{"path":"research/tools-and-analogies-as-engines-of-discovery/krauss-and-the-instrument-first-tradition-sources-as-read-2026-11-05.md","content":"# Krauss and the instrument-first tradition: sources as a helper read them (2026-11-05)\n\nThis is the report of a read-only helper (a language model, Sonnet) made on 2026-11-05 for\n[tools-and-analogies-as-engines-of-discovery.md](../tools-and-analogies-as-engines-of-discovery.md),\nkept verbatim as it came back. Every number carries its source and a read-status tag the helper\nassigned; `[memory]` dates are unverified. The helper was told what the corpus already holds and to\nskip it; its interpretation is marked as its own at the end. What Claude made of it, and what Claude\nthen read itself (the PNAS Nexus paper, the HSSC paper's methods): the main note, § 3 and § 4.\n\n---\n\nRESEARCH REPORT: Krauss, \"The Engine of Scientific Discovery\" (2026), and the instrument-first tradition\n\nRead-status tags: [full text] means I read the whole text. [part] means I read part of it. [abstract only] means only the abstract. [secondary] means a search-engine summary or a fetch-model summary, not the primary text. [memory] means a date I am giving from general knowledge with no source fetched. Treat [memory] dates as unverified. Paywalls and cookie walls blocked several primary sources, and I flag each gap.\n\n1. KRAUSS HIMSELF\n\nPapers found:\n\na) \"New tools drive scientific discovery: evidence from all nobel-prize and major non-nobel breakthroughs\", Humanities & Social Sciences Communications, 20 Apr 2026. https://www.nature.com/articles/s41599-026-06865-1 [part: about 60,000 characters of the article, retrieved via Exa. The Methods section and Extended results did not come through.] This is the core dataset paper.\n- Selection: all 533 Nobel discoveries (1901–2022). Seven textbooks that each give a \"top 100\" list of scientists or discoveries (Tiner 2022, Salter 2021, Gribbin 2008, Rogers 2009, Simmons 2000, Balchin 2014, Haven 2007) were added. \"After excluding duplicates, 302 discoveries remained—and of those, 74 won a Nobel\", leaving 228 non-Nobel, so 533+228 = 761.\n- Definition of \"major discovery\": \"a new experimental, methodological or theoretical breakthrough that marks an entirely new way to understand the world, opens new paths of inquiry and has a proven, lasting impact on science.\"\n- Coding: \"we instead identified the original discovery-making publication for each of the 761 breakthroughs, and we extracted the central method or tool they used... We manually read and extracted data from each of the over 750 discovery papers.\" The criterion for \"central\" is that the tool is \"commonly highlighted by the discoverers throughout the discovery-making publication (and often also in their nobel-prize lectures) as central to the breakthrough; it plays a direct and key role in the discovery that would otherwise not be possible; it is also later widely adopted by other researchers.\"\n- Single author. In the portion I read I found no coder count and no inter-coder reliability figure. The Methods section was not retrieved, so check it.\n- Lag by period: \"from an average of 30 years for discoveries made in the 1800s, to 21 and then 10 years for those in the first and second half of the 1900s and finally 6 years for those between 2000 and 2022.\" The Nobel-only average gap is 14 years. Abstract: \"most discoveries since 1975 are triggered within 4 years of creating the needed tool.\"\n- Immediate links: \"about one in every eight discoveries since 1975\" is an immediate method-to-discovery link, and \"about three in every ten\" arrive within two years. A search summary [secondary] gave 10% and 21% instead, so the two sets of figures disagree. Use the full-text figures.\n- Regression on lag: the gap is about 7 years longer in medicine and biology than in physics. It is about 12 years shorter for 1925–75 and 16 years shorter for 1976–2022, relative to before 1925.