BioFS Protocol v3.18 & Sequentia BioRouter

A BioNFT-gated legal-act filesystem for human biology, its AI-agent attachment, and cascading consent

Daniel Uribe, founder and CEO, GenoBank.io | GenoBank.io Research & Engineering Team

Sequentia Network (chain id 15132025) and the BioRouter protocol

August 11, 2026

Version 3.18

Share this whitepaper: LinkedIn X (Twitter) Download PDF

1. Introduction

1.1 A non-revocable asset held by everyone but the patient

A genome is not a rotatable password. It is identifying, lifelong for the germline, and partially shared with relatives, so a single leaked variant can expose a person who never consented. Yet the prevailing model for genomic data is centralized custody: direct-to-consumer testing firms, academic medical centres, and institutional biobanks aggregate exome and whole-genome data into proprietary silos and license access to pharmaceutical and research partners, while the data subject receives no ongoing consent visibility, no cryptographically verifiable attribution, and no economic participation. The same custody model cannot satisfy the law it operates under. When a patient invokes GDPR Article 17 or CCPA Section 1798.105, a centralized holder cannot produce tamper-evident proof that secondary copies, intermediate BAM alignments, and downstream variant annotations have actually been purged from distributed backups and partner databases.

1.2 The AI-interface failure

The injury is now compounded by a second failure. Laboratories hold a patient’s data in cloud object storage; patients increasingly reach for a conversational AI to interpret it; and the Model Context Protocol (MCP) has become the bus over which agents attach to external data. What is missing is the layer in between. Handing an agent a raw FASTQ or BAM does not work: the file is almost uninterpretable as raw model context, it exceeds any practical window, and, more importantly, it bypasses the consent that governs the underlying biosample. The patient is left with data they cannot use and rights they cannot exercise.

1.3 The systems root, and a second-generation failure

Biology never received the primitives the internet was built on. The network had TCP/IP for a shared address space and POSIX for a shared filesystem grammar, with an inode to name a file and an unlink to remove it. Human biology has neither a portable name for a biosample nor an operation that revokes access to it. GA4GH standards describe how genomic data should be discovered and passed, but they are advisory, and Articles 9 and 17 of the GDPR require controls that execute, not controls that are documented.

A first attempt to close this gap on general-purpose blockchains failed on its own terms. Programmable IP licensing deployed on general Ethereum L1 or L2 EVM networks incurred gas costs around $22.00 per VCF tokenization, saw transaction revert rates above 60% under cohort-scale load, and offered no native primitives for BGZF block compression, tabix indexing, or genomic scoring, forcing heavy off-chain wrappers that reintroduced the centralization the design was meant to remove.

1.4 Thesis

BioFS resolves both failures by treating the law as a design specification and building the filesystem the law implies. Its operations are legal acts: to tokenize is to assert ownership, to grant is to license, to stream is to exercise an access right, and to erase is to execute Article 17. Two primitives carry the model. A biocid names a biosample permanently and independently of where its bytes live, and a revocable BioNFT, the instrument introduced in earlier work on genomic data and non-fungible tokens, carries ownership and consent. A public ledger, Sequentia L1, records identity, consent state, storage routes, and settlement, so it holds rights and pointers, never biological content. An AI agent attaches through MCP and reads specific regions through streamed sub-chunk access, so interpretation happens without bulk sequence ever entering the model. Because every paid read re-exercises the patient’s authorization and every revocation is enforced at the next access, consent here is not a one-time signature but an ongoing, economically live relationship the donor governs: what the companion paper calls Metamorphic Consent. The remainder of this paper specifies the architecture (Section 3), the operations that enforce legal acts (Section 4), the zero-byte agent attachment (Section 5), multi-modal coverage (Section 6), the biophysical scoring substrate (Section 7), the resolution of the immutability-versus-erasure paradox under Article 17 (Section 8), and production deployment and evidence (Sections 9 to 11).

