Matrix Cognition

October 2026

3 articles

Tutorials13 Oct 202612 min read

Hybrid search in Python: BM25 plus embeddings, measured

BM25 and two embedding models fused by RRF and by a weighted blend on SciFact. A published blend weight helped one model and hurt the other.

Code Data
Evaluation9 Oct 202613 min read

Is your test set in the training data? Four overlap rules, measured

The GPT-3, PaLM, Llama 2 and GPT-4 contamination rules in one stdlib script. On GSM8K, 13-word overlap caught 69.1-95.6% of edited copies, 0.8% of rewordings.

Code Data
Paper Readings6 Oct 202615 min read

TORQUE re-implemented: where keeping outliers around a rotation pays

A reading of arXiv 2609.36032v1, re-run on Qwen2.5-0.5B: keeping outliers cut activation error at every budget; the KV-cache gain faded once its bits were paid.

arXiv 2609.36032 Code Data

September 2026

13 articles

Tutorials25 Sep 20265 min read

Counting tokens before the bill: six tokenizers, six kinds of text

The same 20,000 characters came to 4,716 to 33,502 tokens depending on language and tokenizer. Measured counts, the divide-by-four rule tested, a forecast script.

Code Data
Systems22 Sep 202611 min read

What a quantisation format actually costs, derived from the source

Q4_K is not four bits per weight. Exact costs derived from ggml-common.h, joined to llama.cpp's measured size and perplexity tables, and the dominated formats named.

Code Data
Paper Readings18 Sep 20268 min read

Where should a document live: context, KV cache, or weights?

A reading of arXiv 2609.17346v1, which compares in-context documents, KV-cache injection and fine-tuning on five benchmarks. Tables re-derived, code included.

arXiv 2609.17346 Code Data
Tutorials15 Sep 20269 min read

LLM-as-judge done carefully: position bias, length, and agreement

A pairwise judge harness run in both orders over a 48-item golden set, with swap tests, a padding probe, and Cohen's kappa. Code and data included.

Code Data

Earlier in September

Aims and scope

Four kinds of article, each with a different standard of evidence.

Tutorials build something that runs: a retrieval pipeline, an evaluation harness, a fine-tuning run, a serving setup. The code ships with the article as a complete listing, the package versions are named, and the numbers printed in the text came from running that code.

Paper readings take one paper, cited by its arXiv identifier, and separate three things that are usually blurred together: what the authors claim, what they actually measured and under which settings, and what could be reproduced or checked here. A result quoted from a table says which table.

News covers model releases, tooling changes, and benchmark movements, dated from the primary announcement (the lab's post, the model card, the release notes) rather than from secondary coverage, with the link.

Evaluation and systems pieces cover the parts people skip: how to measure a pipeline without fooling yourself, what quantisation does to a specific model, what inference actually costs.

Sections

All 16 articles

Editorial standards

No benchmark numbers from memory. No "reportedly". No capability claims lifted from a press release and restated as fact. No invented quotes from researchers. If something could not be verified on the day of writing, the article says what is unknown instead of guessing. No exclamation marks.

Published under the owner’s direction, with the code, the sources and what could not be checked on the page. How this site is made.

October 2026

Hybrid search in Python: BM25 plus embeddings, measured

Is your test set in the training data? Four overlap rules, measured

TORQUE re-implemented: where keeping outliers around a rotation pays

September 2026

Counting tokens before the bill: six tokenizers, six kinds of text

What a quantisation format actually costs, derived from the source

Where should a document live: context, KV cache, or weights?

LLM-as-judge done carefully: position bias, length, and agreement

Earlier in September