Skip to content

Scope

  • genre: deep-research
  • language: en-GB
  • draft: content/drafts/deep-research-fidelity.md
  • created: 2026-09-15
  • corpus: 214 citekeys, digest a31f0c4b77de
  • draft digest: not recorded (run dossier stamp once the draft is ready)

Reader

The research group, at the meeting where we decide whether the next project proposal commits to a fidelity target. They know the field; what they do not know is where our own corpus contradicts itself, which is the whole reason for the report.

Covers

One question: does this corpus support treating fidelity as a tunable parameter, or as a property fixed by the decision the twin serves?

The report carries the contradiction map as a first-class section -- direct contradictions between sources, which side has the stronger evidence, the question that would resolve it, what every perspective agreed on, and the blind spot none of the searches reached.

Does not cover

A recommendation. The group decides; the report lays out what the corpus supports and how firmly. Stated here because a deep-research report is the kind of document that drifts into advocacy if nobody writes this line down.

Sources outside the corpus. Nothing is fetched from the web, by design. Where the corpus is silent the report says so rather than reaching.

Cost-benefit of fidelity in monetary terms. No paper in the corpus attempts it; naming it here keeps it from looking like an oversight.

Glossary

  • Fidelity -- how closely the model reproduces the asset on the quantities a given decision depends on. The report's central contested term: at least two cited families use it to mean a global accuracy score instead, and every use is marked with which sense is meant.
  • Tunable -- adjustable after deployment without re-deriving the model.
  • Contradiction -- two sources whose claims cannot both hold under the same reading of fidelity. A disagreement that dissolves once the senses are separated is called a terminological clash instead, and the report keeps the two apart.
  • Blind spot -- a question no perspective's searches surfaced any source for. Distinct from a gap, which is a question the corpus raises and does not answer.