[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6},"actionData","errors","routes/hashtag",{"_7":8,"_9":10,"_11":12},"articles",[13],"authors",{},"tag","encryprion",{"_14":15,"_16":17,"_18":19,"_20":21,"_22":23,"_24":25,"_26":27},"content","## A design for erasure-coded, privacy-preserving blob storage\n\n**Design Document** \n**December 5th 2025**\n\n---\n\n## Abstract\n\nThis document describes a distributed storage system built upon Nostr and Blossom infrastructure that provides durable, privacy-preserving storage for immutable blobs through erasure coding across independent servers. The system maintains a hierarchical namespace analogous to a filesystem through content-addressed manifests organized in a Merkle DAG structure. State evolution is tracked via a cryptographically-linked hash chain of commit events, enabling complete auditability and direct disaster recovery. The entire dataset-including all historical state-remains recoverable from a single cryptographic key. This design prioritizes user sovereignty: the owner explicitly controls when changes are committed and which servers store them, and decides when obsolete data is garbage collected.\n\n---\n\n## Table of Contents\n\n1. [Introduction](#1-introduction)\n2. [Design Goals](#2-design-goals)\n3. [Architecture Overview](#3-architecture-overview)\n4. [Block Layer](#4-block-layer)\n5. [Erasure Coding Layer](#5-erasure-coding-layer)\n6. [Encryption Layer](#6-encryption-layer)\n7. [Inode Structure](#7-inode-structure)\n8. [Directory Hierarchy](#8-directory-hierarchy)\n9. [State Management via Hash Chain](#9-state-management-via-hash-chain)\n10. [Single-Key Discovery and Recovery](#10-single-key-discovery-and-recovery)\n11. [Transport Layer](#11-transport-layer)\n12. [Privacy-Preserving Existence Queries](#12-privacy-preserving-existence-queries)\n13. [Garbage Collection](#13-garbage-collection)\n14. [Lifecycle Summary](#14-lifecycle-summary)\n15. [Security Analysis](#15-security-analysis)\n16. [Future Considerations](#16-future-considerations)\n17. [Conclusion](#17-conclusion)\n\n---\n\n## 1. Introduction\n\nThe proliferation of cloud storage services has created a fundamental tension between convenience and sovereignty. Users gain automatic synchronization across devices but surrender control over their data to third parties who may inspect it, monetize it, lose it, or deny access to it. The alternative-self-hosted infrastructure-demands technical expertise and ongoing maintenance that most users cannot provide.\n\nThis system addresses a specific need: reliable long-term storage of personal data across unreliable infrastructure, with complete privacy from storage providers and full recoverability from a single secret. The design assumes that individual storage providers may disappear without warning, yet data should survive as long as a sufficient subset of providers remains operational.\n\nThe architecture leverages two existing protocols. Nostr provides a decentralized identity layer where users control cryptographic keypairs and can publish signed events to any compatible relay. Blossom extends this model to binary data, offering content-addressed blob storage over HTTP with Nostr-based authentication. By combining these protocols with erasure coding and client-side encryption, the system achieves properties that neither protocol provides alone.\n\nThe core requirements are as follows. First, durability: the system must tolerate arbitrary server failures up to a configurable threshold without data loss. Second, privacy: storage servers must learn nothing about stored content, file sizes, directory structure, or access patterns. Third, sovereignty: the user must retain complete control over their data, including explicit authority over when changes are committed and when old data is deleted. Fourth, recoverability: the entire dataset, including its complete history, must be recoverable from a single secret key. Fifth, simplicity: storage servers remain minimal in functionality-they store and retrieve opaque blobs, nothing more.\n\n---\n\n## 2. Design Goals\n\n### 2.1 Threat Model\n\nStorage servers are assumed to be honest-but-curious and unreliable. A server will store data if compensated or incentivized to do so, and may inspect any data it stores, analyzing content, access patterns, and timing. Servers may disappear without notice, whether due to business failure, legal action, or technical problems. Active data corruption is not a concern, since content addressing makes such tampering immediately detectable.\n\nThe system explicitly does not protect against a coordinated attack where more than n - k servers simultaneously destroy their shares of a given block. Such an attack requires either widespread collusion or a correlated failure mode affecting the majority of selected servers. Users concerned about this threat should select servers across diverse jurisdictions and independent operators.\n\nThe system also does not protect against compromise of the user's secret key. An attacker with access to the nsec can decrypt all data and forge new commits, and irreversibly delete the dataset by publishing malicious state updates. Key management remains the user's responsibility.\n\n### 2.2 Non-Goals\n\nSeveral capabilities are explicitly outside the scope of this design.\n\nMutability of stored blobs is not supported. All blobs are immutable and content-addressed. Modifying a file creates new blobs; the old blobs remain until garbage collected. This immutability simplifies consistency and enables deterministic caching, with natural versioning as a side effect.\n\nMulti-user access control is not addressed. This design focuses on single-owner storage where one keypair controls the entire dataset. Sharing files or delegating access to other users would require additional mechanisms not specified here.\n\nReal-time synchronization is not a goal. The system optimizes for backup and archival workloads, where latency is secondary to durability. Changes accumulate locally and are committed in explicit, user-initiated snapshots.\n\nAutomated background maintenance is deferred. While a steward service could handle ongoing verification and repair, this design places that responsibility with the client application. The user must periodically verify data availability and initiate repairs when servers fail.\n\n---\n\n## 3. Architecture Overview\n\nThe system organizes functionality into distinct layers, each with a single responsibility. Data flows downward through these layers on write and upward on read.