Private genomics
Lens readout drag the lensSynthetic · im
The genome here is synthetic: sample “im”.
Strand: GRCh37 chromosomes 1–22 and X (UCSC hg19, 2009), stretched around the six panel sites.
01 · The problem
Syntelea's three-gene result depends on six positions. A consumer DNA file holds hundreds of thousands.
The file is read on your device and never uploaded.
A recipient can check the result came from your file under a named CPIC rule.
We test what DNA models like Evo 2 encode, apart from the CPIC rules.
2,402 records
Synthetic · im02 · Clopidogrel example
Clopidogrel, given after a heart attack or stent, is activated in the liver, mainly by CYP2C19.1 Both copies of the gene affect how much active drug a dose makes.
After ACS or PCI
Standard dosing
Clopidogrel 75 mg/day for everyone
Genotype-guided
CPIC guidance for clinicians1
Select a row. Meters are schematic. *1/*17 and *17/*17 follow the *1/*1 advice; *2/*17 is intermediate.
03 · From file to proof
Steps 1 to 4 run on your device; step 5 runs wherever the proof is checked.
Synthetic · im
A five-step animated figure in isometric pixel blocks, using the synthetic sample genome “im”.
Step 1, Read: the raw text file is parsed in the browser. A rule moves across a flat slab and fills it with 2,402 small squares, one per record, in genome order. Nothing is uploaded.
Step 2, Classify: six orange blocks rise out of the slab, one for each panel position, and stack into three towers labelled CYP2C19, CYP2C9 and SLCO1B1. CPIC rules give CYP2C19 intermediate metabolizer, CYP2C9 intermediate metabolizer, SLCO1B1 decreased function.
Step 3, Commit: every record is salted and hashed, and the hashes fold in pairs into a stepped pyramid. The orange block on top is the Merkle root, computed in this tab, which changes on every visit because the salt is random.
Step 4, Prove: the root is sealed into a dark chip, the Jolt proof, about 85 KB and zero-knowledge. The slab and pyramid below are marked as private input.
Step 5, Verify: the chip slides onto a verifier's card on another device. The card lists the three phenotypes, each checked, the commitment, and the genotypes received: none. A stamp marks the proof as valid after about half a second.
Synthetic sample “im”; the salt is new on each visit. Proof size and times: Syntelea, Apple Silicon Mac, 16 GB. Rules: CPIC 2022 (CYP2C19, SLCO1B1) and 2020 (CYP2C9), cpicpgx.org.
04 · Who it affects
Estimates from CPIC, pooled from published genotyping studies. In 7 of the 9 groups, fewer than half are CYP2C19 normal metabolizers.
of 49,790 UK Biobank participants may have an atypical response to at least one drug, across 14 CPIC pharmacogenes.
CYP2C19 · CPIC estimate
Select a bar for exact shares and the 100 people below.
Gene
Estimate
Phenotype
Out of 100 people
A group label describes ancestry; each person's result comes from their own genotype.
Shares assume Hardy–Weinberg equilibrium; *1 is the remainder, so it is overstated where few alleles were studied. "Likely" and "possible" phenotypes count with their parent; CYP2C9 scores 1.5–1 are intermediate, 0.5–0 poor. "Not normal" excludes indeterminate. The panel-only view is our derivation from CPIC allele frequencies, not a CPIC figure. People-counts are rounded to total 100.
05 · What comes next
Jolt proofs now cover variants, scan scores, phenotypes and cohort counts. The other steps are in development or planned.
Counts read from the proofs run for this site, the proof manifest and the rule registries.
06 · Who it is for
If you work in one of them, we would like to try it with you.
Check a proven CPIC phenotype in a browser.
Screen participants by phenotype, such as CYP2C19 metabolizer status.
Study what Evo 2 encodes about the six panel variants.
Run Jolt in a browser tab with the published setup.
Browser prover
WebAssembly, one thread
Native prover
Optional, multi-core
Proof size
Sent to the verifier
To verify
In a browser
The three-gene proof · Apple Silicon Mac, 16 GB RAM · Jolt rev 5aa807b · zero-knowledge mode
07 · What is published
Download any file here and check it against its published hash. The Rust source is not public yet.
Check a download
Published
Received
File
Only the copy in this tab changes.
