<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Tomoya Nagasawa</title><description>Research and papers by Tomoya Nagasawa: data encodings, databases, zero-knowledge proofs, formal verification and AI.</description><link>https://tn.ms/</link><language>en</language><atom:link href="https://tn.ms/rss.xml" rel="self" type="application/rss+xml"/><item><title>A log-structured store across trust tiers</title><link>https://tn.ms/research/tiered-lsm-store/</link><guid isPermaLink="true">https://tn.ms/research/tiered-lsm-store/</guid><description>A key-value store whose layers sit on tiers of different speed and trust: a volatile tier that answers at once, a durable ordered log, and immutable segments. Reads equal the sequential map, flush and compaction preserve it, tombstones may be dropped exactly at the bottom level, re-encoding with any lossless codec is safe, recovery is exact, finalised state is never rolled back, and a wrong state root has one-step evidence.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Log-structured merge trees</category><category>Write-ahead logging</category><category>Replay verification</category></item><item><title>Demurrage at constant cost</title><link>https://tn.ms/research/demurrage-share-accounting/</link><guid isPermaLink="true">https://tn.ms/research/demurrage-share-accounting/</guid><description>A holding cost on money, applied through share accounting: only the total supply decays, so every holder is charged in proportion at constant cost per operation. Decay is independent of when updates happen exactly when it is geometric; a linear per-interval rule rewards frequent updates by exactly r²ab·S; floor rounding stays within one unit per operation and never pays out more than the supply; supply plus collected pool is conserved.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Demurrage</category><category>Share accounting</category><category>Functional equations</category></item><item><title>Upgrading everything but the store</title><link>https://tn.ms/research/state-logic-upgrades/</link><guid isPermaLink="true">https://tn.ms/research/state-logic-upgrades/</guid><description>An architecture in which one persistent key-value store is fixed and every other component, including the registry that names components, can be replaced. Upgrades never touch stored state; only setters write; flattened keys are faithful exactly when the key encoding is injective; positional layouts collide where keyed ones cannot; a two-step role transfer opens a stale-writer window that an atomic or revoke-first transfer closes; registry replacement keeps every name resolvable under a stated ordering; and the set of components that can change a binding is a reachability closure.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Dynamic software updating</category><category>Access control</category><category>Key-value storage</category></item><item><title>Paper: Bit-Level Columnar Coding of JSON Trees with Append-Only Delta Chains</title><link>https://tn.ms/papers/bit-columnar-coding.pdf</link><guid isPermaLink="true">https://tn.ms/papers/bit-columnar-coding.pdf</guid><description>A JSON encoding that drops byte alignment for variable-width bit fields. One pre-order scan splits a document into twelve column streams coded with width-selecting integer codes, a ±3 delta window and run-length packing; the only count written is the number of leaves. The integer codes are proved prefix-free with closed-form lengths, and the slot and leaf sequences are proved to determine the tree. The same grammar serves as an update format whose length-framed payload chains replay to every version. Sizes computed from the specification on seeded synthetic data: arrays of records take 52–55% of the size of the shortest MessagePack encoding; an interleaved time series is 21% larger.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>paper</category></item><item><title>Paper: Canonical Arithmetization of JSON Documents, with Authenticated Stores and Zero-Knowledge Path Queries</title><link>https://tn.ms/papers/canonical-json-arithmetization.pdf</link><guid isPermaLink="true">https://tn.ms/papers/canonical-json-arithmetization.pdf</guid><description>A canonical encoding of JSON values into elements of the BN254 scalar field, proved injective on a precisely stated domain by giving every stage a left inverse: flattening into (path, leaf) rows with empty containers as leaves, shortlex-sorted paths, type-tagged integers, length-prefixed decimal tokens with run compression, and 75-digit blocks behind a guard digit. On this encoding, path queries and non-existence are defined as relations over a commitment and decided exactly by a walk over the token grammar; the paper states what a circuit must constrain for soundness, with an analytic constraint estimate, and specifies a two-level sparse Merkle store with an injective radix map from ids to leaves. All encoding results are machine-checked in Lean 4.