Dense results
A measured before/after on the same index: what a search returns without Brain's summary layer, and the single consolidated answer it returns with it.
The query
"cyclooxygenase inflammation cancer" — one question against one index, asked twice. The only variable is whether the summary layer is in the way of the results. Measured by Levirge's evaluation harness on BEIR SciFact, a public 5,183-document biomedical benchmark corpus, 15 Aug 2026.
Each step below shows the readable version first, with the raw payload behind a tab — the same bytes an agent receives, from a live pull on the same corpus.
Without the summary layer — eight fragments
Ranked document hits. Each is a real source; none is an answer. Two are only loosely on the subject, so the model pays to read them and then pays again to discard them — every time anyone asks.
1 Cyclooxygenase-Dependent Tumor Growth through Evasion of Immunity 2 Tumor-associated macrophages and high COX-2 → poor prognosis 3 COX-2 inhibitors enable repeated vaccinia virus administration 4 PI3Ks qualify as drug targets for cancer, inflammation … (loosely related) 5 Osteoarthritic cartilage degradation indicators … (loosely related) 6 15-Hydroxyprostaglandin dehydrogenase down-regulated in colorectal cancer 7 Omega-3-derived lipid mediators 8 COPD, neutrophils, cigarette smoke and COX
Thousands of tokens, two of them off the point, and the synthesis is still the model's job.
A live pull for "interleukin-1b inflammasome" on the same corpus. These are the rows a plain vector store returns — documents and facts, no synthesis:
[
{ "id": "700b960dd0a3aad0", "kind": "doc", "score": 0.992,
"title": "Human Monocytes Engage Alternative Inflammasome Pathway" },
{ "id": "df399593abc4edcb", "kind": "doc", "score": 0.989,
"title": "Inflammasome-induced IL-1β secretion in microglia" },
{ "id": "4e6192a75e2477bc", "kind": "fact", "score": 0.939,
"title": "[fact] Inflammasomes are multiprotein complexes that li…" },
{ "id": "86ac2fbf338b865b", "kind": "fact", "score": 0.935,
"title": "[fact] Cytokines downstream of inflammasome activation,…" }
]
Four rows to read and reconcile before answering anything.
With the summary layer — one consolidated answer
The top result is a single entity summary, consolidated from 11 source documents, each claim carrying the source it came from. The agent cites this and stops.
Cyclooxygenase-2 (Cox-2) — an enzyme implicated in inflammatory and pathological processes: viral clearance, lipid-mediator generation, and cancer development. - Viral clearance: Cox-2 inhibitors circumvent viral-clearance limits via neutralising antibodies [cited] - Lipid mediators: bioactive signals (15-epi-LX, 18R-HEPE) via NSAIDs + omega-3 [cited] - Renal vasoconstriction [cited] - Cancer: survivin correlation [cited] - COPD airways [cited]
The summary also carries a small entity-relation graph — cox-2 → survivin, PGE-2, arachidonic-acid, aspirin — so the next question about a neighbouring subject starts from the relations, not another eight-document read.
The same live pull as the previous step, unfiltered. Row 1 is the dense summary, above every raw document and fact — and its kind says what it is, so a model can prefer it deliberately rather than by luck:
[
{ "id": "0a013be2e4b93400", "kind": "summary", "score": 0.997,
"title": "[rule] interleukin-1b is a proinflammatory cytokine who…" },
{ "id": "700b960dd0a3aad0", "kind": "doc", "score": 0.992,
"title": "Human Monocytes Engage Alternative Inflammasome Pathway" },
{ "id": "df399593abc4edcb", "kind": "doc", "score": 0.989,
"title": "Inflammasome-induced IL-1β secretion in microglia" },
{ "id": "4e6192a75e2477bc", "kind": "fact", "score": 0.939,
"title": "[fact] Inflammasomes are multiprotein complexes that li…" },
{ "id": "86ac2fbf338b865b", "kind": "fact", "score": 0.935,
"title": "[fact] Cytokines downstream of inflammasome activation,…" }
]
Row order tells the story: the answer, then the primary sources it was built from, then supporting facts.
Document 0a013be2e4b93400 — row 1 in full, consolidated from 7 sources, every claim carrying the id it came from:
interleukin-1b is a proinflammatory cytokine whose bioactivity is
controlled by activation of the inflammasome.
Major capabilities and subsystems
- proinflammatory cytokine, bioactivity controlled by inflammasome
activation [700b960dd0a3aad0]
- generated through NLRP3 inflammasome activation in Myelodysplastic
Syndromes, driving clonal expansion and pyroptotic death
[1a48fcf7bb8cb4be]
- produced in human monocytes in response to LPS via an alternative
inflammasome pathway [700b960dd0a3aad0]
- controlled in ILC2s in airway inflammation, interacting with IL-4
and IL-12 to steer cell fate [b82d5fcd36bb8472, 90250f9be040c89e]
Invariants / gotchas
- the LPS "alternative inflammasome" uses NLRP3-ASC-caspase-1 but is
devoid of pyroptosome formation, pyroptosis and K⁺-efflux
dependency [df399593abc4edcb]
- microglial IL-1β secretion is only partially caspase-1/4/5-dependent
vs hematopoietic macrophages [df399593abc4edcb]
The "invariants / gotchas" section is the part worth noticing: it carries the exceptions, which is where a synthesis usually goes wrong.
Every entity carries a typed relations graph, queryable on its own — t-cells has 85 edges, breast-cancer 57. Some summaries, this one included, also inline a machine-readable block of them:
{ "relations": [
{ "target": "caspase-1", "type": "part_of", "confidence": 0.9 },
{ "target": "lipopolysaccharide","type": "integrates_with", "confidence": 0.9 }
] }
The next question about a neighbouring subject starts from the graph rather than another full read. The inline block is the occasional bonus; the graph itself is on every entity.
What was measured
Consolidation only helps if the dense answer actually surfaces — and only if it doesn't cost ranking quality. Both were measured on the same corpus and query set.
summaries in top-10 35% — they surface, not buried rank-1 for 25% of queries nDCG@10 0.71 — BM25 baseline 0.665 ● consolidation does not cost retrieval quality
Where this came from
Both result lists come from the same index and the same query. The effect is a property of consolidation, not of this corpus: how dense your answers get depends on how many sources say something about the same subject.
corpus BEIR SciFact — public biomedical benchmark, 5,183 documents variable the summary layer — same index, same query, both runs raw tabs a live pull for "interleukin-1b inflammasome" on the same corpus, 15 Aug 2026 - a second query, because it is the one we hold verbatim payloads for payloads titles excerpted; summary text verbatim apart from one unrelated line - a consolidation defect we logged rather than quietly leave in the example harness Levirge evaluation harness, 15 Aug 2026