Published hashes: assets/zk/manifest.json, assets/zk/core/BUILD.json, data/dms/PROVENANCE.json. Each file is downloaded from this site and hashed with SHA-256 in your browser.
The proof made on Prove and checked on Verify.
Proven natively and checked in a browser.
Public repository planned.
PlannedPublic research data and synthetic test genomes.
Tell us where Syntelea falls short, or which question to try next.
contact@syntelea.com · Please do not send genetic data or health information by email.
How it works
Uploading sends every genotype in the file. Syntelea sends three phenotypes and a proof.
Two ways to share a result
≈600,000: positions in a typical consumer array file. Proof size and times measured on an Apple Silicon Mac, 16 GB.
What each party sees
Pick a recipient to see which records reach them and in what form.
23andMe layout · GRCh37
Synthetic · imRecords near the panel sites
| # | rsidchr:pos | Genotype | Panel site | Leaf hash |
|---|
Whole file · one square per record
Click to inspect
Commitment root
computing…
They hold
Taken on trust
Leaf = SHA-256(0x00 ‖ salt ‖ record), with a 16-byte salt per record.
The genes
Variants are defined in PharmVar; each rule comes from a CPIC guideline.
Gene models are schematic, not to scale. CYP3A5, TPMT, NUDT15 and DPYD are proven on the Platform pages. CYP2D6 is not covered: arrays cannot count gene deletions or duplications.
Prove
The file is read in this tab and never uploaded.
One proof opens files up to 262,144 records, like the samples here. A larger file, such as a full 23andMe or AncestryDNA download, gets one proof per gene and one for the rules.
Synthetic genomes
Reveals phenotypes and a commitment, no genotypes.
Opens the six panel genotypes for anyone to re-check.
Root
A proof opens each panel site with one hash per level of the Merkle tree, so larger files need more bytes.
Measured natively on an Apple M4.
Openings are counted in this browser with the proof's own encoding. Limits: assets/jolt/config.json, assets/zk/extract_panel@1/config.json. Run: assets/zk/benchmarks.json.
Not a diagnosis. Ask a clinician before changing a medication.
Verify
About half a second in a browser. A zero-knowledge proof carries no genotype.
Edits one phenotype and its digest, then checks again.
Not a diagnosis. Ask a clinician before changing a medication.
| Variant | EVEE score | CPIC function |
|---|
Explore
Goodfire's EVEE scores variants from the internal representations of Evo 2, the Arc Institute's DNA model. Syntelea's calls use CPIC rules; these scores are research evidence only.
CYP2C19 *3 score
A stop codon, flagged at 50%, matching CPIC.
CYP2C19 *2 score
ClinVar: likely benign. CPIC: no function.
CYP2C9 order inverted
*2 (decreased) scores above *3 (no function).
Variant detail
Variant
Loading…
Locus · schematic, not to scale
Top five probe shifts
Probe readouts from 0 to 1. They describe the model, not a measured effect.
Panel variants
EVEE is trained on ClinVar disease labels, so its scores are not CPIC function calls.
Lab measurements
Deep mutational scans measure what each variant does to the enzyme itself. Scores on held-out variants are proven against them on the Platform pages.
CYP2C9 · two MaveDB scans
Step 1 of 4
Loading the scans…
MaveDB 00000095-a-1 (activity, Click-seq in yeast) and 00000095-b-1 (abundance, VAMP-seq in human cells), Amorosi et al. 2021, CC0. Both are scaled so the median stop codon scores 0 and the median silent change 1. Measured in cells, not in patients; drawn in this browser.
Planned experiments
A–D are planned and need model access; only public and synthetic sequences would be sent.
Platform
Items not yet inside a Jolt proof are marked In development or Planned.
Eight steps from sequencing reads to lab experiments
Select a step
Replay01 · Reads
CPU only
Sequencing reads (BAM) become variant calls (VCF), with no GPU.
02 · DNA
Arc Institute
A genomic language model scores how well a change fits its sequence context.
03 · Variant
Checked against PharmCAT
Genotypes at fixed panel sites, mapped to star alleles.
04 · Protein
MaveDB · CYP2C9, CYP2C19
Scans measure enzyme activity and abundance for thousands of variants.
05 · Cell
Public data, no model run
Single-cell data shows how cell lines respond to drugs.
06 · Patient
CPIC rules map a diplotype to a phenotype.
