Deep research across 12 fields

Research with evidence attached.

Ask a research question. Hylē reads the literature, runs bounded analyses and returns a cited answer with figures you can check, including where the evidence runs out.

Try

Product

Two papers in. Every claim judged.

One example study, screen by screen: the papers you attach, the plan you approve, a verdict on each claim and the report you read.

Example data
Hylē Home with two example papers attached, the check set to Compare and a focus question typed in.
Plan and quote board: both papers read with one unread figure panel disclosed, standard effort for 30 credits.
Evidence workbench claims graph: eight claims connected by supporting, refuting and neutral paths, each with its judgment.
Research report verdicts at a glance: Supported, Not supported, Already known and Insufficient evidence, each with confidence.
Reader with the exact cited results passage highlighted in the paper extract, beside its parse coverage.

Method

From a question to a result you can check.

Four stages, each leaving a record you can open later.

Research / Target evidence comparisonReady

  1. 01

    Decompose

    Hylē turns your question into hypotheses, boundary conditions and template parameters, so you can see what it is about to look for before it looks.

    PlanBIO-01

    Which candidate target has the strongest convergent evidence for an oncology programme?

    • H1Directional effect across independent cohorts
    • H2Assay comparability between panels
    • ScopeHuman cohorts and organoid models

  2. 02

    Retrieve

    It searches literature and open registries and keeps the provenance of every record: where it came from, how it was accessed, and the passage that matters.

    Evidence ledger3 records

    S1Assay summarySupports

    BR-01 · Europe PMC · open access

    Across the four panels, candidate B’s relative effect was 0.82, 0.73, 0.91 and 0.99, and every uncertainty bound crossed 1.

    Opens at the exact source span

    S2Cohort notes

    BR-02 · OpenAlex

    S3Independent replication

    not found

  3. 03

    Compute

    Parameter sweeps, statistical tests and sensitivity checks run as isolated CPU jobs. Each figure records the inputs, method and version that produced it.

    Fig. 1 · Candidate B vs Ajob 1/1

    CohortnRatio (95% CI)Weight, %0.50.7511.5← Favours BFavours A →Assay panel31214.60.82(0.65–1.04)Organoid1407.50.73(0.52–1.03)Retrospective1,24042.00.91(0.81–1.02)Replication†96035.90.99(0.87–1.13)Pooled2,6521000.91(0.82–1.00)Heterogeneity: τ² = 0.002; χ² = 3.92, df = 3 (P = 0.27); I² = 23%Overall effect: Z = 1.94 (P = 0.05)† Reuses the retrospective assay; not an independent replication (S2)Cohortn · weight, %Ratio(95% CI)0.50.7511.5← Favours BFavours A →Assay panel312 · 14.60.820.65–1.04Organoid140 · 7.50.730.52–1.03Retrospective1,240 · 42.00.910.81–1.02Replication†960 · 35.90.990.87–1.13Pooled2,652 · 1000.910.82–1.00Heterogeneity: τ² = 0.002; I² = 23%; P = 0.27Overall effect: Z = 1.94 (P = 0.05)† Reuses the retrospective assay (S2)
    Random effects (DerSimonian–Laird), log scale; squares sized by weight; n = 2,652 in 4 cohorts. Example data.

  4. 04

    Synthesize

    Figures, intervals and citations come together in a study you can export and share, with the gaps listed next to the conclusion instead of hidden below it.

    Synthesis

    Insufficient evidence

    Candidate B has the most consistent directional signal S1S2, but the evidence is insufficient to select a target.

    Gap: no independent replication S3

  • Citations attached sentence by sentence
  • Unsupported claims flagged, never smoothed over
  • Every figure a reproducible CPU run

Searches open scholarly indexes and keeps the record behind every citation

01 · Sources searched

  • OpenAlexscholarly works
  • arXivpreprints
  • Europe PMClife sciences
  • Unpaywallopen-access copies

02 · Passages

  • S1Cohort recordTable 2 · effect size
  • S2Methods preprint§3.2 · protocol
  • S3Assay summaryFig. 2 · dose–response
  • S4Open-access copyp. 4 · single site only

03 · Claims

  • The effect holds in both cohortsClaim 1 · S1, S2Supported
  • Response rises with doseClaim 2 · S3Supported
  • Independently replicatedClaim 3 · S4Insufficient evidence

Product

Every answer stays tied to its source.

