The prose review of #76 caught the project's foundational phrase overclaiming: texts are admitted to the calibration corpus as "published before generative models existed" / "could have written it", but GPT-3 existed from 2020 and its API was in beta from mid-2020, so a Wikipedia revision stamped 2021 was not written before generative models existed. The claim that holds is written before 2022 — before those models reached the public — which is weaker, true, and still stronger than anything a classifier can say.
#76 fixed the phrase on every surface it touched: README, SKILL, the student sheets, the committee procedure, the corpus README, the generated calibration page, and all eight report messages in both languages. What remains says the old thing:
tools/SignsOfAI.Calibration/Fetch.cs — class summary and the PLOS/Wikipedia/PELIC doc commentsDocs/Calibration/corpus.json — the 50 Wikipedia notes read "before generative models"src/SignsOfAI.Core/Rules/GenreGate.cs, Model/Finding.cs, Model/AnalysisResult.cs, Calibration/PublishedCalibration.cs — doc commentssrc/SignsOfAI.Mcp/Tools/ReportTools.cs — a tool description an MCP client shows verbatimtools/SignsOfAI.Calibration/ParaphraseReport.cs and the generated Docs/PARAPHRASE.mdtests/SignsOfAI.Core.Tests/ChatResidueTests.cs — doc comment
One sweep, one pull request, no numbers change. The MCP tool description is the only one a stranger reads without opening the repository, so it goes first.
The prose review of #76 caught the project's foundational phrase overclaiming: texts are admitted to the calibration corpus as "published before generative models existed" / "could have written it", but GPT-3 existed from 2020 and its API was in beta from mid-2020, so a Wikipedia revision stamped 2021 was not written before generative models existed. The claim that holds is written before 2022 — before those models reached the public — which is weaker, true, and still stronger than anything a classifier can say.
#76 fixed the phrase on every surface it touched: README, SKILL, the student sheets, the committee procedure, the corpus README, the generated calibration page, and all eight report messages in both languages. What remains says the old thing:
tools/SignsOfAI.Calibration/Fetch.cs— class summary and the PLOS/Wikipedia/PELIC doc commentsDocs/Calibration/corpus.json— the 50 Wikipedia notes read "before generative models"src/SignsOfAI.Core/Rules/GenreGate.cs,Model/Finding.cs,Model/AnalysisResult.cs,Calibration/PublishedCalibration.cs— doc commentssrc/SignsOfAI.Mcp/Tools/ReportTools.cs— a tool description an MCP client shows verbatimtools/SignsOfAI.Calibration/ParaphraseReport.csand the generatedDocs/PARAPHRASE.mdtests/SignsOfAI.Core.Tests/ChatResidueTests.cs— doc commentOne sweep, one pull request, no numbers change. The MCP tool description is the only one a stranger reads without opening the repository, so it goes first.