Skip to content

Sciqst as a deep-research provider — evaluation (2026-08-22)

Verdict: do not adopt as a primary deep-research provider — but there is a free corpus of 2,078 public reviews, 338 of which name a dismech disease, and those are worth a look as leads.

Sciqst reviews are short by construction (corpus median 7 references, hard maximum 23 across all 2,078), and there is no API. That rules it out as a substitute for Falcon/Edison, Asta or OpenScientist. It does not rule out mining the existing free corpus, which costs nothing and is the subject of the second half of this page.

What it is

Sciqst (ByronInsight AG, Switzerland) generates literature reviews from PubMed on a posed question. Freemium: 6 free credits, then one-time credit packs from $5. Reviews can be public or private; public ones are server-rendered at https://www.sciqst.com/reviews/<id> with title, abstract and inline PMIDs all present in the HTML, and are listed in sitemap.xml. The landing page cites MIT, Charité, Novartis, Roche, AstraZeneca and USZ as institutional users (the ToS notes these free institutional credits "[do] not signify a formal contractual relationship with the aforementioned institutions").

robots.txt opts model-training crawlers out (GPTBot, ClaudeBot, Google-Extended, CCBot) while explicitly allowing search and user-requested retrieval, with Allow: / for everything except a handful of app endpoints. The survey below is user-requested retrieval of sitemap-listed public pages.

Corpus survey (n = 2,078 public reviews)

Method: sitemap.xml → 2,078 /reviews/<id> URLs (plus 1,894 public mindmaps, surveyed separately below) → fetch each → extract og:title, meta description, and all inline PMIDs. Matching against dismech used the 2,128 kb/disorders/ names and 1,831 stubs/ labels, restricted to names ≥10 characters to suppress junk hits. Results are in sciqst_dismech_review_matches_2026_08_22.tsv (one row per matched review, with the topical orientation label used below).

The ≥10-character filter makes 338 a lower bound. It excludes 62 curated entries — counted by filename stem, which is the key the matcher used; the same cut over the entries' top-level name: field gives 60 — among them Asthma, COVID-19, Epilepsy, Glaucoma, Cholera, Dengue, Botulism, Chordoma, Glioma — i.e. exactly the common short names a clinical-question corpus is full of. The true overlap is larger than 338, which shifts the yield arithmetic below but not the verdict: the 23-reference ceiling is the binding constraint and is independent of matching.

Reference count is the real, durable limitation

References Reviews
0 2
1–5 451
6–10 1,361
11–15 234
16–25 30
26+ 0

Median 7, mean 7.4, maximum 23. No review in the entire public corpus cites more than 23 papers. Against the providers dismech already uses (median citation_count): openai 98, perplexity 51, openscientist 34, falcon/edison 31, claude_code 23, asta 20. Sciqst's maximum is the median of our weakest mainstream provider. A dismech disease entry routinely needs 30–60 distinct citations across pathophysiology, phenotypes, genetics and treatment; a 7-reference review cannot carry one.

Correction: the recency restriction is NOT general

An earlier draft of this evaluation claimed Sciqst "structurally cannot reach" pre-2022 literature, generalizing from three reviews found via web search whose PMIDs were all ≥38.8M. That generalization was wrong. Across the full corpus, by the oldest paper each review cites:

Oldest cited paper Reviews Share
pre-2000 (<11M) 51 2.5%
2000–2008 (11–20M) 215 10.4%
2008–2022 (20–35M) 494 23.8%
2022–2024 (35–38M) 252 12.1%
2024+ (≥38M) 1,064 51.3%

So about half the corpus is recency-only, and the other half reaches back — 36.6% cite something pre-2022 and 12.8% something pre-2008. (Denominator 2,076, not 2,078: two reviews cite nothing and so have no oldest paper. Both figures are recomputed from the counts rather than summed from the rounded shares above, which is where an earlier 36.7% / 12.9% came from.) Note the PMID→year boundaries are approximate and the two lower ones run 1–2 years late: 20M is nearer early 2010 than 2008, and 11M nearer 2001 than 2000; the 35M/2022 and 38M/2024 boundaries are good. Individual reviews cite papers from 1997–2004 freely. The three search-surfaced samples happened to fall in the recency-only half. Reference count, not literature age, is the binding constraint.