The paper, the passage and the run that produced each figure, one click away from the sentence that relies on them.

Synthesis

One question, a cited answer.

Hylē breaks your question into sub-questions, retrieves papers and registries, and writes a synthesis in which every sentence points to its sources.

  • Citations attached sentence by sentence
  • Agreement and disagreement side by side
  • Evidence gaps listed, not hidden
  1. S1Survival recordsPX-01 · Project file
  2. S2Dose–response assayPX-02 · Project file
  3. S3Survival in resistant tumoursnot found
HylēInsufficient evidence R-046 · synthesis

Does the inhibitor’s survival benefit hold in resistant tumours?

At 30 mg/kg the inhibitor extends median survival from 24 to 51 days, but resistant cells need about 29 times more drug for the same effect.

Survival separates by dose across three arms S1. EC₅₀ rises from 14.2 nM in parental cells to 411 nM in the resistant line S2. There are no survival data for resistant tumours, so the benefit may not carry over S3.

aSurvival by doseVehicle10 mg/kg30 mg/kg02550751000102030405060Days after start of treatmentSurvival (%)30 mg/kg vs vehicle: HR 0.21 (95% CI 0.07–0.64)Log-rank P = 0.005, 3 armsNo. at riskVehicle12128310010 mg/kg12126631030 mg/kg121299863bPotency in parental and resistant cellsParentalResistant02550751001101001,00010,000Concentration (nM, log scale)Inhibition (%)EC₅₀ 14.2 nMEC₅₀ 411 nM29-foldaSurvival by doseVehicle10 mg/kg30 mg/kg02550751000102030405060Days after start of treatmentSurvival (%)30 mg/kg vs vehicle: HR 0.21 (95% CI 0.07–0.64)Log-rank P = 0.005, 3 armsNo. at riskVehicle12128310010 mg/kg12126631030 mg/kg121299863bPotency in parental and resistant cellsParentalResistant02550751001101001,00010,000Concentration (nM, log scale)Inhibition (%)EC₅₀ 14.2 nMEC₅₀ 411 nM29-fold
Fig. 1 · a, Kaplan–Meier survival of xenograft-bearing mice, n = 12 per arm; ticks, censored; dotted drops, median survival (24, 36 and 51 days); log-rank test. b, Inhibition, mean ± s.d. of n = 3; four-parameter logistic fit with 95% band. Example data.

Evidence

Open any claim to its source span.

Select a citation to read the exact passage it came from, with the paper and its retrieval record alongside. Claims without a retrievable source are flagged.

Claim 2Resistant cells need about 29 times more drug for the same effect.S2

S2Dose–response assayPX-02 · Project file · workspaceSupports claim 2

Results · paragraph 3

Cells were exposed to ten concentrations from 0.3 nM to 10 µM in triplicate and fitted with a four-parameter logistic model. EC₅₀ was 14.2 nM (95% CI 11.3–17.8) in parental cells and 411 nM (317–534) in the resistant line, while maximal inhibition was similar in both.

Opens at the exact source span

Open the source in the example study

Figures & compute

Figures you can interrogate.

Forest plots, dose–response curves, networks and 3D structures render live, backed by bounded CPU runs you can repeat and export with the method attached.

AlphaFold model of human p53

AF-P04637-F1 · 393 residues

  • Very high > 90
  • Confident 70–90
  • Low 50–70
  • Very low < 50
Open in Figure Lab

Tube width follows secondary structure (DSSP); pLDDT colours per AlphaFold DB. Source: AlphaFold DB, CC BY 4.0.

Fields

Built for twelve fields.

Each field comes with standard research templates, the input formats they read and the figures their studies produce. Pick one to start from a worked example.

Bio

BIO-01–05

Gene and target evidence comparison, expression reanalysis, variant checks and docking triage.