2. Background: the legal status of biological data and the immutability-versus-erasure paradox

We read the law as a specification that constrains the address space, the storage plane, and the consent model a lawful system may implement. Under the GDPR, genetic data is a special category (Article 9) that may be processed only on a narrow basis such as explicit consent, and a data subject holds the right to erasure (Article 17). CCPA Section 1798.105 grants an analogous right to deletion. A lawful genomic system must therefore be able to prove, not merely assert, that a withdrawal took effect.

This requirement collides with the property that makes a blockchain useful. A ledger is valuable precisely because it is append-only and tamper-evident, yet an append-only store cannot forget. The naive resolutions both fail. Writing genomic bytes on-chain, or onto a content-addressed permanent store such as IPFS, makes erasure impossible, because immutable and un-unpinnable storage cannot honor Article 17. Keeping everything in a centralized store restores erasure in principle but destroys the auditability, attribution, and revocation the ledger was there to provide.

BioFS resolves the paradox by separating 2 planes. The ledger holds only pointers, hashes, consent flags, and rights, all pseudonymous and none of it biological content. The biological bytes live in deletable object storage under content addressing. Revoking consent and purging the deletable copy gives effect to Article 17 for that copy, while the residual on-chain anchor is a pointer and a hash, not the biology, and is never claimed as anonymized data. Section 8 specifies this mechanism and its cascade across derivative data. This is also why BioFS does not use IPFS as a storage substrate: a filesystem whose central promise is erasure cannot rest on a store that cannot delete.

3. What BioFS is: the biocid address space, storage planes, and the 5-layer stack

3.1 biocid: a permanent name for a biosample

Every biological dataset in BioFS is addressed by a content identifier of the form biocid://{bioagent}/{owner_wallet}/{biodata_type}/{dataset}; the companion paper writes a simplified 3-part form, biocid://<wallet>/<type>/<identifier>, of this same address. The identifier is the same regardless of which laboratory produced the data, which bucket holds the bytes, or which application reads it, so it functions as the inode and the DNS entry that biology never had, and it is dereferenced only through the consent gate, never as a raw or signed storage URL. Each biocid is registered with a content-free SHA-256 digest of the encrypted payload, so any local copy can be verified against the on-chain digest, and is resolved to live storage locations through the BioRoutes contract. No DNA fingerprint touches the public ledger: the discovery index that lets a peer learn a variant exists is a gated, server-side keyed-HMAC structure held off-chain, and the Bloom filter that speeds a consent check is an access-control pre-filter that can only deny, not a genomic fingerprint.

3.2 Storage planes (Layer 0)

Raw payloads (FASTQ, aligned BAM and CRAM, VCF, and DICOM medical imaging) live in encrypted object storage on Google Cloud Storage and in local POSIX directory trees. Payloads are read-only, write-once, and content-addressed. Each biocid is resolved across 4 route tiers, PRIMARY, SECONDARY, ARCHIVE, and MIRROR, so storage location can migrate without changing the name. Because Article 17 erasure is a first-class guarantee (Section 8), every storage plane is a deletable one; immutable stores are excluded by design.

3.3 The 5-layer protocol stack

BioFS is layered so that on-chain trust anchors, high-speed transport, and agent execution stay strictly isolated.

• Layer 0, storage. Encrypted, content-addressed object storage and POSIX trees holding the biological bytes across the 4 route tiers.
• Layer 1, Sequentia L1 and BioRouter contracts. The single source of truth for identity, consent state, route mapping, and settlement, on CometBFT consensus, with on-chain writes finalizing in roughly 3 seconds (soft, probabilistic); the Sequentia L1 network specification details the consensus and EVM architecture.
• Layer 2, biofs-node transport. A server-side streaming engine in Rust and Go, over QUIC (TLS 1.3) on UDP port 4433, with EIP-712 session handshakes and sub-chunk region access (Section 5).
• Layer 3, BioOS compute and MCP. Agentic execution over bio-vaults, exposing structured bioinformatics tools to frontier LLMs through an MCP server, with no local retention of genomic bytes.
• Layer 4, sovereignty and family vaults. User-facing primitives: EIP-55 custodial biowallets, BIP-32 key derivation, Tangem NFC hardware cards, patient-owned digital-twin adapters, and a proposed family-vault construct over kin-shared variants, not exercised in the reported deployment.