\n\n```\n┌─────────────────────────────────────────────────────────────┐\n│ State Layer │\n│ (hash chain of commits, garbage collection) │\n├─────────────────────────────────────────────────────────────┤\n│ Namespace Layer │\n│ (paths → inodes, Merkle DAG) │\n├─────────────────────────────────────────────────────────────┤\n│ Inode Layer │\n│ (file metadata, block references) │\n├─────────────────────────────────────────────────────────────┤\n│ Encryption Layer │\n│ (per-block keys, authenticated encryption) │\n├─────────────────────────────────────────────────────────────┤\n│ Block Layer │\n│ (fixed-size, padded chunks) │\n├─────────────────────────────────────────────────────────────┤\n│ Erasure Coding Layer │\n│ (Reed-Solomon k-of-n shares) │\n├─────────────────────────────────────────────────────────────┤\n│ Transport Layer │\n│ (Blossom PUT/GET, Nostr events) │\n└─────────────────────────────────────────────────────────────┘\n```\n\nWhen writing a file, the client divides data into fixed-size blocks, encrypts each block with a derived key, erasure-codes each encrypted block into n shares, and uploads those shares to n distinct Blossom servers. The client then constructs an inode containing the metadata needed to reverse this process, encrypts and stores the inode using the same pipeline, updates the directory hierarchy, and publishes a new commit event to the hash chain.\n\nWhen reading a file, the client traverses from the current chain head through the directory structure to locate the target inode, fetches any k of the n shares for each block, decodes and decrypts the blocks, and reassembles the original file.\n\n---\n\n## 4. Block Layer\n\n### 4.1 Fixed-Size Blocks\n\nAll data entering the system is divided into fixed-size blocks before any cryptographic processing. The block size B is a system parameter, typically 256 KiB (262,144 bytes), though implementations may support alternative sizes for specific use cases.\n\nFor a file of size S bytes, the number of blocks is:\n\n```\nN_blocks = ⌈S / B⌉\n```\n\nThe final block is padded to exactly B bytes using a length-prefixed padding scheme. The first two bytes of each block encode the actual content length as a big-endian 16-bit unsigned integer, followed by the content bytes, followed by zero bytes to fill the block:\n\n```\n[content_length: u16_be][content: content_length bytes][padding: zeros]\n```\n\nThis scheme enables unambiguous removal of padding during reconstruction. For the final block of a file, content_length contains `S mod B` (or B if S is an exact multiple). For non-final blocks, content_length equals B - 2, and the two-byte overhead slightly reduces effective capacity.\n\n### 4.2 Privacy Through Uniformity\n\nThe decision to use fixed-size blocks with padding is primarily motivated by privacy. When all stored blobs are exactly the same size, external observers cannot perform traffic analysis based on blob dimensions.\n\nWithout uniform sizing, an adversary observing uploads could distinguish small files from chunks of large files based on byte counts. They could infer file types from characteristic size patterns-a 4.7 GB blob likely represents a DVD image, while a 25 MB blob with specific dimensions suggests a high-resolution photograph. They could correlate related blobs by noticing that blobs uploaded together have sizes summing to a plausible file size. They could identify whether a blob contains user data or system metadata based on typical metadata sizes.\n\nWith uniform blocks, all stored blobs appear identical in size. A 100-byte text file produces the same 256 KiB blob as a chunk of a multi-gigabyte video. Directory metadata and file content are indistinguishable. The only information leaked is the count of blocks, which provides a loose upper bound on total data volume but reveals nothing about how that volume is distributed across files.\n\nThis uniformity has costs. Small files incur substantial padding overhead-a 1 KiB file stored in a 256 KiB block wastes 99.6% of the space, and after erasure coding with overhead factor 1.5x, that 1 KiB file consumes 384 KiB of storage across servers. This design accepts that tradeoff. Systems requiring efficient small-file storage should consider alternative approaches, but such optimizations necessarily leak information about file sizes.\n\n---\n\n## 5. Erasure Coding Layer\n\n### 5.1 Reed-Solomon Coding\n\nEach encrypted block undergoes erasure coding to provide redundancy across multiple storage servers. The system employs Reed-Solomon codes, which belong to the family of Maximum Distance Separable (MDS) codes-they achieve the theoretical optimum for storage efficiency at any given fault tolerance level.\n\nA Reed-Solomon (n, k) code transforms k source symbols into n encoded symbols such that any k of the n symbols suffice to reconstruct the original data. The system can tolerate the loss of any n - k symbols, whether from server failures, network partitions, or data corruption.\n\nThe mathematical foundation rests on arithmetic in a Galois Field. This implementation uses GF(2^8), the field with 256 elements corresponding to byte values 0-255. Field elements are represented as polynomials over GF(2) with degree less than 8, and the field is constructed using the irreducible polynomial:\n\n```\np(x) = x⁸ + x⁴ + x³ + x² + 1\n```\n\nAddition in GF(2^8) is XOR of byte values. Multiplication is more complex but can be implemented efficiently using precomputed logarithm tables. For elements a and b with a,b ≠ 0:\n\n```\na × b = gf_exp[(gf_log[a] + gf_log[b]) mod 255]\n```\n\nwhere gf_log and gf_exp are 256-element lookup tables computed once at initialization using a generator element (typically α = 2).\n\n### 5.2 Encoding Process\n\nEncoding treats each byte position across the k source blocks as coefficients of a polynomial. For byte position i, let b₀, b₁, ..., b_{k-1} be the bytes at position i in each of the k source blocks. These define a polynomial:\n\n```\nP(x) = b₀ + b₁x + b₂x² + ... + b_{k-1}x^{k-1}\n```\n\nThe n encoded shares contain the evaluations of P(x) at n distinct points. Using a systematic encoding where the first k shares contain the original data, the evaluation points are chosen such that P(αⁱ) for i = 0, 1, ..., k-1 reproduces the original bytes (α⁰ = 1, so P(1) = b₀ + b₁ + ... + b_{k-1} requires specific construction).\n\nIn practice, encoding multiplies the source vector by a k × n generator matrix GM derived from a Vandermonde matrix. The entry at row i, column j is α^{ij}. For systematic encoding, this matrix is transformed so that its first k columns form the identity matrix:\n\n```\nGM_systematic = V_{k,k}⁻¹ × V_{k,n}\n```\n\nThe resulting encoded shares consist of the k original blocks (unchanged) followed by n - k parity blocks.\n\n### 5.3 Decoding Process\n\nErasure decoding exploits the key property that erasure locations are known-we know which servers failed, we lack their data. This differs from error correction, where corrupted symbols must first be identified.