</description><pubDate>Thu, 12 Nov 2026 00:00:00 GMT</pubDate><category>paper</category></item><item><title>Small integers at bit granularity</title><link>https://tn.ms/research/integer-codes/</link><guid isPermaLink="true">https://tn.ms/research/integer-codes/</guid><description>Length functions, prefix-freeness and Kraft sums for 2-bit selector integer codes, compared with LEB128 and Elias-γ, and what they cost under a geometric source.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Integer codes</category><category>Prefix codes</category><category>Kraft inequality</category></item><item><title>Trees as columns</title><link>https://tn.ms/research/trees-as-columns/</link><guid isPermaLink="true">https://tn.ms/research/trees-as-columns/</guid><description>A JSON tree written as two pre-order pointer arrays plus typed value columns, packed at the bit level, and an update scheme in which each edit is a small payload addressed against the previous materialized state.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Tree encoding</category><category>Columnar layout</category><category>Delta coding</category><category>Incremental update</category></item><item><title>Arithmetizing JSON</title><link>https://tn.ms/research/arithmetizing-json/</link><guid isPermaLink="true">https://tn.ms/research/arithmetizing-json/</guid><description>A canonical, injective encoding of arbitrary JSON into a short vector of field elements, so that an arithmetic circuit can prove the value at one path of a committed document without revealing the rest.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Canonical encoding</category><category>Arithmetic circuits</category><category>Zero-knowledge proofs</category></item><item><title>Authenticated state with sparse Merkle trees</title><link>https://tn.ms/research/authenticated-state/</link><guid isPermaLink="true">https://tn.ms/research/authenticated-state/</guid><description>A document store committed to by one hash: fixed-depth sparse Merkle trees, an injective map from string ids to leaf indices, batched state-transition proofs and value or predicate queries checked against a single stored root.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Sparse Merkle trees</category><category>Validity proofs</category><category>Arithmetic circuits</category></item><item><title>Programs as data: point-free combinators in JSON</title><link>https://tn.ms/research/programs-as-data/</link><guid isPermaLink="true">https://tn.ms/research/programs-as-data/</guid><description>A small combinator language whose programs are JSON arrays: its reduction rules, what its combinators can and cannot express, and why programs without variables are easy to store, hash and evaluate deterministically.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Combinatory logic</category><category>Point-free style</category><category>Homoiconicity</category></item><item><title>Composing effects: monads and Kleisli arrows</title><link>https://tn.ms/research/kleisli-composition/</link><guid isPermaLink="true">https://tn.ms/research/kleisli-composition/</guid><description>Effectful pipelines as Kleisli arrows: the monad laws are exactly the category laws for arrow composition, which makes it possible to build a pipeline before any value exists, run it in a synchronous or an asynchronous monad, and refactor it without changing what it computes.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Monads</category><category>Kleisli category</category><category>Program composition</category></item><item><title>Treasury-backed issuance as a state machine</title><link>https://tn.ms/research/treasury-backed-issuance/</link><guid isPermaLink="true">https://tn.ms/research/treasury-backed-issuance/</guid><description>A token whose supply is backed by a treasury, modelled as a four-field state machine: a floor invariant that survives every operation for any monotone floor curve, backing that never decreases, an exact bound on what a fee sandwich can earn, a dollar cap with a tight drawdown theorem, a criterion for when a cubic Bézier curve is a function and where it turns vertical, and why a proportional contest for a fixed prize cannot clear at backing. Every combinatorial and arithmetic statement is proved in Lean; the real-analysis steps of the graph criterion are standard and stated as such.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Mechanism design</category><category>Invariants</category><category>Bézier curves</category><category>Tullock contests</category></item><item><title>Keys that are never stored</title><link>https://tn.ms/research/keys-never-stored/</link><guid isPermaLink="true">https://tn.ms/research/keys-never-stored/</guid><description>A wallet key re-derived on demand from factors: derivability sets as monotone access structures, AND/OR key combination realising every monotone formula, a threshold key-derivation service, a two-of-three construction, and what a pairing check does and does not give, all in a symbolic model checked in Lean.