{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "where-new-knowledge-comes-from",
 "title": "It Was Already There: Where New Knowledge Actually Comes From",
 "dek": "Models don’t conjure knowledge out of nothing — they read archives nobody had finished reading and walk rooms the rules imply, and what turns either into knowledge is the check that follows, never the confidence of the proposal.",
 "published": "2026-08-24",
 "series": [
  "guide"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/where-new-knowledge-comes-from/",
  "md_url": "https://research.strata2signal.com/where-new-knowledge-comes-from/index.md",
  "md_sha": "358d9e67e9f5b4acfde281d814c053acc0dfe97eca71776fb000fd46f02142b2",
  "html_sha": "3197f413541a0187f12386d1e7e7b3131fab5529b7602ddfb1d7f19c2257a9d3",
  "anchors_sha": "6aea14c5ab50864096079352b175ea1cf9ada9b41c1535f2cdb28829a94a46a3"
 },
 "short": {
  "paragraph": "Models don't conjure knowledge out of nothing - they read archives nobody had finished reading and walk rooms the rules imply, and what turns either into knowledge is the check that follows, never the confidence of the proposal.",
  "from_dek": true,
  "counts": {
   "words": 3847,
   "minutes": 17,
   "tables": 0,
   "kit": false
  },
  "bullets": [
   {
    "text": "an MIT-led team published a model trained on 2,335",
    "figure": "2,335",
    "cite": "the-first-engine-the-archive-nobody-had-read",
    "quote": "In 2020 an MIT-led team published a model trained on 2,335 molecules, each labelled by whether it stopped E.",
    "span": [
     5094,
     5100
    ]
   },
   {
    "text": "the A-Lab reported successes from 57",
    "figure": "57",
    "cite": "what-the-record-shows",
    "quote": "The abstract now reads 36 compounds from 57 targets.",
    "span": [
     16256,
     16259
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-first-engine-the-archive-nobody-had-read",
   "heading": "The first engine: the archive nobody had read",
   "level": 2,
   "span": [
    4901,
    6932
   ],
   "chunks": [
    [
     4901,
     6932
    ]
   ],
   "chars": 2031,
   "digest": "an MIT-led team trained a model on 2,335 molecules to rank compounds. the model identified halicin, a molecule that killed bacteria in a dish and mouse models, which had previously been a dropped diabetes candidate. the model inherited human priors by learning from tests performed on a library assembled for another purpose.",
   "digest_skipped": null
  },
  {
   "id": "the-second-engine-the-rooms-the-rules-imply",
   "heading": "The second engine: the rooms the rules imply",
   "level": 2,
   "span": [
    6932,
    10535
   ],
   "chunks": [
    [
     6932,
     10535
    ]
   ],
   "chars": 3603,
   "digest": "the second engine uses rules to imply unrecorded events. Kepler's laws predicted a transit of Mercury in 1631, and Dirac's equation implied the positron. AlphaGo's move 37 was a legal move that human models rated at a one in ten thousand chance. models add speed and a way to try unvisited regions.",
   "digest_skipped": null
  },
  {
   "id": "two-kinds-of-there-all-along",
   "heading": "Two kinds of there-all-along",
   "level": 2,
   "span": [
    10535,
    12461
   ],
   "chunks": [
    [
     10535,
     12461
    ]
   ],
   "chars": 1926,
   "digest": "discovery consists of two types: patterns present in the data, unmined, and possibilities implied by the rules, unvisited. one is an archive finally read and the other is a possibility space finally walked. the text notes that wins do not show how often engines propose things the world refuses.",
   "digest_skipped": null
  },
  {
   "id": "what-the-record-shows",
   "heading": "What the record shows",
   "level": 2,
   "span": [
    12461,
    17435
   ],
   "chunks": [
    [
     12461,
     17435
    ]
   ],
   "chars": 4974,
   "digest": "models produce candidates that require external verification to become knowledge. AlphaFold and halicin were verified by experiments. the text discusses how a model's confidence does not distinguish between a discovery and a hallucination, citing disputes over reported successes in GNoME and A-Lab research.",
   "digest_skipped": null
  },
  {
   "id": "where-that-leaves-us",
   "heading": "Where that leaves us",
   "level": 2,
   "span": [
    17435,
    18583
   ],
   "chunks": [
    [
     17435,
     18583
    ]
   ],
   "chars": 1148,
   "digest": "the throughput of proposing has increased, but the deciding process remains slow and external. while mathematics allows a machine to check a machine, determining if a molecule kills a bacterium requires the world to referee. the pattern and the rule exist, but truth is not the finder's call.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    18583,
    20825
   ],
   "chunks": [
    [
     18583,
     20825
    ]
   ],
   "chars": 2242,