Volcano plot and p53 target enrichmentSVG + Canvas · Example dataa, Volcano plot of 18,400 genes, treated vs control (n = 3 per arm); Wald test, Benjamini–Hochberg q; dashed lines |log₂ FC| = 1 and q = 0.05. b, Running enrichment score of 120 p53 target genes along the ranked list; ticks mark each gene. Example data.Volcano plot, 18,400 genes, treated vs control (n = 3 per arm); Benjamini–Hochberg q; dashed lines |log₂ FC| = 1, q = 0.05. Example data.

Templates

  1. 01Gene and target evidence comparison
  2. 02Expression data reanalysis
  3. 03Genomic variant and region check
  4. 04Existing protein structure comparison
  5. 05Molecular property and docking triage

Reads

FASTAVCFBAMh5adPDBSDF

Worked example

Targeted kinase inhibitor resistance mechanisms in EGFR T790M mutant NSCLC

Start from a worked example

Medical

MED-01–05

PICO screening, effect-size meta-analysis, diagnostic accuracy and cohort statistics.

Random-effects forest plot with RoB 2SVG · Example dataRisk ratio for heart-failure hospitalisation, SGLT2 inhibitor vs placebo, 10 trials (n = 29,312); DerSimonian–Laird random effects, 95% CI; squares sized by weight. RoB 2 domains D1–D5 and overall: + low, − some concerns, × high. Example data.Risk ratio for heart-failure hospitalisation, SGLT2 inhibitor vs placebo, 10 trials; random effects (DerSimonian–Laird), 95% CI; squares sized by weight. Example data.

Templates

  1. 01PICO literature screening
  2. 02Effect-size meta-analysis
  3. 03Diagnostic performance comparison
  4. 04Clinical trial registry comparison
  5. 05Observational cohort basics

Reads

RISBibTeXCSVXLSXJSONPDF

Worked example

Efficacy of SGLT2 inhibitors on heart-failure hospitalisation: a PICO-structured evidence review

Start from a worked example

AI / Machine learningAI / ML

AIML-01–05

Benchmark reproduction, ablations, calibration error and surrogate experiment search.

Compute scaling and compute-optimal model sizeSVG · Example dataa, Validation loss vs training compute for eight model sizes; dots, compute frontier; line, power-law fit to the frontier with 95% band, dashed past the data. b, Compute-optimal parameter count from IsoFLOP minima (n = 9 budgets), ±1 s.e.; log–log OLS fit with 95% band. Example data.Validation loss vs training compute for eight model sizes; dots, compute frontier; line, power-law fit with 95% band, dashed past the data. Example data.

Templates

  1. 01Tabular baseline comparison
  2. 02Paper reduction reproduction
  3. 03Error, calibration and group evaluation
  4. 04Search and RAG regression evaluation
  5. 05Small surrogate and active experiment candidates

Reads

CSVParquetJSONLNPZYAMLARFF

Worked example

Transformer attention sparsity scaling across depth and context length

Start from a worked example

Engineering

ENG-01–05

Design parameter sweeps, measurement calibration, vibration anomalies and thermal PDE sweeps.

Welch PSD and envelope spectrumSVG · Example dataa, Welch PSD of drive-end acceleration, healthy vs outer-race fault (12 kHz, 4,096-point Hann, 16 averages); shaft 1× = 29.95 Hz, BPFO = 107.4 Hz. b, Envelope spectrum of the faulted signal (2.4–4.8 kHz band); dashed, BPFO harmonics with ±1× sidebands. Example data.Welch PSD of drive-end acceleration, healthy vs outer-race fault (12 kHz, 16 averages); shaft 1× = 29.95 Hz, BPFO = 107.4 Hz. Example data.

Templates

  1. 01Design parameter sweep
  2. 02Measurement-based model calibration
  3. 03Vibration and signal anomaly check
  4. 04Small thermal and structural PDE comparison
  5. 05Chemical process mass-balance and reaction sweep

Reads

CSVWAVNPYMSHXDMFYAML

Worked example

Fatigue-life prediction scatter across composite layup configurations

Start from a worked example

Physical science

SCI-01–05

ODE reductions, model cross-checks, spectral peak alignment and lattice structure search.