\n\nGiven any k received shares, reconstruction proceeds as follows. Form a k × k matrix by selecting the columns of GM corresponding to the received shares. This submatrix is guaranteed to be invertible due to the MDS property. Compute its inverse. Multiply the received shares by this inverse to obtain the original source blocks.\n\nThe computational complexity of classical matrix-based decoding is O(k³) for the matrix inversion plus O(k²) for the matrix-vector multiplication. For the block sizes and redundancy parameters typical in this system, decoding completes in milliseconds on modern hardware. Implementations requiring higher throughput can employ FFT-based algorithms achieving O(n log n) complexity.\n\n### 5.4 Parameter Selection\n\nThe choice of n and k determines the tradeoff between storage overhead and fault tolerance, with operational complexity scaling accordingly.\n\n| k | n | Overhead | Tolerance | Servers Required |\n|---|---|----------|-----------|------------------|\n| 2 | 3 | 1.50× | 1 failure | 3 |\n| 3 | 5 | 1.67× | 2 failures | 5 |\n| 4 | 6 | 1.50× | 2 failures | 6 |\n| 4 | 7 | 1.75× | 3 failures | 7 |\n| 6 | 9 | 1.50× | 3 failures | 9 |\n\nFor personal storage, a (2, 3) or (3, 5) configuration provides a reasonable balance. The former tolerates one server failure with 50% overhead; the latter tolerates two failures with 67% overhead. Users with access to more servers or heightened durability requirements may choose higher parameters.\n\nThe system should store shares from the same block on distinct servers to maximize independence. If two shares land on the same server, that server's failure removes two shares, halving effective fault tolerance.\n\n### 5.5 Share Addressing\n\nEach share is content-addressed by the SHA-256 hash of its bytes:\n\n```\nshare_id = SHA256(share_bytes)\n```\n\nBlossom servers store and retrieve shares solely by this identifier. They possess no information about which file or user a share belongs to. The share_id serves as both the storage key and the integrity check-if a server returns data whose hash doesn't match the requested ID, the data is corrupt or fraudulent and must be discarded.\n\n---\n\n## 6. Encryption Layer\n\n### 6.1 Key Hierarchy\n\nEncryption employs a hierarchical key derivation scheme rooted in the user's Nostr identity. This hierarchy enables fine-grained key management while maintaining the single-key recovery property.\n\n```\nnsec (Nostr secret key)\n │\n └─► Master Storage Key (derived via HKDF)\n │\n └─► Per-File Key (random, stored encrypted in inode)\n │\n └─► Per-Block Key (derived via HKDF from file key + block index)\n```\n\nThe master storage key is derived deterministically from the nsec:\n\n```\nmaster_key = HKDF-SHA256(\n IKM = nsec,\n salt = \"nostr-storage-v1\",\n info = \"master-key\",\n length = 32\n)\n```\n\nThis derivation is deterministic-the same nsec always produces the same master key-enabling recovery without storing additional secrets.\n\nEach file receives a randomly generated 256-bit key at creation time. This per-file key is stored within the file's inode, encrypted to the master key. Random per-file keys ensure that identical files produce different ciphertexts, preventing content-based correlation.\n\nPer-block keys are derived from the file key to avoid nonce reuse concerns:\n\n```\nblock_key = HKDF-SHA256(\n IKM = file_key,\n salt = \"block-encryption\",\n info = block_index_as_u64_be,\n length = 32\n)\n```\n\n### 6.2 Authenticated Encryption\n\nEach block is encrypted using XChaCha20-Poly1305, an authenticated encryption scheme providing both confidentiality and integrity. This algorithm was selected for several reasons.\n\nXChaCha20-Poly1305 uses a 192-bit nonce, large enough that random generation is safe without coordination. The probability of nonce collision reaches 50% only after approximately 2^96 encryptions with the same key-far beyond any realistic usage. This property simplifies implementation since blocks can use random nonces without tracking which nonces have been used.\n\nThe algorithm performs well in software without hardware acceleration. On systems lacking AES-NI instructions, ChaCha20 outperforms AES by a factor of three or more. Since storage systems may run on diverse hardware including mobile devices and low-power servers, software performance matters.\n\nPoly1305 authentication detects any modification to the ciphertext. An attacker who alters even a single bit will cause decryption to fail with overwhelming probability. This integrity protection operates independently of the content-addressing scheme and provides defense in depth.\n\nThe encryption process for block i with file key K_f:\n\n```\nblock_key = HKDF-SHA256(K_f, \"block-encryption\", i, 32)\nnonce = random_bytes(24)\nciphertext = XChaCha20-Poly1305-Encrypt(block_key, nonce, plaintext_block)\nencrypted_block = nonce || ciphertext\n```\n\nThe 24-byte nonce is prepended to the ciphertext so decryption can extract it. The authentication tag (16 bytes) is appended by the AEAD construction. Total overhead per block is 40 bytes (24-byte nonce + 16-byte tag), negligible relative to the 256 KiB block size.\n\n### 6.3 Metadata Encryption\n\nFile inodes and directory entries contain sensitive metadata-filenames, sizes, timestamps, and structural relationships. This metadata receives the same encryption treatment as file content.\n\nWhen storing an inode, the client serializes it to JSON, treats the JSON as file content, and processes it through the same block/encrypt/erasure-code pipeline. The resulting shares are indistinguishable from file data shares. Directory blobs undergo identical processing.\n\nThis recursive structure means servers observe only uniform encrypted blocks. They cannot determine whether a block contains a photograph, a text document, a directory listing, or another inode. The type and structure of stored data is completely opaque.\n\n---\n\n## 7. Inode Structure\n\nAn inode contains all information necessary to reconstruct a file. After decryption, an inode is a JSON object with the following structure:\n\n```json\n{\n \"version\": 1,\n \"type\": \"file\",\n \"size\": 10485760,\n \"created\": 1701820800,\n \"modified\": 1701907200,\n \"key\": \"\",\n \"blocks\": [\n {\n \"index\": 0,\n \"hash\": \"\",\n \"shares\": [\n {\"id\": \"\", \"server\": \" https://blossom1.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom2.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom3.example.com\"}\n ]\n },\n {\n \"index\": 1,\n \"hash\": \"\",\n \"shares\": [\n {\"id\": \"\", \"server\": \" https://blossom1.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom2.