits beside them, asking whether the same move works on a different kind of learner.

The shared contract is the comparison point. Task-specific quality and systems numbers remain local to the paper and evaluation that reports them. Four subjects cover the family contract and the task papers, and each of them sits under exactly one subject. The generalisation test is reported separately, because it tests the contract rather than a task.

The family contract stays stable while addressing, update rules and decoding differ by task. The subject map keeps those instantiations separate, with four subjects and exactly one subject per task paper. The generalisation test carries the contract to a structurally different learner and reports its evidence per host.

Named entity recognition with exact structured decoding

One compiled artifact, identical on every architecture

You can read what each paper claims and how it was tested, and ask for the draft behind it. Neither of those means a journal, conference, or independent reviewer has accepted the work.

Treat the drafts as review inputs. External acceptance, independent replication, and production assurance require their own evidence and are not asserted here.

Methods and measurements are presented in the source drafts. Their scope, environment, baselines, and limitations travel with each claim.

All six entries are working papers dated July 2026.

The date describes the draft snapshot, not a publication event.

Every entry has a full draft, and Dweve releases it on request.

The drafts are with external reviewers, so nothing is published here. Ask through the contact form and we release the paper on approval.

Reported results belong to the task, dataset, configuration, and environment named in the paper.

A result in one paper is not promoted to a family-wide guarantee.

The repository does not establish a journal, conference, or peer-review outcome for these drafts.

The library therefore uses working-paper language throughout.

Publication or peer-review status will only be stated when the repository supports it.

This register shows where each title comes from and how to get the paper itself. There are no extra papers hidden behind marketing names. Every entry names one source directory, and the draft that directory holds is the one we send when you ask for it.

Use the register to verify that the library entry, the detail route, and the draft we release all refer to the same working paper. Each title names one authoritative source directory, so a claim on this page can be traced to the document that supports it.

The register binds each canonical slug to its authoritative source directory. Detail pages carry the abstract, section outline, and stated status; the complete draft stays with external reviewers and is released on request. No entry appears here without a source directory behind it, which is what makes the register checkable rather than asserted.

One draft per source, released on request

Every listed paper names its authoritative source directory, and its draft is released on request.

The summaries help you choose. The complete draft carries the methods, results, limitations, and references, and we release it when you ask for it.

Move from the family contract to the task paper, then take the stated evidence boundary into due diligence.

The library keeps the family definition, the task instantiations, the generalisation test, and their source paths in one reviewable sequence.

The six titles and dates come from the authoritative paper sources. The shelf does not rename old marketing material or create papers around product claims. What is listed here is the complete current working set, and nothing sits in it that the source directory does not carry.

Each detail route offers a concise orientation. The draft remains the complete record of methods, reported measurements, limitations, figures, and references, so a summary is a way in rather than a substitute for the paper.