07 · Cohort
Jolt
Participants prove phenotypes on their own devices; an aggregator counts the proofs.
08 · Test
A prediction is hashed and timestamped before anyone measures it.
Panel sites
Each site must match PharmCAT's position, rsID and bases, and lift from GRCh37 to the same GRCh38 position.
Our sites on PharmCAT's positions, one gene per row
Read at load from data/pharmcat/sitesets-crosscheck.json and PharmCAT 3.4.0 positions. T0 sites only.
Measurements
Every statement, proved once natively and checked again in a browser. Proving time follows the trace; the proof stays under 90 KB.
One proof per statement
Two labels
A Jolt proof shows that a rule ran as written. It does not make the biology true.
Examples
Select an exampleSelect a cell
| Biology ↓ Cryptography → | ProvenProvenJolt proof | SignedSignedLab, device or user signature | AssumedAssumedStated, not backed |
|---|---|---|---|
| T0 · RulePublished guideline | No example | ||
| T1 · MeasurementLab or instrument data | |||
| T2 · Model + controlsModel with baseline and held-out test | No example | ||
| T3 · HypothesisInterpretation, not validated |
Example
CPIC phenotype, 3 genes
Why T0 · Rule
Versioned CPIC rules: CYP2C19 and SLCO1B1 from 2022, CYP2C9 from 2020.
Why Proven
Jolt proves the rule ran on the committed genotypes. The verifier sees phenotypes, never a genotype.
Example
Warfarin dose band
Why T0 · Rule
The dose band under CPIC 2017 and the IWPC formula (Klein et al. 2009), with VKORC1 and CYP4F2 as inputs.
Why Proven
The band is computed inside Jolt with all clinical inputs private.
IWPC (Klein et al.) 2009, NEJM · Johnson et al. 2017, CPIC warfarin guideline
For comparison
Signed lab report
Why T0 · Rule
The same kind of rule, so the same tier.
Why Signed
A signature names who vouches for the result. It does not show the rule ran correctly, and the reader sees the genotype.
Example
Genotypes from the committed file
Why T1 · Measurement
A genotype is a measurement.
Why Proven
Merkle paths prove each panel genotype sits in the committed file. Whether the file is accurate is a separate claim.
Example
Signed commitment
Why T1 · Measurement
A raw DNA file is measurement data.
Why Signed
A P-256 signature over the commitment, checked inside the proof, moves the input from Assumed to Signed. The demo key belongs to a demo lab, not a laboratory.
Limit
The file matches your DNA
Why T1 · Measurement
The file is measurement data.
Why Assumed
Syntelea cannot check whose file it is or how accurate it is. The limits ledger in About lists this.
Example
Variant calls from reads
Why T1 · Measurement
Calls made by samtools and bcftools from sequencing reads are measurements.
Why Assumed
Calling runs on a CPU, outside the proof. The calls are committed, and everything after them can be proven.
Example
Score on a committed test set
Why T2 · Model + controls
A model scored on held-out data is what T2 means.
Why Proven
The test set is committed first, and each score is proven over that commitment. Its timestamp stays in a local log until a public one is approved.
Research
EVEE pathogenicity score
Why T2 · Model + controls
Evaluated on ClinVar disease labels, not on drug-response function.
Why Assumed
Explore shows a published snapshot. The run itself is not proven.
Example
Reading 64 SAE features
Why T3 · Hypothesis
What a feature means is an interpretation.
Why Proven
About 262 thousand multiply-adds per position (4,096 × 64) over committed activations; the model run itself stays Assumed. The activations are synthetic until the hosted model runs.
Example
Preregistered prediction
Why T3 · Hypothesis
A prediction is untested by definition.
Why Signed
Its hash is timestamped before results exist. That shows it was fixed in advance, not that it is right.
Limit
What an SAE feature means
Why T3 · Hypothesis
Only causal tests, ablating or steering the feature, could lift it to T2.
Why Assumed
No cryptography can make a meaning true.
Empty cells: no example yet.
Platform · Modules
Open a module for its plan.
Items per module, by status
One square per item in the plans below. An item is built when its proof statement was proven and checked for this page; read at load from assets/data/platform/runs.json.
More genes, medicine interactions, dosing.
Three-gene proof. CYP2C19, CYP2C9 and SLCO1B1 phenotypes under CPIC, proven with Jolt.