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom3.example.com\"}\n ]\n }\n ],\n \"erasure\": {\n \"algorithm\": \"reed-solomon\",\n \"k\": 2,\n \"n\": 3,\n \"field\": \"gf256\"\n }\n}\n```\n\nThe `key` field contains the per-file encryption key, itself encrypted using NIP-44 to the owner's public key. This double encryption-file content encrypted to the file key, file key encrypted to the owner key-enables potential future extensions where file keys could be shared with other parties without exposing the master key.\n\nThe `hash` field in each block entry contains the SHA-256 hash of the plaintext block before encryption. This enables integrity verification after decryption: if the decrypted block's hash doesn't match, either the ciphertext was corrupted, the wrong key was used, or the inode itself is corrupt.\n\nThe `shares` array maps share indices to their content addresses and storage locations. During reconstruction, the client attempts to fetch shares in order, stopping once k shares are successfully retrieved.\n\nInodes themselves are stored as blobs following the identical pipeline. An inode is serialized, padded, encrypted, erasure-coded, and distributed. The resulting structure is referenced by its content hash, forming a node in the Merkle DAG.\n\n---\n\n## 8. Directory Hierarchy\n\n### 8.1 Directories as Encrypted Blobs\n\nA directory is a file whose decrypted contents enumerate named entries and their corresponding inode references. After decryption, a directory blob contains:\n\n```json\n{\n \"version\": 1,\n \"type\": \"directory\",\n \"created\": 1701820800,\n \"modified\": 1701907200,\n \"entries\": {\n \"photos\": {\n \"type\": \"directory\",\n \"inode\": \"\"\n },\n \"documents\": {\n \"type\": \"directory\",\n \"inode\": \"\"\n },\n \"notes.txt\": {\n \"type\": \"file\",\n \"inode\": \"\"\n }\n }\n}\n```\n\nThis directory blob is encrypted and stored identically to file inodes-same block size, same encryption, same erasure coding. From a server's perspective, all blobs are indistinguishable.\n\nEntry names are stored within the encrypted blob, invisible to servers. An observer sees only uniform encrypted blocks, with no information about file count, names, or directory structure. They see only a collection of uniform encrypted blocks.\n\n### 8.2 Merkle DAG Structure\n\nThe directory hierarchy forms a Merkle Directed Acyclic Graph (DAG). Each node-whether file inode or directory-is identified by the content hash of its encrypted representation. Parent nodes contain the hashes of their children.\n\n```\n ┌─────────────────┐\n │ Root Directory │\n │ hash: 0xABC │\n └────────┬────────┘\n │\n ┌────────────────┼────────────────┐\n │ │ │\n ▼ ▼ ▼\n ┌───────────────┐ ┌───────────────┐ ┌───────────────┐\n │ photos/ │ │ documents/ │ │ notes.txt │\n │ hash: 0xDEF │ │ hash: 0x123 │ │ hash: 0x456 │\n └───────┬───────┘ └───────┬───────┘ └───────────────┘\n │ │\n ▼ ▼\n ┌───────────────┐ ┌───────────────┐\n │ image1.jpg │ │ report.pdf │\n │ hash: 0x789 │ │ hash: 0xFED │\n └───────────────┘ └───────────────┘\n```\n\nThis structure provides several important properties. Any node's hash authenticates its entire subtree-if an attacker modifies any descendant, the hashes will not match during traversal. The structure can be verified incrementally; a client can validate a path from root to a specific file without fetching the entire tree. Unchanged subtrees share storage; updating one file doesn't require re-uploading siblings.\n\n### 8.3 Path Resolution\n\nTo resolve a path like `/photos/image1.jpg`, the client:\n\n1. Obtains the root directory hash from the current chain head\n2. Fetches and decrypts the root directory blob\n3. Looks up \"photos\" in the entries, obtaining hash 0xDEF\n4. Fetches and decrypts the photos directory blob \n5. Looks up \"image1.jpg\" in the entries, obtaining hash 0x789\n6. Fetches and decrypts the image1.jpg inode\n7. Uses the inode to fetch, decode, decrypt, and reassemble the file\n\nEach step requires fetching k shares and then decoding and decrypting. The total number of blob fetches equals the path depth plus one (for the file inode). Deep hierarchies incur proportionally more round trips.\n\n---\n\n## 9. State Management via Hash Chain\n\n### 9.1 The Problem with Mutable Pointers\n\nA content-addressed storage system requires at least one mutable pointer to locate the current root. Without such a pointer, clients would need to know the root hash through some out-of-band mechanism, and that hash would change with every update.\n\nA naive approach uses a Nostr replaceable event (kinds 30000-39999) to store the current root hash. The most recent event for a given (kind, pubkey, d-tag) tuple supersedes all previous events. This works but has limitations.\n\nReplaceable events provide no history. Once superseded, the previous root is lost unless the client independently preserved it. There is no audit trail showing when changes occurred or what the previous states were. If two clients simultaneously commit changes, one will silently overwrite the other with no indication of conflict.\n\nConflict detection is impossible. If device A commits while device B is offline, and then device B commits without fetching A's update, B's commit will reference an outdated parent state. With simple replaceable events, this produces silent data loss-A's changes vanish without warning.\n\n### 9.2 Hash Chain of Commits\n\nThis design employs a hash chain of commit events, building an ordered history from the first commit forward. Each commit references its predecessor, forming a cryptographically-linked sequence analogous to a blockchain or git history.\n\nA commit event has the following structure:\n\n```json\n{\n \"kind\": 30097,\n \"pubkey\": \"\",\n \"created_at\": 1701907200,\n \"tags\": [\n [\"d\", \"storage-chain\"],\n [\"prev\", \"\"],\n [\"root\", \"\"],\n [\"seq\", \"42\"]\n ],\n \"content\": \"\",\n \"sig\": \"\"\n}\n```\n\nThe `prev` tag contains the event ID of the immediately preceding commit, creating the chain. The `seq` tag provides a monotonically increasing sequence number for quick ordering. The `root` tag contains the content hash of the current root directory blob.\n\nThe encrypted `content` field contains additional metadata:\n\n```json\n{\n \"root_inode\": {\n \"hash\": \"\",\n \"shares\": [\n {\"id\": \"\", \"server\": \" https://blossom1.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom2.example.com\"},\n {\"id\": \"\", \"server\": \" https://blossom3.example.com\"}\n ]\n },\n \"erasure\": {\"k\": 2, \"n\": 3},\n \"garbage\": [\"\", \"\"],\n \"message\": \"Added vacation photos\"\n}\n```\n\nThe `garbage` array lists blob hashes that are no longer referenced as of this commit and may be deleted. The optional `message` field allows human-readable commit descriptions.