The free corpus: what actually overlaps with dismech

338 of the 2,078 reviews name a dismech disease — 329 matching a curated kb/disorders/ entry, 9 matching an open stubs/ entry (though most of those stubs turn out to be stale; see below).

The catch is topical. Classifying the 338 by title:

Orientation Count Share
clinical / therapeutic (management, efficacy, dosing, X vs Y, guidelines) 145 43%
other / descriptive 116 34%
mixed mechanism + clinical 46 14%
mechanism-leaning 31 9%

Sciqst's user base is evidently writing clinical practice questions — "Optimal NOAC Selection for Elderly Frail Patients with Atrial Fibrillation", "Comparative Efficacy and Safety of Apixaban Versus Rivaroxaban", "Permissive Hypercapnia". dismech is a pathophysiology knowledge base. So the effective yield is not 338 but roughly the 31 mechanism-leaning reviews plus the better half of the 46 mixed — call it 40–50 reviews of plausible interest, each carrying ≤13 references.

Best of the mechanism-leaning set. This is an editorial pick, not a strict top-10 by reference count — it skips a 12-reference Cystic Fibrosis row and an 11-reference Hepatocellular Carcinoma row whose subjects (cystic lung disease imaging, radiotherapy) are not mechanism content despite the classifier's label. Re-cut it yourself from the orientation column:

Refs dismech entry Review
13 Glomerulonephritis Mechanisms of Albuminuria in Diabetic Nephropathy
12 Metabolic Dysfunction-Associated… Incretin-Based Therapies for MASLD
11 Non-Small Cell Lung Cancer Mechanisms of Osimertinib Resistance in NSCLC
10 Optic Neuritis Mechanistic Insights and Clinical Implications of Ethambutol…
10 Brucellosis Brucellosis: Insights into Epidemiology, Pathogenesis, and Public Health Implications
9 IgA Nephropathy Advances in Understanding and Managing IgA Nephropathy
8 Polycystic Ovary Syndrome Complex Pathophysiology of PCOS
8 Hemochromatosis Genetic and Clinical Insights into Hereditary Hemochromatosis
8 ADPKD Advances in Understanding ADPKD
7 Keratoconus Advances in Understanding the Pathophysiology of Keratoconus

The 9 that hit the stub queue

These are the 9 rows that matched an open stubs/ entry, covering 7 distinct diseases. Read the table with a caveat: only 3 of those 7 are genuinely uncurated — interstitial cystitis, trigeminal neuralgia and Muckle-Wells. The other 4 diseases (5 of the 9 rows) are stale stubs whose disease has since been curated under a different name — pheochromocytoma → Pheochromocytoma_Paraganglioma.yaml, myelofibrosis → Primary_Myelofibrosis.yaml, X-linked hypophosphatemic rickets → X-Linked_Hypophosphatemia.yaml, pancreatitis → Chronic_Pancreatitis.yaml plus kb/modules/pancreatitis_acinar_autodigestion.yaml. That is the expected drift CLAUDE.md describes, cleared by just tidy-stubs --apply; it is a side-finding of this survey, not a problem with the corpus.

Refs Stub disease Review
17 interstitial cystitis Advances in Understanding and Treating Interstitial Cystitis/BPS
12 myelofibrosis Advancements in the Management of Myelofibrosis: From JAK Inhibitors…
10 trigeminal neuralgia Advances in Balloon Compression Techniques…
9 Muckle-Wells syndrome Muckle-Wells Syndrome: Current Insights and Future Directions
8 X-linked hypophosphatemic rickets Comprehensive Insights into X-linked Hypophosphatemic Rickets
7 pheochromocytoma Metastatic Pheochromocytoma: Advancements and Challenges
7 pheochromocytoma Biochemical Screening for Pheochromocytoma
3 pancreatitis The Role of Antibiotics in the Management of Pancreatitis
1 trigeminal neuralgia The A931T Variant in the TRPM7 Channel