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Access structures</category><category>Threshold cryptography</category><category>Symbolic protocol models</category></item><item><title>Checking a batch of state accesses offline</title><link>https://tn.ms/research/offline-memory-checking/</link><guid isPermaLink="true">https://tn.ms/research/offline-memory-checking/</guid><description>Offline memory checking for a key–value state: a batch of transactions is verified from a log sorted by key and version, using only checks on single rows, checks on consecutive rows and one permutation argument. A Lean proof shows that the checks force every access to have seen the value of its key just before it and the claimed post-state to equal sequential execution, and that distinct versions cannot be dropped.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Memory checking</category><category>Permutation arguments</category><category>Verifiable computation</category></item><item><title>Group key schedules as reachability</title><link>https://tn.ms/research/group-key-schedules/</link><guid isPermaLink="true">https://tn.ms/research/group-key-schedules/</guid><description>A symbolic theory of group key schedules over an ordered, authenticated log: secrecy as non-reachability in a derivation graph, healing as confinement of wrap keys, the collapse of top-down derived trees, a lower bound on the number of published rules (wraps or KDF edges) under a separation hypothesis, and the epochs a long-term recovery key exposes, with two schedules worked out and every statement checked in Lean.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Key schedules</category><category>Symbolic security</category><category>Lower bounds</category></item><item><title>Metered verified primitives: data stores whose cost is known before they run</title><link>https://tn.ms/research/metered-verified-primitives/</link><guid isPermaLink="true">https://tn.ms/research/metered-verified-primitives/</guid><description>A data primitive as a specification, an executor that reports a usage vector, and a bound computed before execution, tied by three laws. A gate that refuses over-budget requests is sound, atomic and never charges more than the limit, and the laws are closed under composition, so a store assembled from primitives inherits both its meaning and its price.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Resource bounds</category><category>Refinement</category><category>Ordered indexes</category></item><item><title>Auditing a signed log that not everyone may read</title><link>https://tn.ms/research/private-log-audit/</link><guid isPermaLink="true">https://tn.ms/research/private-log-audit/</guid><description>A signed, hash-chained log whose rules also limit who may read each event. Validity is a fold over a permission state, a fraud proof about one step convicts the signer and cannot frame an honest one, the parties who can check a step are characterised exactly for those who know the state, and two machine-checked counterexamples show that readers and auditors differ. A zero-knowledge fraud proof discloses only its claim. Everything is proved in an idealised symbolic model and checked in Lean.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Hash chains</category><category>Fraud proofs</category><category>Access control</category><category>Zero knowledge</category></item><item><title>Rotating mailboxes over opportunistic forwarding</title><link>https://tn.ms/research/rotating-mailboxes/</link><guid isPermaLink="true">https://tn.ms/research/rotating-mailboxes/</guid><description>Signed packets, mailbox ids rotated per epoch by a pseudorandom function of a shared secret, a skew-tolerant polling window and TTL-bounded flooding over contacts, combined so that authentic messages reach their recipient with no server, account or fixed address. Each property is proved for the combination in an idealised symbolic model, with the core statements checked in Lean.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Opportunistic forwarding</category><category>Temporal graphs</category><category>Symbolic cryptography</category></item><item><title>Taking powers away from a message host</title><link>https://tn.ms/research/host-powers/</link><guid isPermaLink="true">https://tn.ms/research/host-powers/</guid><description>A host that stores and relays a group&apos;s messages can read, alter, reorder, drop, break the group&apos;s rules, exclude members and disappear. Combining end-to-end encryption, signed hash-chained heads exchanged among readers, rule checking and migration of a self-certifying log removes, exposes or escapes each of these powers under stated conditions. Proved in a symbolic model without time, together with a limit no combination of this kind can overcome: withholding an entry from every reader leaves no evidence. Serving a stale view is a second, stated limit outside the model.