\n\n### 9.3 Commit Process\n\nCreating a new commit follows this sequence:\n\n1. Fetch the current chain head from Nostr relays\n2. Verify local changes are based on this head (detect conflicts)\n3. Upload all new blobs (file blocks, inodes, directories)\n4. Construct the new root directory referencing updated content\n5. Upload the new root directory blob\n6. Create a commit event with `prev` pointing to the fetched head\n7. Sign and publish the commit event to Nostr relays\n\nIf step 2 reveals that the local state diverges from the chain head-because another device committed in the interim-the client must reconcile before proceeding. Reconciliation strategies include:\n\n- **Abort**: Discard local changes, fetch remote state, let user redo changes\n- **Merge**: If changes affect disjoint subtrees, automatically merge\n- **Fork**: Create a branch, defer reconciliation to user\n\nThe appropriate strategy depends on the application. For personal backup, aborting with user notification is often sufficient. More sophisticated applications might implement git-like merging.\n\n### 9.4 Snapshot-Based Workflow\n\nThe hash chain naturally supports an explicit save model. Users accumulate changes locally-adding files, modifying documents, reorganizing directories-without network activity. These changes exist only on the local device.\n\nWhen the user explicitly saves (clicks a button, invokes a command), the client:\n\n1. Collects all pending local changes\n2. Uploads the changed blobs\n3. Publishes a single commit event encompassing all changes\n\nThis batching reduces network traffic and avoids intermediate states, giving users clear checkpoints. The resulting history shows meaningful snapshots with meaningful descriptions (\"Added tax documents for 2024\"), discarding the stream of micro-changes.\n\nBetween saves, the local state may be lost if the device fails. This is acceptable for a backup-oriented system-unsaved changes are analogous to unsaved edits in a document editor. Users who want continuous protection should save frequently.\n\n### 9.5 Chain Traversal and History\n\nThe complete history is recoverable by walking the chain backward from the head. Each commit's `prev` tag leads to its predecessor until reaching the genesis commit (which has no `prev` tag or a null value).\n\n```\nHEAD ──prev──► Commit N-1 ──prev──► Commit N-2 ──prev──► ... ──prev──► Genesis\n```\n\nClients can implement time-travel functionality: given any historical commit, they can reconstruct the exact filesystem state at that point by using the commit's root hash to traverse the Merkle DAG.\n\nThis history has storage implications. Old commits reference old blobs which must be retained for history to remain valid. Users who don't need history can garbage collect aggressively. Users who value history must retain more data. Section 13 discusses garbage collection in detail.\n\n---\n\n## 10. Single-Key Discovery and Recovery\n\n### 10.1 Recovery Process\n\nDisaster recovery requires only the owner's Nostr secret key (nsec). No backup files, no secondary credentials, no trusted third party. The recovery process:\n\n1. **Derive public key**: Compute npub from nsec using secp256k1\n2. **Query relays**: Request kind 30097 events with author = npub and d-tag = \"storage-chain\"\n3. **Find chain head**: Identify the event with highest sequence number (or most recent created_at)\n4. **Derive master key**: Compute master storage key from nsec via HKDF\n5. **Decrypt commit**: Decrypt the commit's content field using NIP-44\n6. **Fetch root**: Download k shares of the root directory inode using URLs from the commit\n7. **Decode and decrypt**: Erasure-decode and decrypt the root directory\n8. **Traverse**: Recursively fetch any desired files through the directory structure\n\nThe Nostr relay network serves as the discovery layer. Relays are interchangeable-the client can query any relay that might have stored the owner's events. Since events are signed, their authenticity is verifiable regardless of which relay provides them.\n\n### 10.2 Relay Selection\n\nRecovery reliability depends on commit events being retrievable from at least one relay. Users should publish commits to multiple relays and periodically verify that relays still hold their events.\n\nRelay selection strategies include:\n\n- **Personal relays**: Relays the user operates or trusts, likely to retain events long-term\n- **Paid relays**: Commercial relays with retention guarantees\n- **Popular public relays**: High availability but may prune old events\n\nThe commit events are small (a few kilobytes) and don't grow with dataset size. Storing them across many relays is inexpensive and dramatically improves recovery reliability.\n\n### 10.3 Blob Server Discovery\n\nCommit events contain URLs for the servers storing each blob's shares. If a server disappears, its URL becomes invalid, but the blob can still be reconstructed from surviving shares on other servers.\n\nThe system doesn't specify a discovery mechanism for finding new servers. In practice:\n\n- Users manually configure a list of preferred servers\n- Users may operate their own Blossom server for guaranteed availability\n- Future extensions could use a Nostr-based server registry\n\nIf insufficient shares remain available for any blob, that blob is unrecoverable. Users should monitor server health and re-upload shares to replacement servers before reaching this point.\n\n---\n\n## 11. Transport Layer\n\n### 11.1 Blossom Protocol\n\nBlossom servers provide content-addressed blob storage over HTTP, authenticated via Nostr events. The protocol is intentionally minimal-servers store bytes and retrieve bytes, nothing more.\n\nCore endpoints:\n\n| Method | Path | Description |\n|--------|------|-------------|\n| GET | `/{sha256}` | Retrieve blob by hash |\n| HEAD | `/{sha256}` | Check blob existence |\n| PUT | `/upload` | Store a new blob |\n| DELETE | `/{sha256}` | Remove a blob |\n| GET | `/list/{pubkey}` | List blobs uploaded by a pubkey |\n\nThe `{sha256}` path component is the lowercase hex-encoded SHA-256 hash of the blob's contents. An optional file extension may be appended (e.g., `/{sha256}.pdf`) for MIME type hinting, but servers identify blobs solely by hash.\n\n### 11.2 Authentication\n\nWrite operations (PUT, DELETE) require authentication via a Nostr event included in the Authorization header:\n\n```\nAuthorization: Nostr \n```\n\nThe authorization event has kind 24242 and must include:\n\n```json\n{\n \"kind\": 24242,\n \"created_at\": 1701907200,\n \"tags\": [\n [\"t\", \"upload\"],\n [\"x\", \"\"],\n [\"expiration\", \"1701910800\"]\n ],\n \"content\": \"Uploading backup data\",\n \"sig\": \"\"\n}\n```\n\nThe `t` tag specifies the authorized action: \"upload\", \"delete\", or \"list\". The `x` tag binds the authorization to a specific blob hash, preventing replay attacks. The `expiration` tag limits the authorization's validity window.\n\nServers verify the signature, confirm the kind is 24242, check that the action matches the `t` tag, validate that the current time is before expiration, and for uploads/deletes, verify the `x` tag matches the blob hash.