Browser and native proving. The same proof made in a browser tab or by the local prover.
Verifier. Checks a proof in the browser; no genotype is received.
Rule registry. Hashed, versioned CPIC tables. A CPIC update means a new table hash and a new proof from the same commitment.
More genes. CYP3A5, TPMT, NUDT15 and DPYD, plus more CYP2C19 and CYP2C9 alleles. Coverage is read from each file.
Medicine interactions. A private medication list adjusts the phenotype; only the result and the rule hash are public.
Warfarin dose band. CPIC 2017 + IWPC in fixed-point arithmetic, with VKORC1 and CYP4F2 as inputs; only the band is public.
PharmVar definitions. Allele definitions beyond CPIC's; needs a free API key, not set up yet.
The demo computes in this browser; its statements are proven with Jolt. Synthetic inputs, not for prescribing or dosing.
Commitment, share links, signed inputs.
Closes two gaps in the limits ledger: whether the input is genuine, and who holds the proof.
Salted Merkle commitment. Every record of the file is committed; a proof opens only the panel sites.
Share links. A link that opens the six panel sites, each with its Merkle path to the root. No proof runs.
Lab signature. A P-256 signature over the commitment, checked inside the proof. The demo key belongs to a demo lab, not a laboratory.
Holder binding. A passkey binds the proof to its holder, who signs a fresh challenge on request.
Expiry and revocation. A validity period and an entry on a public status list, with no personal data.
Build conversion. GRCh37 positions lifted to GRCh38 inside the proof, against a hashed chain file.
The demo computes in this browser; the lab-signature check is proven with Jolt.
Prove one fact, nothing more.
Trial eligibility. Meets a study's genetic criteria, with no genotype revealed. A missing call counts as unknown.
Proof of absence. None of the variants in a published site set, always shown with that set's hash.
Polygenic score band. PGS Catalog weights, proven natively; only the band is public. Consumer array files cover too few of the score's sites and are refused.
Carrier match. Two people learn only whether they share a variant, by private set intersection beside the proofs.
Synthetic genomes. The demo computes in this browser; the eligibility check is proven with Jolt.
Population counts, no central database.
Contribution proof. One participant's phenotype proof with a nullifier, so each person is counted once and not named.
Proven aggregation. Each contribution proof is checked; privacy noise is proven, and nothing is published below a minimum count.
Aggregator service. Collects contributions for a regional PGx atlas; the demo cohort is synthetic.
Federated simulation. Three NVIDIA FLARE nodes on one machine, each contribution with a Jolt proof. A simulation, not a hospital setup.
Before this
Provenance (M2), or forged genomes get counted.
Synthetic participants. One cohort is proven end to end on that page.
A function axis, tested on scans.
Explore showed that a disease-risk score does not measure drug-response function. Explore+ tests model scores against measured function.
Baselines on held-out scans. Held-out sets from the CYP2C9 and CYP2C19 scans are committed; baseline scores over them are proven.
Held-out probes. Probes scored on positions and a gene they never saw in training; each score is proven.
Function maps. Deep mutational scans from MaveDB (CC0): CYP2C9 activity and abundance, CYP2C19 abundance. Drawn in the browser, outside any proof.
EVEE comparison. EVEE's record calls CYP2C19 *2 synonymous and carries ClinVar's "likely benign", though CPIC calls it no function.
Clinical calibration. Scores mapped to PP3/BP4 evidence strengths (Pejaver et al. 2022).
Evo 2 activations. From the hosted 7B model. The client is built and tested; the call needs an API key, not set up yet.
Causal tests. Ablating or steering SAE features, the only way to lift T3 to T2.
Before this
Preregistration (M9), so the criteria are fixed first.
Public cell data, no model run.
Shows public cell data downstream of a gene, with the evidence tier at each step.
Atlas slices. Single shards of Tahoe-100M (CC0), one drug and cell line at a time.
Evidence card. Shows where the tier drops from gene to cell, with no model run claimed.
Before this
Preregistration (M9), so the criteria are fixed first.
Proven read-outs, scores and small models.
Separates what can be proven about a model from what can only be stated.
Proven SAE read-out. 64 chosen features over committed activations. The model run stays Assumed; the activations are synthetic until the hosted call runs.