3.4 The BioRouter contract suite

The BioRouter suite spans contracts deployed on Sequentia L1 (Chain ID 15132025) and contracts still on the roadmap; the Status column marks each, and the companion paper gives every deployed address with its current on-chain state.

The gate of record is the checkPermission predicate, evaluated against a grant table that carries each grant’s transaction hash and block number, with the Bloom filter as a pre-filter that can only deny. The BioPIL license algebra (roadmap) is designed to enumerate 7 license types, Non-Commercial (0), Commercial Use (1), Research Use (2), Clinical Use (3), AI Training (4), Pharma Research (5), and GDPR-Revocable (6), each carrying a time-bounded lease and a royalty value in basis points from 0 to 10,000; until it ships, license scope is enforced off-chain against the checkPermission gate. Treasury settlement accrues to 0x5C9d00f5BC59037A038A2936F0F614b770f947A7.

4. Filesystem operations as legal acts

BioFS exposes 58 or more command verbs through biofs-cli, grouped into 11 functional domains. The design principle is that a verb is a legal act that the filesystem enforces and records against pre-formed rights; the filesystem executes a right, it does not adjudicate whether the right exists.

• Identity and session (login, whoami, researcher) establish who is acting, via Web3 wallet signature, Magic link, passkey, ORCID, or Tangem card.
• Tokenization (tokenize, tokenize-biosample, tokenize-fastqs) is the act of asserting ownership: it mints a BioNFT, records the content-free payload digest and the off-chain discovery-index entry, and sets the BioPIL license terms.
• Access and permissions (access, share, cred, ticket) are the acts of licensing: they grant scoped, time-bounded rights to named delegate wallets and accredited laboratories.
• Read (get, cat, view, resolve, stream) exercises an access right, validating the BioPIL license and active consent before any byte is returned.
• Erase (rm, bionft revoke, access revoke-consent) executes Article 17: it revokes consent, burns access tokens, and triggers the purge cascade of Section 8.
• Compute, twins, spectroscopy, clinical, biowallets, payments, and lineage round out the surface (Appendix A).

The two load-bearing flows are the open path and the erase path. On open, a client presents an EIP-712 signature, the node resolves the biocid, validates the on-chain license and consent bound to it, and only then does a region stream begin. On erase, a single patient signature sets the consent flag on the biocid and the revocation cascades to every derivative. The full verb reference, with sub-verbs and execution behavior for all 11 domains, is given in Appendix A.

5. Zero-byte AI attachment: biofs-node, QUIC sub-chunk access, and MCP

biofs-node, implemented in Rust and Go, is the server-side engine that links the on-chain BioRouter to physical storage and to AI agents. It communicates over QUIC (TLS 1.3) on UDP port 4433, and each session opens with an EIP-712 wallet-signature handshake that validates BioNFT ownership or an active BioPIL license before any stream is established.

To serve 10 to 100 GB whole-genome BAM and BGZF VCF files without a full download, biofs-node uses ROSA, a read-only segmented-access engine that parses BGZF block headers and tabix index structures in memory, so a client can stream a single chromosomal region, for example chr17:43,044,295-43,125,483 for BRCA1, in one range request. Batch queries are decoded from a single compound wire frame, so license verification, file info, region reads, and per-gene variant lookups complete in one round trip.

The attachment point for AI is the Model Context Protocol. Through a BioFS MCP server, a frontier LLM invokes structured bioinformatics tools (for example bio_stream_chunk and workspace_classify) that return regions, variants, and derived summaries, never raw sequence, and retain no genomic bytes locally. The agent reads skills and regions, it does not cat a FASTQ into the prompt. Because the same interface answers the GA4GH data-access verbs, BioFS can serve as an executable layer for GA4GH DRS retrieval, Beacon discovery, and Passport-scoped authorization, giving advisory standards an enforcing runtime. Attachment is not open by default: human researchers and AI agents must register under ERC-8004, with KYC or attestation sufficient for the legal capacity to receive grants, and carry a non-transferable identity and staked reputation before a Governor can grant them purpose-bound access, so an agent at a negative reputation score is economically excluded rather than free to re-mint a fresh identity.

