Writing from Sophon
Notes on office agents, the products their work lands in, and how the company runs. Each post records its publication date when released and is corrected in place when the facts change.
-
Check a two-PDF comparison against both sources
Verify a two-PDF comparison using passages from both sources. Practice equivalence, conflicts and scoped one-source claims with fictional notes.
-
Check a PDF chart's numbers before reusing them
Check PDF chart values against their source. Practice separating labelled values, calculations and unknowns with three fictional cards and a review record.
-
Check a PDF podcast against its source, one timestamp at a time
Record an audio version and interval, then check the claim against the same PDF version, page and passage. Use a fictional example and a manual review log.
-
Check scanned PDF text before trusting an AI summary
Compare a scanned page with its recognized text. Use a fictional exercise to catch identifier, unit and approval errors before summarizing a document.
-
Adobe adds PDF tools to Claude: check the file, not just the reply
Adobe announced Acrobat tools in Claude. Use a fictional three-page exercise to check page order, requested edits and unchanged text before sharing a PDF.
-
What a template page owes a reader
Our 3,926 template pages render about 90 distinct words each. What we measured before adding more, why the sample was weighted, and what it found.
-
Our store names do not match our legal name
Our 16 September 2026 record of store and account names, with a 14 October update on PDF AI's US App Store identity. The original record is preserved below.
-
Local by default, cloud as fallback: how Cue's polish step ended up on the user's machine
Cue's dictation polish step was designed cloud-first with a local fallback. A 227-sample benchmark reversed that. What the case study records about the decision.
-
What the DeepMind case study says about Cue, and what it does not
Google DeepMind published a case study on Cue in 2026: what it measured, what runs where, how we quote it, and the claims it does not support.
-
Counting installs: why the home page says 1.41M+ and not a larger number
Google Play publishes install ranges, the App Store publishes none, and estimators publish figures nobody can check. How the home page numbers are built from that.
-
How the voice-agent benchmark landscape was built, and why it is not a leaderboard
Thirteen public benchmarks, one row each, with capability coverage and licences from first-party sources. What the dataset answers and what it refuses to.
Follow new posts with the RSS feed.