Proven leaderboard. A test set committed before any entry; each score proven over it. The timestamp is local until a public one is approved.
Proven small models. Two fixed-point scan models, each with a model card and a proof that binds its weights.
Proven scores and read-outs on that page; the leaderboard demo there is synthetic.
Predictions fixed before the measurement.
Ties a model's predictions to the measurements that test them.
Preregistered predictions. Hashed and timestamped before measurement; a later measurement moves a claim from T3 to T1, with the record of when it was fixed.
Syntelea does not generate, design or export biological sequences.
Timestamps, agent tools, a test set.
Makes reproducibility checkable, for people and for AI agents.
Timestamps. Hypotheses hashed and timestamped before results exist. Local log today; no public stamp yet.
syntelea-mcp. Read-only tools for AI agents; no tool reaches a private genome or a prover key.
SynteleaBench-PGx. GeT-RM consensus calls as test vectors; our accuracy over 1,981 calls from 281 samples is proven.
Clinical formats. FHIR Genomics Reporting and CDS Hooks output, each marked research, not a diagnosis.
Public log entries. An entry in a public transparency log is permanent, so none has been written yet.
Public repository. Code, data licences and the sequence policy in one public repository.
Agent skills. The same tools as agent skills, built with evaluation cases. Submitting them needs a public repository first.
Figures with proofs. Each figure carries its data, code and run hashes.
SHA-256 in this browser; nothing is written to a public log.
From sequencing reads to a proof, no GPU.
Adds the steps before the DNA file. The browser path stays as it is.
Reads to VCF without a GPU. A CYP2C19 region of a public BAM, fetched and called without a GPU; the calls match that sample's GeT-RM consensus.
VCF to proof. Calls are normalised, committed and proven: seven genes from one GeT-RM genome.
Reference files. Genome in a Bottle and GeT-RM samples, each file stored with its source, checksum and licence.
The reads run in M10 is read from assets/data/platform/reads.json. No module is a clinical service.
Platform · A
M1 adds medicines and dose bands, M2 adds lab and holder signatures, M3 shares one yes/no answer, and M4 counts phenotypes across a cohort.
Module M1 · Registry
Where each gene sits, which sites its CPIC rule table reads, and that table's hash.
GRCh38
ReplaySelected gene
–
Rule table version
A new table changes only its own hash. The site set and the genome commitment stay the same.
CPIC guidelines
Computed in this browser from the published site-set and rule-table bytes. The Jolt proof that binds a table hash: .
assets/sitesets (SYSS v1). Coordinates from the Ensembl REST snapshot, alleles from the CPIC API, both fetched 2026-10-22.assets/rulesets (SYRT v1), built from the CPIC API snapshot of 2026-10-22. The v0.2 tables reproduce the rules of prove_panels@0.2 on all 16,275 genotype combinations./v1/guideline, fetched 2026-10-22.Module M1 · Layer A
Versioned CPIC tables, private medicine lists and warfarin dose bands.
Synthetic genome
Gene
Opens only the panel leaves, never the file.
Private medicine list
CYP2C9 · from the genome
–
VKORC1 −1639G>A · rs9923231
Chosen by hand: the synthetic samples only carry the six sites of the three-gene panel.
Self-reported race, as IWPC coded it
A coarse social category in the 2009 model, standing in for unmeasured factors. Not ancestry, not a gene.
From the medicine list
Shared with tab 02.
Rule tables · SHA-256 in this browser
ReplaySynthetic · imLoading the published tables…
Recompute from the same commitment
Press RecomputeOnce in the proof
Research demo on synthetic inputs. Not for prescribing or dosing.
Every term in integers · coefficients × 10⁻⁴
√(mg/week) scaleFixed point vs floating point
Loading the measurements…
Module M2 · Layer A
Lab signature, holder check, expiry and revocation.
The lab signs the file's Merkle root, so one changed letter breaks the signature.
Synthetic · im
Signed message · 93 bytes
The holder's phone signs a fresh nonce, so a copied bundle fails.
–waitingwaitingThe lab signs a public list, one bit per bundle. Verifiers fetch all of it, so the lab cannot tell which bundle is checked.
Valid for 180 days
Status list · 128 bits
List signature –
Computed in this browser. The lab-signature check is proven with Jolt; the holder and validity checks are in development. Keys are made in this tab; the lab is a demo key, not a laboratory. Data: synthetic genome im and the engine's salted SHA-256 Merkle tree (assets/js/engine.js); message layout from core/crates/syntelea-core/src/v3/attest.rs.