6. One address space, three modalities

The biocid form and the consent gate are identical across data types. Genomic files (FASTQ, BAM, VCF), DICOM medical imaging studies, and FHIR clinical records are each named by a biocid, bound to the patient wallet, and read only through the consent gate. A patient’s AI agent therefore reaches genome, image, and record through one attachment rather than 3 integrations. The limit worth stating plainly is that a provider that retains its own copy of a study outside the vault is outside the gate for that copy; BioFS governs the copy written to the patient, and its guarantees are guarantees about that copy.

7. Biophysical spectroscopy substrate: Cosic-RRM variant scoring

Beyond storage and access, the filesystem can compute directly on the sequences it holds. A distinctive capability of BioFS v3.18 is native protein biophysical spectroscopy for missense variant scoring, which supplies physical evidence orthogonal to statistical classifiers such as AlphaMissense and REVEL.

The Resonant Recognition Model (RRM) of Cosic represents a protein as a discrete signal by mapping each residue to its Electron-Ion Interaction Potential (EIIP), the average energy of valence electrons:

\[ \text{EIIP} = 0.25\, Z^{*} \cdot \frac{\sin(1.04\pi Z^{*})}{2\pi} \]

where \(Z^{*}\) is the average valence-electron number per atom of the side chain. The 20 amino acids map to fixed constants (for example Leu = 0.0000, Glu = 0.0057, Cys = 0.0829, Phe = 0.0946 Ry). The characteristic frequency \(f_c\) of a functional protein family is the prominent peak of the cross-spectrum of the discrete Fourier transforms of \(M\) homologous members:

\[ X_m(k) = \sum_{n=0}^{N-1} x_m(n)\, e^{-j 2\pi k n / N}, \qquad S(k) = \prod_{m=1}^{M} |X_m(k)| \]

A missense variant at position \(n_0\) induces a spectral energy shift at \(f_c\):

\[ \Delta E\% = \frac{|X_{\text{mut}}(f_c)|^2 - |X_{\text{wt}}(f_c)|^2}{|X_{\text{wt}}(f_c)|^2}\times 100\% \]

The Piezoelectric Signal Model complements the RRM by evaluating side-chain dipole coupling, and BioFS applies a Morlet continuous wavelet transform to localize resonance disruption along the backbone. A variant that produces \(\Delta E\% > 15\%\) or severe wavelet-power dissipation at \(f_c\) is flagged by biofs fourier-score as a biophysical pathogenicity candidate for review. The related verbs (rrm-consensus, psm-consensus, rrm-distribution, rrm-train) compute family frequencies and train an XGBoost ensemble that combines these physical features with AlphaMissense, REVEL, and PrimateAI scores.

8. Resolving immutability versus erasure under GDPR Article 17

Genomic analysis is iterative, so BioFS models data as an acyclic directed graph of transformations, a derivative-lineage graph: FASTQ to BAM under GPU alignment, BAM to VCF under variant calling, VCF to an annotated SQLite store, and that store to a patient digital-twin adapter. Every derivative records its parent biocid and creator wallet in its on-chain manifest, so the lineage of any object is explicit.

That lineage is what makes erasure provable rather than promised. When a patient runs biofs rm <biocid> or invokes consent revocation, three things happen. The primary consent flag is set to revoked on-chain in the ConsentManager. The gate then rejects every future license verification for the parent biocid and for all registered child biocids in its derivative-lineage tree. And biofs-node purges cached BGZF indexes and derivative stores on its next synchronization cycle, closes active streams, and instructs attached inference nodes to drop cached context. Because the ledger held only pointers and hashes, and the biological bytes lived in deletable storage, destroying the deletable copies gives effect to Article 17 for those copies. The residual on-chain anchor is a pointer and a hash; it is not the biology, and it is not offered as anonymized data. This is the concrete reason a permanent store such as IPFS is unsuitable: it cannot participate in the purge.

