Four kinds of claim
A dossier that flattens what a person said, what others reported, what the evidence shows, and what the model guesses into one voice is worse than useless. Soulscrape keeps them separate:
- fact — documented in the evidence, cited to its source.
- stated_belief — the subject's own stated position, cited.
- pattern — a revealed regularity the evidence supports but the subject may not have stated.
- speculation — a labeled guess, kept only when it is honest to include and marked as such.
Confidence describes support within the evidence examined. It never upgrades a guess into a finding, and contradictions are preserved rather than resolved.
Authorization is a property of the ask
Possessing data is not authorization. Messages, exports, and public information arrive with a context, and the skill writes the intended use, audience, and missing evidence into a question packet before reading. Private corpora stay in your agent environment.
Public web research is off by default. When you turn it on — by asking, or by naming a URL — the skill follows your instructions on sources, time window, and depth, records every finding with its URL, access date, passage, and identity binding, and stops when the questions are answered.
What a dossier is not
- Not the person, and not the subject's own page — a dated interpretation published by someone else.
- Not a voice imitator, a character assessment, or a basis for consequential decisions about a real person.
- Not finished — it is revisable, and it says what it did not cover.
A bounded reasoning proxy is possible only when the subject has authorized that use. Everything else — a collaboration guide, a research brief, a public index — works because it stays honest about what it is.