Module M3 · Layer A
The device answers one yes/no question about the genome.
The sponsor asks
Stays on the device
Sent to the sponsor
Proof of absence · three published site sets
Synthetic genome
Loading the site-set registry…
M3 · status
Module M4 · Layer A
Phenotype counts across a cohort, with no central genome database.
Contributions
Privacy budget · GeT-RM CYP2C19
Published count and its 95% noise range
Loading the GeT-RM table…
Chance a bucket is published, by ε
One cohort, proved end to end · synthetic participants
Limit. Someone with two secrets counts twice. One secret per sample is an assumption, and the ledger lists it as one.
M4 · status
Platform · B
M5 sets measured CYP2C9 and CYP2C19 function beside model scores. M6 follows CYP2C19 *2 from the CPIC rule toward cell data, tier by tier.
Module M5 · Layer B
Deep mutational scans as ground truth for model scores.
Every amino-acid change, measured
Score set
Panel alleles
Whole protein
Window · measured · drag the frame above to move it
Arrow keys move between variants; the readout below follows.
MaveDB score sets 95-a-1, 95-b-1 and 1199-a-1 (CC0), drawn in this browser; not part of any proof. 1 = wild type, 0 = a typical stop codon. Measured in yeast or human cell lines, not in patients.
Pathogenicity next to measured function
EVEE scores disease risk learned from ClinVar labels; CPIC and the scans describe the enzyme. EVEE snapshot; scans from MaveDB.
Work plan
Module M6 · Layer B
One allele followed toward public cell data, with the evidence tier at each step.
CYP2C19 *2 and clopidogrel
Work plan
No output here classifies a genome or advises on a medicine.
Platform · C
A sparse-autoencoder (SAE) read-out and a benchmark score are small enough to prove with Jolt. The model run itself stays outside the proof.
Module M7 · Layer C
The activations are committed; the proof covers the read-out and the rule.
The read-out takes one committed activation vector and computes the chosen features: 4,096 multiply-adds each.1,2
Cycles per proof · log scale
Native onlyMeasured rows: native proofs, median of three runs; panel and rule rows on a real 23andMe file, read-out rows on synthetic activations. Not counted for the read-out: the SAE weights, built into the guest at 8 bits (256 KiB), and the top-k step over all 32,768 features, which a read-out of chosen features does not prove.1,3
Hidden test sets ask teams to trust the organiser.4 Here a hash of the set is published before any submission.
Organiser · 48 hidden labels
GeT-RM · CYP2C19Commitment · SHA-256 of set + salt
…
Submissions · three teams
SyntheticReveal · anyone re-checks
Labels: CYP2C19 no-function carriers by CPIC allele function, among GeT-RM reference samples with a consensus call; 48 drawn at random in this browser. Team predictions are synthetic. SHA-256 is computed here; real scores proven with Jolt are below.5
Presets
Scores proven over committed test sets
assets/zk/benchmarks.json (Apple M4, 16 GB, native, median of three). Read-out rows and bound: measured on the two published S11 guests (w5.s11_boundary, synthetic activations); the PLATFORM.md §3.2 budget of 0.2 M + 20–33 k cycles per feature was the estimate. Browser and native classes: PLATFORM.md §1. Trace limit: 4,194,304 = 2²².assets/data/platform/bench.json, derived from data/getrm/consensus.json and the CPIC allele table. Source: CDC. Use of this material does not imply endorsement by CDC, HHS or the US Government.Platform · E
Reads from a public genome are fetched by byte range, called on a CPU and committed like a DNA file. The calls are then proven with Jolt.
Module M10 · Layer E
One recorded run on public data, drawn from its own output.
Reads to a commitment · NA07348 · CYP2C19
Panel site
01 · Byte ranges read from the BAM file
02 · Reads over the reference
03 · Genotype call
04 · VCF record
05 · Salted Merkle commitment
06 · Proof
syntelea-reads on an Apple Silicon Mac, CPU only. Every mark is read from its output; reads shown are the ones bcftools counted (mapping and base quality ≥ 20).1–4Seven genes from the same genome, each proven
Prove a DNA file in the browser
assets/data/platform/reads.json.Platform · D1
M8 registers what a lab assay should show before anyone runs it. Syntelea does not generate, design or export biological sequences.
