Essays
Reading, evidence and the digital archive
Essays on the questions behind Archēglyph: when a translation can be trusted, how to read a theme as evidence, what makes a machine-read quotation citable — and, for those who want it, the machinery underneath.
Reading, evidence and method
For anyone working in the archive: what machine reading can and cannot give you, and how to cite what it does.
-
A translation is not the source
A machine translation is a claim about a text, produced by a named model on a named day — not the text itself, any more than OCR output is the page.
-
Downstream of Trove: analysis in the corpus stack
Digitisation projects like Trove, Chronicling America, and Europeana produce the corpus; Archēglyph produces the analysis on top of it — one stack, not competitors.
-
Orthogonal to LLM 'deep research'
Deep-research agents synthesise; Archēglyph indexes. Different problems, different workflows — knowing which you need keeps your citations defensible.
-
The citable-claim test
A simple test for a research tool's output: can you, in one click, see the page a claim came from? If not, it's for exploration, not for scholarship.
-
Why Archēglyph cannot hallucinate
Hallucination is a property of generative systems. Archēglyph isn't one: every surfaced line was already in the source corpus, and we show you the page.
-
Reading clusters as a researcher
The Archēglyph cluster view leads with quotations, not scatterplots. Here is how to use it — and why the scatterplot is behind a toggle.
-
What a good provenance badge looks like
UX writing about the transparency contract: what goes inside a provenance badge, what gets omitted, and why the re-run affordance sits right next to it.
-
Transparency is a feature
Why every extracted text block in Archēglyph shows the model that produced it, and why we treat that disclosure as product surface rather than footer text.
For the technically curious
The choices under the hood, argued in the open.
-
Why we snapshot per dataset
The decision behind Archēglyph's dataset snapshot: one tarball bundling a tantivy index, a zvec store, and a sqlite catalogue — why the unit is the dataset.
-
Choosing an embedding model for digital humanities
A practical comparison of MiniLM-L6-v2 and BGE-small-en-v1.5 for DH corpora: what each optimises for and when the extra dimensions earn their keep.
-
VLM vs OCR: when to pick what
Notes from the newspapers prototype on when Tesseract is still right, when a vision-language model earns its cost, and how to tell before a full run.