Daylight Commons 3.8.0

Daylight Commons 3.8.0

Argument lens (under Discourse checks, its own switch, off by default; everything runs on your device). It splits a post into premises, minor premises, assumptions, conclusions and turns at the words that join them, in English, Portuguese, Spanish, French and Latin, and groups arguments into lists, chains and turns. On Summa I q.2 a.2 it finds every break a careful reader marked by hand. Roles show only once two people's annotations confirm them; until then each part reads as a segment. The 60 passages to annotate and the annotation page ship with this release.

Claim types. Each sentence is named as a question, quotation, definition, sourced statement, classification, inference, opinion, or factual claim with or without a source. On a held-out set labelled by two models, 15 of 16 were right.

Inference gaps. It marks claims about what someone else knows or intends, and conclusions that bring in words their premises lack. Held out: 4 of 5 marks right, 4 of 4 found.

Reframing. It flags a post that names the same people as the sources beneath it but puts them in an act none of those sources states, or states only to deny.

Every reading goes into the study parts tree, and each has a side-swap test that requires the same answer with political sides exchanged.

Limits. The figures come from model labels on short sentences, not people's annotations, and the argument roles have not been measured against people yet. The calibration study is preregistered and not run.

Verify with node scripts/verify-release.mjs <folder>. Key id 0e5e 1984 0565 840c 977e 0f03 6c9a c021.