\n- R² = 0.92 between the year of the tool and the year of the discovery.\n- Inconsistency: Fig. 1a says 734 discoveries since 1575. Elsewhere the figure is 761.\n- Other figures: microscopes \"uncovered about 7% of science's major discoveries\". Tool-as-discovery accounts for 28% of Nobel discoveries.\n- Causation claim: \"this is as causal as we can get\", offered as a \"quasi-experimental\" design with five pieces of evidence.\n- Not found in this paper (checked by text search): \"discard\", the observation-hypothesis-test sequence, and multiples. I did not find the 1% discarded figure or the one-third figure in any primary text I read.\n- The 1% figure appears only in a search-engine summary [secondary], with the Galvani 1791 example. Its likely source is the book, which I could not read.\n\nb) \"Redefining the scientific method: as the use of sophisticated scientific methods that extend our mind\", PNAS Nexus 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC10981393/ [secondary: a fetch-model summary of the full text]\n- Same 761 discoveries (533 Nobel and 228 from the seven textbooks).\n- Coded for observation, experiment and hypothesis test. The \"classic method\" means all three.\n- Observation appears in 94%, hypothesis testing in 81%, experimentation in 75%. The classic method was applied in 71% overall (75% since 1900).\n- The abstract says 25% did not use the classic method: 6% no observation, 23% no experiment, 17% no hypothesis test. The body text implies 25% and 19%, so the abstract and body are internally inconsistent.\n- The summary reports no number of coders and no reliability check. Sole author.\n- This is the likely source of Griffin's \"about one third\": the closest thing I found is a quarter not using the full method. The one-third figure itself is unverified.\n\nc) \"New scientific fields are triggered by powerful new methods\", HSSC 2025, DOI 10.1057/s41599-025-05797-6. https://researchonline.lse.ac.uk/id/eprint/129835 [abstract only]. It covers over 350 fields. About a quarter of the fields are themselves methods.\n\nd) \"Predictable serendipity\", Scientometrics 2026. https://www.researchgate.net/publication/400366195 [part]. \"A serendipitous moment was reported in about one in six major discoveries.\" He argues that serendipitous discoveries are tool-triggered discoveries that multiple researchers can make.\n\ne) *Science of Science* (OUP 2024/25, open access) [not read]. https://academic.oup.com/book/58964. He calls it \"the new-methods-drive-science theory\" [secondary].\n\nf) The book itself. The OUP page gave no content. The PDF at personal.lse.ac.uk exceeded the fetch size limit. I did not read the book, so the 1% and one-third figures are unconfirmed there.\n\n2. CRITICS AND REVIEWS\n\nI found four reviews (listed at http://personal.lse.ac.uk/kraussa/The_Engine_of_Scientific_Discovery/). I read one in full.\n\n- Lutz Bornmann, Scientometrics, 9 Jul 2026. https://link.springer.com/article/10.1007/s11192-026-05739-2 [full text; short]. This is the only critique I could read. He calls the work \"a highly valuable contribution\" but says it simplifies \"by placing greater focus on a specific aspect\". He applies the Anna Karenina Principle: a new method is \"undoubtedly a crucial, perhaps necessary, factor, but it is not uniquely sufficient\". Breakthroughs also need replication, peer reception and a plausible theory. He notes Krauss treats funding, teams and theory as \"secondary factors\". He raises no objection on selection bias, coding or hindsight.\n- Roger Deulofeu Batllori, Metascience, 10 Jul 2026. https://link.springer.com/article/10.1007/s11016-026-01141-z [reference list only; the body was not retrievable]. The references include Chalmers 2003, \"The theory-dependence of the use of instruments in science\" (Phil. Sci. 70(3)), and Hacking 1983. That suggests the review engages the theory-ladenness of instruments, but I cannot confirm what it argues.