Interstitial cystitis, trigeminal neuralgia and Muckle-Wells are the genuinely uncurated ones. The interstitial cystitis review (17 references) is the largest item in this stub subset — corpus-wide the largest dismech-relevant match is a 22-reference review, The Impact of the COVID-19 Pandemic on Clostridioides difficile Infection (CDI) Acquisition and Outcomes, against the already-curated Clostridioides difficile Infection. Note that X-Linked_Hypophosphatemia is already a worked conformer of defective_skeletal_mineralization (it declares both the phosphopenic-arm and mineralization-front nodes), not a conformer target — another instance of the stale-stub drift above.

The mindmap corpus (n = 1,865)

Sciqst also publishes 1,894 free "interactive AI mind maps" at /mindmaps/<id> (the sitemap lists 1,895 entries, one of which is the index page). These are not a second form of literature review. Surveyed the same way (1,865 parsed successfully; 29 fetch errors):

  • 90,426 total nodes, median 47 per map, range 9–182.
  • Zero citations. Corpus-wide, not a single mindmap cites a single PMID or DOI. Reviews at least give you a reference list; mindmaps give you nothing to verify against.
  • 71% are pure trees (edges = nodes − 1); the mean map has 1.01 edges beyond a spanning tree, i.e. occasional cross-links, no real graph structure.
  • The graph is embedded in the page as a var mindmapNodes JSON array and rendered client-side with vis-network, so it is trivially extractable — each node carries name, category, connections, and fixed x/y coordinates.
  • Only 73 maps name a dismech disease (72 curated, 1 stub), and those are heavily duplicated — Heart Failure ×9, Atrial Fibrillation ×8, Long COVID ×4, Ischemic Stroke ×4. Versus 338 matching reviews, the mindmaps are a much thinner overlap. Rows are in sciqst_dismech_mindmap_matches_2026_08_22.tsv.

The disqualifying property: edges are untyped

A node's category records only its depth tier (Main Topic, Subtopic, Detail, Sub_Detail, Default). Edges carry no predicate at all. The full Osteoporosis map (IflXOSqUjp10, 75 nodes) is a textbook chapter outline:

Osteoporosis
├── Definition
├── Causes ── Genetic Factors / Lifestyle Factors / Medical Conditions
│             └── Medications ── Bisphosphonates, Hormone Therapy
├── Symptoms ── Bone Pain / Fractures / Height Loss
├── Diagnosis ── Bone Density Test / X-rays / Blood Tests
├── Treatment ── Lifestyle Changes / Surgical Options
└── Prevention ── Calcium and Vitamin D / Exercise / Healthy Lifestyle

No RANKL, no osteoclast, no bone remodeling — nothing of what kb/modules/osteoporosis_bone_resorption.yaml models (remodeling imbalance → RANKL-driven osteoclastogenesis → increased resorption → impaired formation → net bone loss). It also lists bisphosphonates under "Causes", which is backwards — they are the treatment. Uncited, so nothing catches it.

The best case is a map whose prompt explicitly asked for mechanism — "What is the pathophysiology of left bundle branch block". It does contain a latent causal chain (Delayed Ventricular Activation → Asynchronous Contraction → Reduced Cardiac Output). But the same untyped edge also expresses Left Bundle Branch Block → Symptoms → Fatigue (manifestation) and → Treatment → Medications → Beta Blockers (treated-by). Causal, manifestation and therapeutic relations are indistinguishable, and the deepest tier is padding (Beta Blockers → Reduce Heart Rate, Low Salt Diet).

Only 13 of the 1,865 parsed maps have a mechanism word in the title at all.

Content is not as boilerplate as the Osteoporosis example suggests — 77.4% of the 49,351 distinct node labels appear in exactly one map, and only 10.6% of maps carry four or more of Definition/Causes/Symptoms/Diagnosis/Treatment/Prevention. The problem is not that maps are identical; it is that the relations are unlabelled and the claims are unsourced.