Logistic-map bifurcation and Lyapunov exponentSVG + Canvas · Computeda, Long-run states of the logistic map xₙ₊₁ = r xₙ(1 − xₙ) for 2.8 ≤ r ≤ 4 (400 transient, 400 recorded iterates per r; shading, visit density). b, Lyapunov exponent λ(r) from 1,200 iterates; λ > 0 (shaded) marks chaos; dotted lines, period-doubling points and the period-3 window. Computed.a, Long-run states of the logistic map for 2.8 ≤ r ≤ 4. b, Lyapunov exponent λ(r); λ > 0 (shaded) marks chaos. Computed.

Templates

  1. 01ODE paper reduction reproduction
  2. 02Experimental model comparison
  3. 03Spectral peak and condition comparison
  4. 04Molecular and crystal structure exploration
  5. 05Experimental design and power exploration

Reads

CSVJCAMP-DXCIFXYZmmCIFPDF

Worked example

Stability regions of a predator-prey model under seasonal forcing

Start from a worked example

Finance & quant

FIN-01–05

Strategy robustness, macro hypotheses, factor exposure and yield-curve shifts.

Growth of $1 and underwater drawdownSVG · Example dataa, Growth of $1, daily, Jan 2007 – Jun 2025 (n = 4,825 trading days), log scale; grey bands, benchmark drawdowns beyond 20%. b, Drawdown from the running peak; labels mark each maximum drawdown. Sharpe ratio with zero risk-free rate. Example data.a, Growth of $1, daily 2007–2025, log scale. b, Drawdown from the running peak; labels mark each maximum. Example data.

Templates

  1. 01Trading strategy robustness check
  2. 02Macro and company hypothesis comparison
  3. 03Portfolio risk and shock comparison
  4. 04Bond and yield-curve sensitivity
  5. 05Event study and factor exposure

Reads

CSVParquetJSONXLSXXBRLPDF

Worked example

Tail-risk estimates for equity portfolios under regime-switching volatility

Start from a worked example

Business

BIZ-01–05

Market structure, unit economics sensitivity, cohort retention and demand elasticity.

Cohort retention triangle and CAC paybackSVG · Example dataa, Share of customers retained by month since acquisition, 12 monthly cohorts (n = 14,836 customers). b, Cumulative contribution margin per customer ÷ CAC by quarter cohort ($58 ARPU, 78% gross margin); bands, 95% interval; dashed, projected; dots, payback month. Example data.Share of customers retained by month since acquisition, 12 monthly cohorts of 2025 (n = 14,836 customers). Example data.

Templates

  1. 01Competition and market evidence comparison
  2. 02Unit economics and P&L sensitivity
  3. 03Cohort retention analysis
  4. 04A/B experiment result check
  5. 05Demand and inventory scenarios

Reads

CSVXLSXParquetJSONLHTMLURL

Worked example

Contribution-margin sensitivity to churn and acquisition cost in subscription products

Start from a worked example

Policy

POL-01–05

Evidence synthesis, budget impact, quasi-experimental checks and regional disparity.

Difference-in-differences and event studySVG · Example dataa, Maternal employment rate by quarter, subsidy vs comparison municipalities (n = 1,248; 512 treated), 95% CI; dashed, treated counterfactual from the comparison trend. b, Event-study estimates relative to the quarter before the subsidy (k = −1), two-way fixed effects, municipality-clustered 95% CI. Example data.a, Maternal employment, subsidy vs comparison municipalities (n = 1,248), 95% CI. b, Event-study estimates, clustered 95% CI. Example data.

Templates

  1. 01Policy evidence comparison
  2. 02Cost-effectiveness and budget impact
  3. 03Quasi-experimental policy effect check
  4. 04Regional access and disparity map
  5. 05Policy implementation and indicator tracking

Reads

CSVParquetGeoJSONShapefileGeoPackageRIS

Worked example

Quasi-experimental evaluation of universal childcare subsidies on maternal employment

Start from a worked example

International relationsIntl relations

IR-01–05

Actor networks, sanctions exposure, country risk indicators and treaty text comparison.

Choropleth of trade change, by country and regionSVG · Example data · Natural Earth 1:110ma, Change in goods trade with a sanctioned economy between 2021 and 2024 (%), by partner country; hatched, no data; Equal Earth projection, Natural Earth 1:110m. b, The same values by region (n = 132 of 176 countries with data): points, countries; box, interquartile range; tick and label, median; line, range. Example data.Change in goods trade with a sanctioned economy, 2021–2024 (%), by partner country; hatched, no data. Natural Earth 1:110m. Example data.