\n\n### 11.3 Server Responses\n\nSuccessful upload returns a blob descriptor:\n\n```json\n{\n \"url\": \" https://cdn.blossom.example/abc123def456...\",\n \"sha256\": \"abc123def456...\",\n \"size\": 262144,\n \"type\": \"application/octet-stream\",\n \"uploaded\": 1701907200\n}\n```\n\nThe URL may differ from the upload endpoint if the server uses a CDN or different domain for retrieval.\n\nGET requests return the raw blob bytes with appropriate headers:\n\n```\nContent-Type: application/octet-stream\nContent-Length: 262144\nX-Content-Sha256: abc123def456...\n```\n\nHEAD requests return the same headers without the body, enabling existence checks without downloading content.\n\n### 11.4 Server Interchangeability\n\nBlossom servers are interchangeable and fungible. A blob uploaded to server A can be retrieved from server B if server B also has it. The content hash serves as a universal identifier across all servers.\n\nThis interchangeability enables several patterns:\n\n- **Mirroring**: Upload the same blob to multiple servers for redundancy\n- **Migration**: Move from one server to another by re-uploading\n- **CDN integration**: Servers can replicate blobs to edge locations\n- **Opportunistic caching**: Clients can check multiple servers and use whichever responds fastest\n\nThe system's erasure coding distributes shares across servers, so migration requires uploading only shares to replacement servers, not the full reconstructed blob.\n\n---\n\n## 12. Privacy-Preserving Existence Queries\n\n### 12.1 The Query Privacy Problem\n\nBefore uploading a blob, a client might want to check if the server already has it-to avoid redundant uploads and save bandwidth. The naive approach sends a HEAD request with the blob's hash. But this reveals information: the server learns which hashes the client is interested in.\n\nOver time, a curious server could build a profile of which blobs a client accesses and when, identifying blobs that are queried in close temporal proximity. Even though blob contents are encrypted, access patterns leak information.\n\n### 12.2 Binary Fuse Filters\n\nBinary fuse filters enable existence checking without revealing the query. The server periodically publishes a compact filter representing the set of all blob hashes it stores. Clients download this filter and perform membership tests locally.\n\nA binary fuse filter is a probabilistic data structure supporting set membership queries. For a set of n elements, the filter consumes approximately 9n bits (just over 1 byte per element) and answers \"is x in the set?\" with:\n\n- **No false negatives**: If x is in the set, the filter always returns true\n- **Tunable false positives**: The filter may incorrectly return true for elements not in the set, with probability configurable to ~1%\n\nThe query algorithm computes three hash functions of the queried element, XORs the filter values at those positions, and compares to a fingerprint:\n\n```\nquery(x):\n h0 = hash0(x) mod segment_length\n h1 = hash1(x) mod segment_length + segment_length\n h2 = hash2(x) mod segment_length + 2 * segment_length\n return filter[h0] XOR filter[h1] XOR filter[h2] == fingerprint(x)\n```\n\nCritically, this computation happens entirely on the client. The server never sees the query input.\n\n### 12.3 Filter Distribution\n\nServers generate and publish their filters periodically, hourly or daily depending on churn rate. A filter for a server with 1 million blobs requires approximately 1.1 MB, small enough for clients to download and cache.\n\nThe filter might be distributed via:\n\n- HTTP endpoint on the Blossom server (`GET /filter`)\n- Nostr event published by the server's identity\n- Out-of-band distribution (IPFS, BitTorrent, etc.)\n\nClients download filters for their configured servers, then perform local lookups when deciding whether to upload. If the filter indicates a blob might exist (accounting for false positives), the client can either skip uploading or perform a HEAD request to confirm.\n\n### 12.4 Limitations\n\nBinary fuse filters are immutable once constructed. They cannot be updated incrementally-adding a new element requires rebuilding the entire filter. This is acceptable for periodic batch updates but prevents real-time synchronization.\n\nFalse positives mean clients occasionally skip uploads for blobs the server lacks. The subsequent fetch will fail, exposing the false positive; the client can then upload. With a 1% false positive rate, this occurs rarely and has minimal impact.\n\nThe filter reveals the total number of blobs the server stores (derivable from filter size). It doesn't reveal which specific blobs or any information about their contents.\n\n---\n\n## 13. Garbage Collection\n\n### 13.1 The Accumulation Problem\n\nContent-addressed immutable storage naturally accumulates data. Updating a file creates new blobs; the old blobs persist. The directory structure uses copy-on-write semantics, so modifying a deeply nested file creates new blobs for every ancestor directory up to the root. Without cleanup, storage consumption grows monotonically even if the logical dataset size remains constant.\n\nThis design places garbage collection responsibility entirely with the client. The system does not automatically delete anything. Users must explicitly choose to delete obsolete data, accepting the tradeoff between storage costs and history preservation.\n\n### 13.2 Reference Tracking\n\nThe client maintains knowledge of which blobs are reachable from each commit. A blob is garbage if it's unreachable from any commit the user wishes to preserve.\n\nComputing reachability requires traversing the Merkle DAG from each preserved commit's root. For a given set of root hashes, reachable blobs are:\n\n```\nreachable = {}\nfor root in preserved_roots:\n traverse(root, reachable)\n\ndef traverse(hash, reachable):\n if hash in reachable:\n return\n reachable.add(hash)\n blob = fetch_and_decrypt(hash)\n for child_hash in extract_references(blob):\n traverse(child_hash, reachable)\n```\n\nBlobs not in the reachable set are candidates for deletion.\n\n### 13.3 Deletion Strategies\n\nSeveral strategies for garbage collection exist, offering different tradeoffs:\n\n**Keep everything**: Never delete blobs. Storage grows unboundedly, but complete history is preserved. Suitable for archival use cases where history has intrinsic value.\n\n**Keep recent history**: Preserve the last N commits or commits from the last M days. Delete blobs unreachable from this window. Balances storage cost against useful history depth.\n\n**Keep only current**: Preserve only the chain head. Delete all blobs unreachable from the current state. Minimizes storage but loses all history. Recovery options are limited if the current state is corrupted.