9. Deployment and evaluation

BioFS v3.18 is in production on Sequentia L1, with the deployed contracts of Section 3.4 live at their listed addresses and the checkPermission gate as the authorization of record. We evaluated it against a centralized-biobank baseline and against the prior first-generation EVM implementation.

Methodology. These are single-operator point measurements from one production deployment (42 laboratories, 8,547 consented biosamples under active BioNFTs), not distributions and not an independent audit. Two figures are independently checkable: the gas costs are deterministic products of fixed Solidity 0.8.x execution paths, recoverable by anyone through eth_estimateGas or a transaction receipt against the public contract addresses, and biocid resolution latency is the warm-read wall-clock of a resolver lookup plus gate evaluation, reproducible against the public read endpoint. Region-extraction latency is a single BRCA1 range request against whole-genome BAM and BGZF payloads. A third-party benchmark and a formal contract audit are on the roadmap.

Threat model and limits. The ledger is trust-minimized for identity, consent, and settlement under CometBFT. Three limits remain explicit. Laboratory attestation that a biocid corresponds to a real biosample is an oracle input, mitigated by ERC-8004 laboratory identity and the reputation system but not eliminated. Key management is a real risk for non-technical patients, mitigated by custodial biowallets and a roadmap social-recovery path rather than solved. And a per-wallet model governs a datum, not a bloodline: one member’s revocation cannot recall a familial inference already released, and one member’s disclosure can expose non-consenting relatives, a limit the proposed family-vault narrows but does not eliminate. Erasure guarantees bind the copies BioFS governs, not copies a provider retains outside the vault.

10. Discussion: economics, adoption, and compliance

The economic model makes consent an ongoing economic relationship rather than a one-time surrender. Every paid read routes through the on-chain PaymentRouter, which settles a 70/20/10 split to lab, annotator, and platform atomically in one transaction, using pull-payment accounting so no reverting payee can block another, and deposits a settled-payment receipt into the governing BioNFT. That receipt is what makes a patient’s Biodata Dividend a provable entitlement rather than a promise: at cohort scale, Shapley-based attribution apportions each dataset’s marginal contribution, estimated by Monte Carlo permutation sampling and recomputable by any contributor from the on-chain event history. Because authorization is revocable at the BioNFT and settlement is atomic, a donor who withdraws consent stops both the data flow and the revenue split in one act. Payment never substitutes for assent: Biodata Dividends are benefit-sharing for a governed, revocable data-use relationship rather than payment for a specimen, and for economically vulnerable donors the imbalance between subject and controller is a residual risk to the freely-given basis of consent that a research-ethics review, not the architecture alone, must weigh.

Adoption does not require any institution to surrender its role. The end state is one in which laboratories write to the patient wallet by default at the moment of sequencing, so a patient is enrolled in their own vault as a byproduct of the diagnostic workup rather than as a discretionary act. The credible barriers are institutional inertia, key management for non-technical patients, the oracle trust in laboratory attestation, and regulatory acceptance of on-chain revocation as satisfying Article 17. Multi-regime compliance (GDPR, CCPA, and analogous regimes) is addressed by the same dual-plane design rather than by a per-jurisdiction rebuild.

11. Conclusion and recommendations

BioFS Protocol v3.18 and the Sequentia BioRouter provide an economically viable and legally coherent substrate for sovereign genomic data governance. By treating the law as a design specification, naming biosamples with content-addressed biocids, gating every read on a revocable BioNFT, streaming regions over QUIC without bulk dump, and cascading erasure across derivative lineage, BioFS gives patients control that executes and gives AI agents a lawful attachment point. We make 3 recommendations. Standards bodies should give GA4GH interfaces an enforcing runtime rather than leaving them advisory. Laboratories should adopt the biocid write path so ownership begins at the point of sequencing. And AI hosts should attach to biological data through a consent gate such as the BioFS MCP server, so the question of whose agent reads the genome is answered by the patient’s key. We invite implementers to build on the BioRouter.