\n- James Evans, Nature, 17 Aug 2026. https://www.nature.com/articles/d41586-026-02529-x [headline only]. It frames the book around AI as a designer of tools.\n- Tapper, The Observer, \"The science of serendipity\". The domain was blocked. Not read.\n- I found no independent check of his coding on a sample, and no historian-of-science critique. The inconsistencies in b) are my own observation from the fetch summary, not a published criticism.\n\n3. THE INSTRUMENT-FIRST TRADITION (one claim each; none of these sources was fetched except where noted)\n\n- Dyson, *Imagined Worlds* (1997). \"There are two kinds of scientific revolutions, those driven by new tools and those driven by new concepts.\" [secondary: a blog quote plus a Guardian count of about 20 tool-driven against 7 concept-driven revolutions in 500 years]. He blames Kuhn for the assumption that all revolutions are conceptual. Dyson, \"Is science mostly driven by ideas or by tools?\", Science 338:1426 (2012) [bibliographic record only]. Sources say he calls both pictures true and neither complete (Galison for tools, Kuhn for concepts). Dyson adds a count over centuries and an explicit \"both true\" position. Krauss argues the tool side nearly exclusively.\n- Price, \"Of sealing wax and string\" (1984) [memory, unverified]. Claim: instruments and technology precede and drive science (\"artifactual revolution\"). He adds the instrument-to-science cycle and a role for technology as science's own product, which Krauss's one-way arrow lacks.\n- Galison, *Image and Logic* (1997) [memory]. Claim: instrument traditions (image and logic) have their own continuity and can outlive theory change. He adds the layered, partly autonomous life of instrument cultures and their trading zones with theorists. Krauss cites Galison 1987 only as a case-study approach, per the paper's text.\n- Rheinberger, *Toward a History of Epistemic Things* (1997) [memory]. Claim: experimental systems generate \"epistemic things\" that nobody could specify beforehand. He adds that the tool and the object it reveals are co-produced.\n- Hacking, *Representing and Intervening* (1983) [memory; also cited by Batllori]. Claim: experimentation has a life of its own. His case is that instruments produce phenomena before there is theory for them.\n\n4. CASES\n\nKrauss-sourced rows come from the 2026 HSSC paper [part]. All other rows are [memory].\n\n| Tool/method | Developed | Discovery | Year | Lag (yrs) | Source |\n|---|---|---|---|---|---|\n| Hooke's 300x microscope | 1662 | Cells | 1665 | 3 | Krauss 2026 |\n| Telescope | c. 1608–09 | Jovian moons | 1610 | about 1–2 | memory |\n| Cathode ray tube (improved) | 1896 | Electron | 1897 | 1 | Krauss 2026 |\n| Interferometer | 1881 | Michelson–Morley null result | 1887 | about 6 | memory |\n| Cloud chamber | 1911 | Positron | 1932 | 21 | Krauss 2026 |\n| X-ray crystallography | 1913 | DNA double helix | 1953 | 40 | Krauss 2026 |\n| Electron microscope | 1933 | First virus images (TMV) | about 1939 | about 6 | memory |\n| Partition chromatography | 1941 | Insulin sequence (Sanger) | 1951–55 | 10–14 | memory |\n| Bubble chamber | 1952 | Quarks | 1964 | 12 | Krauss 2026 |\n| Optical tweezers | 1970 | Used in biology | 1987 | 17 (a lag case) | Krauss 2026 |\n| SDS-PAGE gel | about 1970 | Cyclin | 1982 | 12 (Hunt: \"could have been done any time\") | Krauss 2026 |\n| PDB | 1971 | AlphaFold (CASP14) | 2020 | about 49 | memory |\n| fMRI (BOLD) | 1990–92 | e.g. fusiform face area | 1997 | about 5–7 | memory |\n| Echelle spectrograph | 1993 | First exoplanet | 1995 | 2 | Krauss 2026 |\n| CRISPR | 2012 | First approved therapy | 2023 | about 11 | memory |\n| ERA5 reanalysis | 2018–19 | Pangu/GraphCast AI weather | 2022–23 | about 4 | memory |\n\nI could not source the PCR (1983) downstream case. H. pylori (1982) is a poor fit for \"tool first\": the culture was an accident of a long incubation, so it is better treated as a counterexample.