nucleation

LIT-988: Scientific discovery in a model-centric framework: Reproducibility, innovation, and epistemic diversity

2 min read

Properties6
StatusActive — read 2026-10-14 (NOTE-792) in full; worth reading for a formal model in which every inquiry step is a sound model comparison and the community's consensus still sits on a false model, with results that reproduce almost perfectly, because of which alternatives its research strategies put up. It proves that replication cannot change where the consensus goes. Read it knowing that the consensus is memoryless and held by no one, that its scientists neither believe nor learn, and that its "epistemic diversity optimizes" means only that an equal strategy mix is never the worst and varies least. Filed for QUESTION-047; THEORY-295 holds what it shows.
Version2
Tagsphilosophy-of-science, epistemology, social-science
Cited as “grounds” byCLAIM-470 — Inquiry is a form of specification: a question fixes which differences would count as an answer, an inquiry adds the operations that make those differences accessible, and what it finds can stabilize into a specification that constrains later inquiry; so inquiry stands to its findings as a specification to its realizations, with one difference, that an inquiry can revise what an answer could be
Cited as “paper” byNOTE-792 — Devezer et al. — Scientific discovery in a model-centric framework
Cited as “source” byTHEORY-295 — In a community whose consensus model is replaced whenever a single sound pairwise model comparison favours the challenger, which models get proposed can hold the consensus on a false model while results reproduce almost perfectly, so reproducibility need not track convergence to truth, THEORY-814 — In scientific communities that settle on a consensus model by sound model comparisons, high reproducibility of results is not evidence that the consensus is true, because which alternatives get proposed can hold a false consensus whose results reproduce almost perfectly

Graph View

Berna Devezer, Luis G. Nardin, Bert Baumgaertner and Erkan Ozge Buzbas (2019), PLOS ONE 14(5), e0216125.

Key takeaways

From the reading, NOTE-792.

  • Consensus as a process. The “global model” is the last winner of a pairwise AIC or SC comparison on fresh data. The no-replication process is a Markov chain over 14 linear models. The consensus has no memory, and no scientist holds a belief.
  • Reproducibility is causally idle. A replication re-tests the same pair on independent data, so its outcome sums out of the transition probabilities (S23). The reproducibility rate cannot affect the long-run distribution, first passage time or stickiness.
  • Reproducible and wrong. Boundary testers, who only add interactions, reach near-perfect reproducibility while spending 25% (AIC) or 48% (SC) of their time at the true model. Their correlation between reproducibility and time at the truth is −0.06, against 0.35 to 0.69 in other populations.
  • Connectedness speeds discovery. Mavericks, who propose from the whole space, reach the truth fastest. Under hard strategies a boundary-tester population takes about 1,600 steps.
  • Diversity as insurance. The equal mix is never the worst population on any property and has the lowest variability. It is not the best on time at the truth, and it fails under high noise.
  • Method matters. SC spends longer at the truth than AIC (medians 27% and 20%) but varies more.

Standing in the record

Filed on 2026-10-12 for QUESTION-047 and read on 2026-10-14 in NOTE-792. What it shows is held as THEORY-295: with every test sound, which alternatives a community puts up can hold its consensus on a false model that its replications keep confirming. This is the record’s nearest model to the question’s design. It is about a standing view, not a belief. Its “epistemic diversity” is a permanent mix of proposal strategies, distinct from the transient diversity of THEORY-148. The seeded body’s claim that diversity “sets whether the community finds the truth” was too strong: which strategies are present, and the noise level, set it.

Backlinks