Appendix A: biofs-cli command reference (11 domains)

Domain 1, local configuration and session. config (set/get/list); login (–magic, –wallet, –passkey, –orcid); logout; whoami; researcher (register, status).

Domain 2, BioNFT tokenization and asset management. tokenize; tokenize-biosample; tokenize-fastqs; ls/list; get/download; cat; rm/erase; inspect/info; bionft (view, status, revoke); context (create, publish, verify, revoke).

Domain 3, access control and permissions. access (request, grant, revoke, list, check, revoke-consent); share; shares; labnfts/labs; lab (refresh-coverage, status); ticket (list, revoke); cred (issue, status, list, revoke).

Domain 4, storage routing and FUSE vault. vault (setup, mount, status, unmount); mount/umount; resolve; route (check, heal); upload-fastq.

Domain 5, BioOS compute and GPU pipelines. pipeline (run-wes, run-somatic); job (create, status, results, list, submit-clara, recall, reconcile); agent-health; cohort-pipeline; annotate; imaging (pull).

Domain 6, patient digital twins and AI intelligence. twin (train, status, compare); cancermap (regen, status); fluency (build, state); ancestry/somos (ingest, somos, status, validate-cohort, repair-cohort); dissect; match.

Domain 7, biophysical resonances and Cosic-RRM spectroscopy. fourier-score; rrm-consensus; psm-consensus; wavelet-consensus; tokenize-spectrum; bode; rrm-distribution; rrm-train; cohort-fourier-score; cohort-train.

Domain 8, clinical genomics, ACMG, and external integrations. clinical; cohort-acmg; variants; myvariant; mavedb-ingest (run, status, stats, query); mychart (connect, list, view, stats, register, refresh).

Domain 9, custodial biowallets and family vaults. biowallet (create, list, bind, family); family (create, derive, list); claim.

Domain 10, x402 micropayments and agent economy. payment/pay; agent (register, list, health).

Domain 11, derivative lineage and verification. lineage; verify; view.

References

1. Cosic, I. (1994). Resonant Recognition Model of Macromolecular Interactions. Progress in Biophysics and Molecular Biology, 62(3), 243-272.
2. Cosic, I., Cosic, D., & Lazar, K. (2016). Is It Possible to Predict Biological Function of Macromolecules Using Resonant Recognition Model? IEEE Transactions on NanoBioscience, 15(4), 316-322.
3. Richards, S., et al. (2015). Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the ACMG and AMP. Genetics in Medicine, 17(5), 405-424.
4. Global Alliance for Genomics and Health. (2024). GA4GH Passport Specification v2.0. GA4GH Regulatory and Ethics Work Stream.
5. European Union. (2016). General Data Protection Regulation, Articles 9 and 17. Regulation (EU) 2016/679.
6. State of California. (2018). California Consumer Privacy Act, Section 1798.105.
7. Coinbase Developer Platform. (2024). x402 HTTP Payment Protocol Specification.
8. Buchman, E., Kwon, J., & Milosevic, Z. (2018). The latest gossip on BFT consensus (Tendermint / CometBFT).
9. Cheng, Y., et al. (2023). Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science, 381(6664).
10. Ioannidis, N. M., et al. (2016). REVEL: an ensemble method for predicting the pathogenicity of rare missense variants. American Journal of Human Genetics, 99(4), 877-885.
11. Uribe, D. (2026). BioFS: A Web3 Legal-Act Filesystem for Human Biology and AI Agents. The Journal of the British Blockchain Association (companion paper).
12. Uribe, D., & Waters, G. (2020). Privacy Laws, Genomic Data and Non-Fungible Tokens. The Journal of the British Blockchain Association, 3(2).
13. Uribe, D., & GenoBank.io Engineering Team. (2025). Sequentia L1 Network Specification and CometBFT EVM Architecture. GenoBank.io Whitepapers.
14. Sundaram, L., et al. (2018). Predicting the clinical impact of human mutation with deep neural networks. Nature Genetics, 50(8), 1161-1170.
15. Ethereum Improvement Proposals. (2025). ERC-8004: Trustless Agents.