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.
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.
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
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
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.
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
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.
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.