Templates

  1. 01Event and actor evidence map
  2. 02Trade and sanctions exposure scenarios
  3. 03Country risk indicator comparison
  4. 04Diplomatic document framing comparison
  5. 05International cooperation network change

Reads

CSVJSONGraphMLParquetTEI-XMLHTML

Worked example

Network effects of targeted sanctions on bilateral trade flows

Start from a worked example

Social science

SOC-01–05

Survey comparisons, interview codebooks, network topology and inter-coder agreement.

Question-wording experiment, diverging Likert barsSVG · Example dataa, Agreement with six items under two wordings (split ballot; form A vs B, n = 2,412, 1,206 per form); bars centred on the neutral midpoint, numbers in %. b, Difference in mean agreement, B − A, on the 1–5 scale with 95% CI. Example data.Agreement with four items under two wordings (split ballot, form A vs B; n = 2,412); bars centred on the neutral midpoint, numbers in %. Example data.

Templates

  1. 01Survey basics and group comparison
  2. 02Qualitative interview coding
  3. 03Social network analysis
  4. 04Social panel association comparison
  5. 05Coder agreement and measurement check

Reads

CSVSAVDTAGraphMLDOCXVTT

Worked example

Question-wording effects on reported political attitudes

Start from a worked example

Law & jurisprudenceLaw

LAW-01–05

Statute and precedent mapping, amendment diffs, citation networks and regulatory gaps.

Citation network of appellate decisionsSVG · Example dataa, Citation network of 48 decisions on data-protection liability, 2004–2024, placed by year and court; node area ∝ times cited; dark edges, citations of the most-cited decision. b, Cumulative citations of the four most-cited decisions. Example data; identifiers are not real case citations.Citation network of 48 decisions on data-protection liability, 2004–2024, by year and court; node area ∝ times cited. Example data; identifiers are not real case citations.

Templates

  1. 01Issue-based statute and case evidence map
  2. 02Clause and amendment comparison
  3. 03US case citation network check
  4. 04Contract obligation and deadline aid
  5. 05Regulatory requirement and evidence gap check

Reads

PDFHTMLXMLDOCXTXTJSON

Worked example

Citation network of appellate precedents on data-protection liability

Start from a worked example

Humanities

HUM-01–05

Argument collation, diachronic corpus use, edition collation and historical relations.

Topic prevalence and keyness, 1800–1899SVG · Example dataa, Expected prevalence of six of 12 topics (structural topic model, K = 12) by year of publication, 1,412 travel narratives, 1800–1899; line, 7-year mean; band, 95% interval. b, Keyness, 1850–99 vs 1800–49 (log₂ ratio of relative frequency). Example data.Expected prevalence of four of 12 topics (structural topic model) by year, 1,412 travel narratives, 1800–1899; band, 95% interval. Example data.

Templates

  1. 01Historical and text argument comparison
  2. 02Corpus usage and change exploration
  3. 03Edition and translation comparison
  4. 04Historical person and place relations
  5. 05OCR and transcription review aid

Reads

TEI-XMLTXTEPUBALTO-XMLGeoJSONTIFF

Worked example

Lexical shift in nineteenth-century travel writing across a digitised text corpus

Start from a worked example

Principles

Keep the limits attached.

A research tool is only useful if you can see where it could be wrong.

  1. 01

    Citations stay with the claim

    Sources remain part of the result, not a list appended at the end.

  2. 02

    Uncertainty stays visible

    When evidence is thin or conflicting, the result says so.

  3. 03

    Runs are reproducible

    Each figure records the inputs, method and version that produced it.

S1Assay summaryEurope PMC
S2Cohort notesOpenAlex
S4Methods preprintarXiv
S3Independent replicationnot found

Claim

Candidate B shows the most consistent directional signal across cohorts.

Supported

Directional signal only; target selection remains Insufficient evidence without independent replication.

Every sentence keeps its threads. The dashed one is the evidence that is missing.

Start with a clear question.

Ask it once. Open every source behind the answer.

Get started See pricing