Module M8 · Layer D
A prediction is hashed first; its evidence tier rises only with a measurement.
A pathogenicity score does not measure enzyme function, so each prediction is fixed before an assay tests it.
CYP2C19
01 · Registered
–
SHA-256, this tab
02 · Measured
No measurement yet
No laboratory has run it
03 · Evidence tier
T3 · Hypothesis
Until a measurement exists
Variant
Sites and CPIC function from the engine's rule tables;1 EVEE values from the snapshot Explore uses.2 The digest is computed in this browser; the log client is in development.
Platform · D2
M9 fixes each study plan with a SHA-256 digest before the run, so a criterion changed later shows up. Benchmarks and agent tools use public data only.
Module M9 · Layer D
Keeps Syntelea's own studies checkable, from plan to figure.
Log entry
Draft · nothing sealedStart from a Syntelea plan
Edit any field: the digest changes with every character.
Computed in this browser with your clock; nothing is sent. The log client, a local Rekor1 plus OpenTimestamps,2 is in development. Nothing is written to the public Rekor log yet, because its entries are permanent.
GeT-RM published consensus genotypes for 363 cell-line samples.3 Here they are counted against the panel.
Gene
Step
Input
Counted in this browser from data/getrm/consensus.json (CDC GeT-RM). Our CPIC tables' accuracy over 1,981 of these calls is proven with Jolt, over a test set committed first.
Samples with public whole-genome reads can be called from those reads and compared with their gold calls.
Counted in this browser from data/getrm/consensus.json; rows are the loci with a T0 rule table in assets/rulesets/registry.json. The checked call comes from one syntelea-reads run (assets/data/platform/reads.json); no Jolt proof is involved.
Each set in data/ records the SHA-256 of its files. Pick a file: this tab fetches it and hashes it again.
data/
File
Records read from each PROVENANCE.json. Hashing uses the browser's own SHA-256; no Jolt proof is involved.
Syntelea's MCP server gives agents read-only tools over public data. No tool reaches a DNA file, a commitment key or a prover key.
Try a request
Also in M9
Platform · Plan
Black blocks are proof statements, orange the proven result; dashed blocks are still to come.
Work plan
Built means inside a Jolt proof and tested today; the rest comes in waves, or is planned.
Statuses follow PLATFORM.md section 2 (feasibility of every item) and the build waves of section 4.
Proof statements
Each meter has three steps: specified, guest built, proven. Select a statement to see what it takes from the others.
Proof statements
Measured
Each proof is padded to a size set only by its number of sites, so it looks the same whether your file has those sites or not.
Panel-site proof (S1) · native runs
Limits
A feature that breaks one of them does not ship.
A site with no call stays unknown, never normal
M1 · M3One proof holds at most 43 panel sites (31 from a VCF); a larger panel is split into several proofs
M1 · M10CYP2D6 copy number and HLA typing need sequencing callers, not array files
M1Polygenic scores carry ancestry bias; only a band leaves the device8
M3Small cohort counts are suppressed9
M4Several phenotypes together can identify a person
M1 · M3A hosted model's forward pass is assumed; only the read-out is proven6
M5 · M7Dosing and clinical outputs stay labelled research10 11
M1 · M9Syntelea does not generate, design or export biological sequences
M8Retrieved 2026-10-21 and 2026-10-22.
About
A project that applies zero-knowledge proofs to pharmacogenomics and tests what DNA models encode.
CPIC rules for seven genes on committed genotypes, proven with Jolt.
Public model scores beside CPIC function, for research only.
Proofs for small models, scores and SAE read-outs over committed data. The model run itself stays assumed.
Roadmap
The first proof works, and so does every proof statement added in waves 2 to 5. Wave 6, publishing the code, is still to come; the Plan lists every statement.
Read at load from assets/zk/manifest.json, assets/zk/core/BUILD.json, assets/rulesets/registry.json, assets/sitesets/registry.json, data/*/PROVENANCE.json and the reads run. Genes in the first proof come from the engine. HLA and UGT1A1 are left out: calling them needs whole-genome data. Waves: PLATFORM.md, section 4.
Limits
Contact
contact@syntelea.com · Please do not email genetic or health data.