   "digest": "discoveries are found rather than conjured. engine one finds patterns in archives, while engine two finds entries implied by rules. confidence is not a tell for truth, as an unchecked discovery is indistinguishable from a hallucination. the check, not the proposal, constitutes knowledge.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work-and-see-it-live",
   "heading": "How to check our work — and see it live",
   "level": 2,
   "span": [
    20825,
    22449
   ],
   "chunks": [
    [
     20825,
     22449
    ]
   ],
   "chars": 1624,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    22449,
    23644
   ],
   "chunks": [
    [
     22449,
     23644
    ]
   ],
   "chars": 1195,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-7d242b0d",
   "q": "how did the model identify the molecule halicin?",
   "a": "The model was trained on 2,335 molecules labeled by their ability to stop E. coli growth. When pointed at a library of 6,111 shelved compounds, it ranked a molecule previously used as a diabetes candidate near the top, despite its structural difference from traditional antibiotics.",
   "cites": [
    "the-first-engine-the-archive-nobody-had-read"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-3ba6b166",
   "q": "what distinguishes the second engine from the first?",
   "a": "While the first engine finds patterns already present in data, the second engine relies on rules that imply unrecorded entries. It identifies possibilities that are not in any existing row of a corpus but are necessitated by the underlying generating rule.",
   "cites": [
    "the-second-engine-the-rooms-the-rules-imply"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-d9022172",
   "q": "what are the two categories of discovery described?",
   "a": "Discoveries are categorized as either patterns present in the data but unmined, such as molecular shapes in a training set, or possibilities implied by rules but unvisited, such as the predicted position of a planet or the existence of a new particle.",
   "cites": [
    "two-kinds-of-there-all-along"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-17e1d83f",
   "q": "how is a candidate transformed into actual knowledge?",
   "a": "A candidate becomes knowledge through external verification. While models propose possibilities, the world acts as a referee through physical experiments, such as growing crystals or testing molecules in dishes, to confirm if the proposal matches reality.",
   "cites": [
    "what-the-record-shows"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-c7cb9a4c",
   "q": "what happens when a model's confidence is high but unverified?",
   "a": "An unchecked discovery is indistinguishable from a hallucination or an invented citation. High confidence scores do not differentiate between a valid prediction and a plausible-sounding invention if the claim has not been tested against the physical world.",
   "cites": [
    "what-the-record-shows"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-9eb0afb1",
   "q": "what role does the speed of proposing play in the discovery loop?",
   "a": "The throughput of the proposing half of the loop has increased significantly. Models can now read archives and explore possibility spaces much faster than humans, though the deciding half of the loop remains slow and external.",
   "cites": [
    "where-that-leaves-us"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a5ea66f4",
   "q": "what are the main takeaways regarding model discoveries?",
   "a": "Discoveries are found rather than conjured. Models act as engines that either mine existing archives or walk implied possibility spaces, but they only produce candidates; the final knowledge requires external verification to ensure the proposal is true.",
   "cites": [
    "what-to-take-with-you"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "three-at-the-table",
   "why": "shares ground with § The second engine: the rooms the rules imply · § The model that was proving it all along"
  },
  {
   "slug": "how-a-vision-model-sees",
   "why": "shares ground with § What the record shows · § Finding three: what it saw, against what was there — six animals, against an answer sheet"
  },
  {
   "slug": "compression-is-the-objective",
   "why": "shares ground with § Two kinds of there-all-along · § When you look closer"
  }
 ],
 "thanks": "",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 6,
  "digests_kept": 6,
  "chips_written": 7,
  "chips_kept": 7,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0
  },
  "new_noun_tokens_checked": [
   "000",
   "10",
   "111",
   "1631",
   "2",
   "335",
   "37",
   "57",
   "6",
   "A-Lab",
   "Dirac's",
   "E",
   "GNoME",
   "MIT-led",
   "Mercury"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the team's own model of human play rated the move at roughly one chance in 10,000",
    "detail": "terminal token not verbatim in the cited span"
   }
  ]
 },
 "generated_utc": "2026-09-09T01:27:31Z"
}