\n\n**Explicit snapshots**: Mark specific commits as preserved (e.g., monthly snapshots, pre-migration backups). Delete blobs unreachable from any preserved commit.\n\n### 13.4 Deletion Process\n\nTo delete garbage blobs:\n\n1. Compute the set of blob hashes to delete\n2. For each hash and each server storing a share of that blob:\n - Generate a deletion authorization event\n - Send DELETE request with authorization\n3. Update the garbage list in the next commit to reflect completed deletions\n\nThe commit's `garbage` field serves as an announcement of intent. It signals to future clients examining history that these blobs were deliberately deleted and should not be considered missing or corrupted.\n\nDeletion authorization uses the same Nostr event mechanism as uploads:\n\n```json\n{\n \"kind\": 24242,\n \"tags\": [\n [\"t\", \"delete\"],\n [\"x\", \"\"],\n [\"expiration\", \"1701910800\"]\n ],\n \"content\": \"Garbage collection\",\n \"sig\": \"\"\n}\n```\n\n### 13.5 Metadata Event Garbage Collection\n\nThe hash chain of commit events also accumulates over time. Old commit events may be pruned from relays to reduce storage, but this requires care.\n\nSafe deletion criteria for commit events:\n\n- The commit's blobs have been garbage collected (no point keeping metadata for deleted data)\n- The commit is not the chain head or a preserved snapshot\n- Sufficient time has passed that no client might be traversing through it\n\nIn practice, commit events are small (kilobytes) and relay storage is cheap. Most users can retain their complete commit history indefinitely. Users with extremely long histories or storage-constrained relays can prune old commits, accepting that history before the pruning point becomes inaccessible.\n\n---\n\n## 14. Lifecycle Summary\n\n### 14.1 Initial Setup\n\nA new user performs one-time setup:\n\n1. Generate or import a Nostr keypair (nsec/npub)\n2. Derive the master storage key from nsec\n3. Configure preferred Blossom servers\n4. Configure preferred Nostr relays\n5. Create an empty root directory\n6. Publish the genesis commit event\n\n### 14.2 Adding Files\n\nTo add a file to the storage system:\n\n1. Read the file content\n2. Divide into fixed-size blocks with padding\n3. Generate a random per-file encryption key\n4. For each block:\n - Derive the block encryption key\n - Encrypt with XChaCha20-Poly1305\n - Erasure-code into n shares\n - Upload shares to n servers\n5. Construct the inode with block metadata\n6. Encrypt and upload the inode blob\n7. Update the parent directory to include the new entry\n8. Recursively update ancestors to the root\n9. Stage changes for the next commit\n\n### 14.3 Committing Changes\n\nWhen the user saves:\n\n1. Fetch current chain head from relays\n2. Verify local changes are based on this head\n3. Upload all staged blobs (files, inodes, directories)\n4. Construct commit event with new root and prev reference\n5. Sign and publish commit to relays\n6. Clear local staged changes\n\n### 14.4 Reading Files\n\nTo read a file by path:\n\n1. Fetch current chain head\n2. Decrypt commit to obtain root blob location\n3. Traverse directory structure to target inode\n4. For each block in the inode:\n - Attempt to fetch k shares from listed servers\n - Erasure-decode to recover encrypted block\n - Decrypt with derived block key\n5. Concatenate blocks and remove padding\n6. Return file contents\n\n### 14.5 Verification and Repair\n\nPeriodically, the client should verify data availability:\n\n1. For each blob referenced by the current state:\n - For each share of that blob:\n - Send HEAD request to check existence\n - Optionally, GET and verify hash matches\n2. If any blob has fewer than k available shares:\n - Fetch k surviving shares\n - Erasure-decode to recover the block\n - Re-encode to generate missing shares\n - Upload replacement shares to new servers\n - Update inode with new share locations\n - Commit the updated inodes\n\n### 14.6 Garbage Collection\n\nWhen storage costs warrant cleanup:\n\n1. Decide which commits to preserve\n2. Compute reachable blob set from preserved commits\n3. Identify unreachable blobs\n4. Delete unreachable blobs from servers\n5. Optionally delete obsolete commit events from relays\n6. Record garbage collection in next commit\n\n---\n\n## 15. Security Analysis\n\n### 15.1 Confidentiality\n\nStorage servers observe only uniformly-sized encrypted blobs. They cannot determine:\n\n- File contents (encrypted with XChaCha20-Poly1305)\n- File sizes (obscured by fixed block padding)\n- File types (all blocks are indistinguishable)\n- Filenames (stored in encrypted directory blobs)\n- Directory structure (directories are encrypted like files)\n- Relationships between blobs (no plaintext linking)\n\nThe encryption is semantic-identical plaintexts produce different ciphertexts due to random per-file keys. Servers cannot detect when users store the same content.\n\n### 15.2 Integrity\n\nContent addressing provides integrity at multiple levels. Share hashes verify individual share integrity. Block hashes (stored in inodes) verify decrypted block integrity. The Merkle DAG structure verifies structural integrity-any modification to any blob changes the root hash.\n\nPoly1305 authentication tags detect ciphertext tampering. Even if an attacker modifies stored ciphertext in a way that produces a valid hash, decryption will fail authentication.\n\n### 15.3 Availability\n\nErasure coding ensures availability despite server failures. With (k, n) parameters, data survives the loss of any n - k servers. The hash chain ensures commit history survives relay churn as long as at least one relay retains the events.\n\nThe system does not provide availability against censorship or targeted attacks where adversaries deliberately destroy more than n - k shares simultaneously.\n\n### 15.4 Authentication\n\nNostr signatures authenticate all state changes. Only the holder of the nsec can publish valid commit events. Blossom authorization events prevent unauthorized uploads or deletions.\n\nRelays and servers can verify signature validity but cannot forge signatures. A compromised relay could refuse to serve events (availability attack) but cannot produce fake commits (integrity preserved).\n\n### 15.5 Key Compromise\n\nIf the nsec is compromised, all security properties fail. The attacker can:\n\n- Decrypt all current and historical data\n- Publish malicious commits\n- Delete data from servers\n- Irrevocably destroy the dataset\n\nKey management is outside this system's scope. Users should employ standard practices: hardware security modules, secure backup procedures, passphrase protection.\n\n---\n\n## 16. Future Considerations\n\n### 16.1 Payment Integration\n\nStorage servers require compensation for resources consumed. Integration with payment systems would enable sustainable server operation.