\n\nControls (theory first, tool later), all [memory]:\n- Einstein 1916, gravitational waves; LIGO detection 2015: about 99 years.\n- Higgs 1964; LHC 2012: 48 years.\n- Bell 1964; Aspect 1982: 18 years.\n\n5. FAILURE SIDE\n\n- I found no quantified estimate of how many tools or methods are developed and abandoned. A search for method abandonment returned only library-science usage-trend studies, which are not relevant to this question.\n- Krauss's dataset conditions on success by construction: tools are only coded when a Nobel or top-100 discovery used them. His definition of \"central\" also requires the tool to be \"later widely adopted\", which makes failed tools unobservable.\n- Ke, Ferrara, Radicchi & Flammini, PNAS 2015, arXiv 1505.06454 [via ar5iv; fetch-model summary, so [secondary]].\n  - Data: APS 463,348 papers (1893–2009) and Web of Science 35,174,034 papers (1900–2011).\n  - Papers with negative B: 4.68% (APS) and 6.56% (WoS). Positive-B papers that never fell below half their peak yearly citations: 9.56% (APS) and 14.26% (WoS).\n  - Maximum B: 11,600 (WoS, Freundlich 1906, awakened 2002) and 1,722 (APS, Zener 1951, awakened 1994). Longest sleep: Pearson 1901, awakened 2002.\n  - About 80% of top sleepers draw at least 75% of their citations from other disciplines.\n  - The paper gives no single overall sleeping-beauty fraction. It argues they are \"far from exceptional\" [secondary].\n- van Raan 2004 (Scientometrics): I could only confirm the bibliographic record. A Nature news piece quoted him on \"6.5%\", which I could not trace to the paper. Do not cite that figure.\n- Nobody found measures sleeping beauties among methods specifically.\n\n6. DATA INFRASTRUCTURE\n\n- I found no study crediting databases with downstream Nobels. The closest is an advocate's claim by Stephen Burley (RCSB PDB director) about Nobel odds in chemistry [secondary; not a bibliometric result].\n- PDB: \"$12 billion\" replacement cost, based on US$100,000 per structure [secondary: https://cdn.rcsb.org/wwpdb/docs/about/outreach-content/2016-wwpdb-sci-data-con.pdf]. Rutgers estimated an economic impact above $5.5 billion a year. Use across 150 disciplines and integration with over 400 resources [secondary, Rutgers/news-medical search results].\n- EMBL-EBI: use value £5.5 billion per annum in the 2021 report, \"more than £3 billion higher\" than 2015–16 [secondary: https://www.embl.org/documents/wp-content/uploads/2021/10/EMBL-EBI-impact-report-2021.pdf].\n- SDSS: over 5,000 refereed papers mentioning SDSS, cited over 200,000 times (2013), and an estimate of over 8,000 papers (Plant & Hanisch 2020) [secondary: https://blog.sdss.org/2013/03/26/sdss-has-now-been-used-by-over-5000-refereed-papers/]. These measure use, not discoveries.\n- Nothing found on GenBank.\n\nMY INTERPRETATION (not sourced)\n\n- The R² of 0.92 is weak evidence: discovery years and tool years both rise over time, so the correlation is partly a trend artefact. The informative figure is the lag distribution.\n- Coding is circular at the edges. \"Central\" is defined by what the discoverer highlighted and by later adoption, so a tool-less discovery cannot appear. Single-author coding with no reliability figure leaves Griffin's success bias and hindsight points open.\n- The shrinking lag may partly reflect that the \"tool year\" is easier to date recently. It may also reflect survivorship: only tools that were later adopted get coded.\n- The controls (99, 48 and 18 years) show the tool-to-theory lag is longer and irregular in the other direction, consistent with a Dyson-style \"both are true\".\n- Still to do: read the book PDF's lag tables, Krauss's Methods section, the Batllori review and the Observer piece. These are the best leads on the 1% and one-third figures.\n"}