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Fork consistency</category><category>Hash chains</category><category>Accountability</category></item><item><title>Fast test answers that stay sound</title><link>https://tn.ms/research/sound-fast-testing/</link><guid isPermaLink="true">https://tn.ms/research/sound-fast-testing/</guid><description>After a change, re-run only the test cases that could have seen it, cheapest tier first. Recording each case&apos;s read set while it runs, selecting cases by intersection with the changed locations, and ordering tiers by cost combine into one procedure. A Lean proof shows that it reports exactly the verdicts of a full re-run, after any number of changes, while executing no more cases than the change reaches.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Regression test selection</category><category>Dynamic read sets</category><category>Refinement</category></item><item><title>Many writers, one main line</title><link>https://tn.ms/research/agent-merge-queue/</link><guid isPermaLink="true">https://tn.ms/research/agent-merge-queue/</guid><description>Many autonomous agents change one shared state at once. Declared footprints, time-bounded leases, a batching merge queue that bisects on failure, and task groupings with prerequisites combine into one admission protocol. A Lean proof shows that the main line only ever holds checked states, that no admitted change is lost, that disjoint work never conflicts, and how many check runs a drain costs.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Merge queues</category><category>Group testing</category><category>Leases</category><category>Optimistic validation</category></item><item><title>Metered non-interference over sealed fields</title><link>https://tn.ms/research/metered-noninterference/</link><guid isPermaLink="true">https://tn.ms/research/metered-noninterference/</guid><description>A typed rule and query language over sealed fields, with metered evaluation and typed declassification, and a logical-relations proof that results, decisions, receipts and charged usage depend on sealed content only through ciphertext handles, lengths and declassified fields.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Information flow</category><category>Logical relations</category><category>Resource metering</category></item><item><title>Checking a log from its signed head</title><link>https://tn.ms/research/transparency-proofs/</link><guid isPermaLink="true">https://tn.ms/research/transparency-proofs/</guid><description>An outsider who holds only a public key and a signed tree head can check that an entry is in a log, that a row is in a state, and that a later head extends an earlier one. Merkle tree hashing with leaf and node domain separation, inclusion and consistency proofs, a row proof and signed heads are combined, and every soundness theorem is proved in the form &quot;the claim holds, or this computable extractor returns a collision of the hash&quot;. Unforgeability is a hypothesis, never an axiom. Everything is checked in Lean.</description><pubDate>Wed, 11 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>Merkle trees</category><category>Transparency logs</category><category>Inclusion proofs</category><category>Consistency proofs</category><category>Signed tree heads</category></item><item><title>Delta packing: nine numbers in sixteen bits</title><link>https://tn.ms/research/delta-packing/</link><guid isPermaLink="true">https://tn.ms/research/delta-packing/</guid><description>Runs of identical deltas and repeated types collapse into a single count-and-value pair. Here are the exact bits a bit-level columnar encoding writes, plus the case where this stops working.</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>research</category><category>encoding</category><category>columnar</category><category>compression</category></item><item><title>Twelve columns: turning a JSON tree into a bitstream</title><link>https://tn.ms/research/twelve-columns/</link><guid isPermaLink="true">https://tn.ms/research/twelve-columns/</guid><description>A bit-level columnar encoding scans a document once, gives every container and key a slot, and splits the result into twelve column streams. Earlier columns tell the decoder how to read later ones.</description><pubDate>Tue, 20 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>encoding</category><category>columnar</category><category>databases</category></item><item><title>Bits, not bytes: where byte-aligned integer codes leave space on the table</title><link>https://tn.ms/research/bits-not-bytes/</link><guid isPermaLink="true">https://tn.ms/research/bits-not-bytes/</guid><description>Byte-aligned integer encodings waste most of a byte on small numbers. A 2-bit width selector recovers most of it, and you can try it yourself below.</description><pubDate>Fri, 02 Oct 2026 00:00:00 GMT</pubDate><category>research</category><category>encoding</category><category>integer codes</category></item></channel></rss>