\n\nPossibilities include per-byte pricing with Lightning Network micropayments, subscription models with ecash or traditional payment, and storage markets where servers compete on price and reliability. Payment integration should not compromise privacy-payments should not link to specific blobs or reveal access patterns.\n\n### 16.2 Steward Services\n\nA steward service could handle verification and repair automatically, running continuously without user intervention. This requires delegating sufficient authority to perform repairs without granting full account access.\n\nPotential approaches include threshold signatures where the steward holds one share of a multisig, capability tokens authorizing specific repair actions, and read-only access combined with user approval for repairs. Steward design involves complex trust tradeoffs and is deferred to future work.\n\n### 16.3 Multi-Device Synchronization\n\nThe current design supports multiple devices through the commit chain, but conflict resolution is minimal. Enhanced multi-device support might include automatic merging for non-conflicting changes, three-way merge for file-level conflicts, operational transformation for collaborative editing, and CRDT-based structures for specific data types.\n\n### 16.4 Deduplication\n\nContent addressing naturally deduplicates identical files-they hash to the same blob. Block-level deduplication across files is more complex. Content-defined chunking using rolling hashes (Rabin fingerprinting) could identify common blocks across similar files, reducing storage for versioned documents or near-duplicates.\n\n### 16.5 Proof of Retrievability\n\nMore sophisticated cryptographic proofs could enable efficient verification without downloading data. Proof of Retrievability (PoR) schemes allow servers to prove they hold data by responding to challenges. This could reduce verification bandwidth from O(data size) to O(security parameter).\n\n---\n\n## 17. Conclusion\n\nThis design provides a practical architecture for durable, private, personal storage built on existing Nostr and Blossom infrastructure. A layered architecture separates concerns: fixed-size blocks provide privacy through uniformity, erasure coding provides durability through redundancy, authenticated encryption provides confidentiality and integrity. Efficient updates come from the Merkle DAG, with auditable history and conflict detection from the hash chain.\n\nThe system achieves its core requirements. Durability is provided through erasure coding-data survives arbitrary server failures up to the configured threshold. Storage providers learn nothing about content, sizes, structure, or access patterns beyond gross storage volume. Sovereignty is preserved-users control when changes commit and when old data is deleted. Recoverability is complete-the entire dataset and its history can be reconstructed from a single secret key.\n\nImmutable, content-addressed blobs simplify consistency and enable deterministic caching. An explicit save model gives users clear checkpoints, avoiding the complexity of real-time synchronization. History and conflict detection come from the hash chain, with no trusted timestamps or consensus required.\n\nSignificant work remains for production deployment. Payment integration and automated maintenance are areas for future development. But the core architecture presented here provides a solid foundation for personal data storage that respects user sovereignty while leveraging decentralized infrastructure.\n\n---\n\n## References\n\n1. Nostr Protocol. NIP-01: Basic Protocol Flow Description. https://github.com/nostr-protocol/nips/blob/master/01.md\n\n2. Blossom Protocol. BUD-01: Server Specification. https://github.com/hzrd149/blossom\n\n3. IETF RFC 8439. ChaCha20 and Poly1305 for IETF Protocols. https://tools.ietf.org/html/rfc8439\n\n4. IETF RFC 5869. HMAC-based Extract-and-Expand Key Derivation Function (HKDF). https://tools.ietf.org/html/rfc5869\n\n5. IETF RFC 5510. Reed-Solomon Forward Error Correction (FEC) Schemes. https://tools.ietf.org/html/rfc5510\n\n6. Graf, T. M., & Lemire, D. (2022). Binary Fuse Filters: Fast and Smaller Than Xor Filters. ACM Journal of Experimental Algorithmics.\n\n7. BIP-340. Schnorr Signatures for secp256k1. https://bips.dev/340/\n\n8. Nostr Protocol. NIP-44: Encrypted Payloads (Versioned). https://github.com/nostr-protocol/nips/blob/master/44.md\n\n---\n\n## Appendix A: notation reference\n\n| Symbol | Meaning |\n|--------|---------|\n| B | Block size in bytes (typically 262,144) |\n| k | Erasure coding: minimum shares for reconstruction |\n| n | Erasure coding: total shares per block |\n| nsec | Nostr secret key (32 bytes) |\n| npub | Nostr public key (32-byte x-coordinate) |\n| GF(2^8) | Galois Field with 256 elements |\n| H(x) | SHA-256 hash function |\n| HKDF | HMAC-based Key Derivation Function |\n| XChaCha20 | Extended-nonce ChaCha20 stream cipher |\n| Poly1305 | One-time authenticator MAC |\n\n## Appendix B: recommended parameters\n\n**Block size**: 256 KiB (262,144 bytes) \nProvides good balance between padding overhead and chunking granularity.\n\n**Erasure coding**: (k=3, n=5) \nTolerates 2 server failures with 67% storage overhead. Requires 5 independent servers.\n\n**Encryption**: XChaCha20-Poly1305 \n192-bit nonces enable safe random generation. Software performance excellent without hardware acceleration.\n\n**Key derivation**: HKDF-SHA256 \nStandard, well-analyzed, widely implemented.\n\n**Commit distribution**: 5+ Nostr relays \nEnsures commit retrievability despite relay failures.\n\n**Verification interval**: Weekly \nBalance between catching failures early and minimizing bandwidth.\n\n","created_at",1777543606,"id","18411b3f6a839c5fb446dae355a38fe08858e402ef23ad60b4613f009363e704","kind",30023,"pubkey","b7ed68b062de6b4a12e51fd5285c1e1e0ed0e5128cda93ab11b4150b55ed32fc","sig","01e1f0635deac264339f7be9f1943a49a2388b494d82aef9ad5389cd7cc83155a6fa68df1c5385d6d160a5a881e03ca79130243bf673f1e419bd32a0f030b647","tags",[28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43],[63,64],[61,62],[59,60],[57,58],[55,56],[44,54],[44,53],[44,52],[44,51],[44,50],[44,49],[44,48],[44,47],[44,46],[44,45],[44,12],"t","file-system","self-sovereignty","decentralization","cryptography","encryption","blossom","nostr","garland","freedom-tech","austrian-economics","published_at","1765093221","summary","The latest iteration of Garland, a privacy-preserving distributed storage system built on Nostr and Blossom, using erasure coding for durability and single-key recovery.","title","Nostr-Native Distributed Storage System","image","https://image.nostr.build/92c1ebd52e5169812978639f2825dea4bc85ab6418d6cfbbc36004c6872691e0.jpg","d","c4d2a3114166f8f5"]