Contrast with a dismech pathograph

pathographs/*.json edges carry predicate, causal_link_type, and a description, and nodes carry meta.evidence with resolvable references:

{"causal_link_type": "DIRECT", "predicate": "causes",
 "source": "Inherited ADSHE gene architecture",
 "target": "Nicotinic acetylcholine receptor dysfunction",
 "description": "CHRNA4, CHRNA2, and CHRNB2 variants alter receptor function."}

A Sciqst mindmap edge is "connections": ["Asynchronous Contraction"]. There is no lossless mapping from the second into the first: the predicate and the evidence are not missing fields to be filled in, they are the entire content of a dismech edge.

Verdict on mindmaps

No usable content. Do not ingest them. An uncited, untyped concept tree is the one artifact type dismech is specifically built to be the opposite of, and converting one would mean inventing both the predicate and the evidence — which is the fabrication mode the evidence SOP exists to prevent.

Two secondary observations, neither a reason to ingest:

  • As a demand signal, resist the temptation. The 1,865 map titles are real user queries and show what clinicians actually ask about. But weighting the stub queue by that is the same mistake as the retired ranked dashboard (issue #8969) — a cheap popularity-correlated feature that tracks how common a topic is rather than whether it is worth curating.
  • The interaction model is worth stealing, though. vis-network with a 2D/3D orbit toggle, node size scaled by depth tier (30px at level 1 down to 10px at level 10), hold-1s-to-expand progressive disclosure, click-to-focus and double-click-to-search. dismech pathographs render via dagre/mermaid and get hard to read at high node counts; progressive expansion by depth is a real answer to that. That is a UI idea to file, not data to import.

Integration blockers (unchanged)

  • No API, no docs, no export. dismech generates reports through deep-research-client (pinned >=0.2.10); a just research-disorder sciqst <Disease> recipe needs a provider adapter upstream, and there is no documented interface to write one against. Capture must be manual, in the shape of the existing manual provider slug.
  • The search is not disclosed. No PubMed query, keyword set, or date range appears on a review page; the only metadata is a timestamp and "A generated literature review based on recent scholarly papers". No model is named.
  • Provenance is third-party. A public review was generated by some other user, for their question, with settings we cannot see. That is weaker provenance than any report in research/ today, all of which we ran ourselves.
  • ToS. Silent on automated access and on accuracy. It permits claiming authorship of a generated review "permitted that it does not violate any existing copyrights", forbids implying anyone other than the member created or endorsed it, and grants ByronInsight a "limited, non-exclusive, worldwide, fully paid, perpetual, irrevocable, transferable right and license to use…Your Content". Nothing blocks reading a public review and following its PMIDs; re-publishing someone else's generated review into research/ is a different act and is not clearly licensed.

Recommendation

  1. Do not register sciqst in DEEP_RESEARCH_PROVIDERS and do not commit third-party reviews into research/. The provenance and licensing are wrong for that, and a registry entry drives a legend and filter chip on the public research index. Do not ingest mindmaps in any form — see above.
  2. Treat the 338 matched reviews as a PMID lead list, nothing more. The useful content of a Sciqst review, for us, is its reference list — its own entries are PMID + title with no abstract text, so a curator must fetch the real abstract anyway. Following a lead costs one just fetch-reference call and carries no provenance debt.
  3. Start with the interstitial cystitis and Muckle-Wells reviews if anyone wants to test the value of the lead-list idea on an uncurated disease.
  4. Standard discipline applies unchanged and non-negotiably: every PMID re-fetched with just fetch-reference, every snippet verified with just count-verified-snippets, every ontology term through just validate-terms.

What would change the verdict

  1. A documented API or bulk export.
  2. Reference counts in the 25+ range — currently impossible; 23 is the observed corpus ceiling.
  3. A disclosed, controllable PubMed query. The site advertises "customizable PubMed search keywords" behind login; if that also lifts the result-count cap, the central objection here weakens. Testing it needs an account with credits, which this evaluation did not have.

Item 2 is the blocker. Items 1 and 3 would not matter if reviews stay at 7 references.