Longer work

Papers and articles

Slower-form research and technical writing. This section stays sparse until the work deserves the space.

[PAPERS]

Indexed work

independently publishedpaper2026-09-08

Nidus: Explicit Application Composition for Rust Backend Services

A backend application needs a composition model as well as a request handler. Its components must declare dependencies, initialize shared resources, expose routes, and coordinate operational behavior. Nidus organizes these responsibilities through explicit modules, typed providers, and controller metadata, while retaining Axum routing, Tower middleware, and Tokio execution. This paper explains the design through three separate stages: generated Rust declarations, application construction, and request execution. It describes the guarantees provided at each stage, follows a small feature from module declaration to handler invocation, and identifies the costs and limits of the abstraction. The implementation uses a runtime container keyed by Rust types and validates module structure during bootstrap. Module exports describe composition contracts; they do not enforce runtime access isolation. The account is anchored to Nidus 1.0.17 and supported by source inspection and focused executable checks.

independently publishedpaper2026-09-08

The Exact Price of Mergeability for Capacity-Parametric Counter Replay

An exact characterization of the information needed to summarize saturating counter logs when the replay capacity is chosen later from the integers 0 through N. Final-value queries from zero have 2^N observational equivalence classes and require exactly N logical bits. Supporting the ordered merge of independently summarized chunks raises the necessary state count to N·2^N + 1, attained by a canonical, associative summary. For N ≥ 1, the minimum fixed-width information content is N + floor(log₂ N) + 1 bits: at N = 64, merging costs seven additional bits. The lower bound requires only correctness after one binary merge. Semantic, reachability, counting, and lower-bound results are checked in Lean 4. These are logical information bounds, with no claim of an optimized packed implementation or throughput improvement; historical novelty remains unestablished in light of related work on saturation and hysteresis.

independently publishedpaper2026-07-24

Rigidity of Pattern-Avoiding Breadth-First Reading Words of Increasing Trees

Two rigidity theorems for pattern-avoiding breadth-first reading words of increasing trees resolve three conjectures recorded in the OEIS. For 312-avoiding permutations, realizability on an increasing unary-binary tree is equivalent to the binary-heap inequality, with the identity map giving the conjectured bijection. For 231-avoiding words of odd length, unary-binary realizability implies full binary realizability. Both collapses fail for the pattern 321, with explicit minimal counterexamples. The proofs rely on a promotion cascade lemma that holds without a pattern hypothesis. The paper also extends six OEIS sequences by 28 new terms and accompanies the results with self-contained proofs, exhaustive computation, and complete Lean 4 formalizations over genuine inductive trees.

independently publishedpaper2026-07-06

Locus: Owner-Drained Chunk Mailboxes for KV-Block Recycling in CPU LLM Inference

A technical white paper on Locus, a domain memory pool for CPU LLM serving whose remote-free path is a per-worker lock-free chunk mailbox: a finished request's KV blocks are returned as one atomic push, and the pool owner drains every mailbox off the allocation hot path, so freeing never contends on a shared queue and the design carries zero tuning parameters. Developed under a falsification-first methodology, it is evaluated on LOCUS-EVAL v1, a frozen four-workload suite of deterministic serving-shaped KV traces run against jemalloc, mimalloc, and system malloc on Apple Silicon, where the mailbox ranks first on every workload. A self-critical touch-parity audit corrects the headline: at a one-byte-per-block parity the advantage is 1.6x to 2.7x over mimalloc on the three trace workloads and compresses to about 1.15x over system malloc once full KV writes dominate, with a transient burst-cancellation footprint of up to 1.5x the theoretical peak. The scope is deliberately narrow (a single macOS host, synthetic traces, an allocator-level measurement), with NUMA locality and end-to-end serving integration stated as future work.

independently publishedpaper2026-06-07

The Missing HTTP Verb: STRUT

A position paper proposing STRUT, an idempotent but unsafe HTTP method for server-driven, minimal-input resource creation. The paper defines candidate semantics, compares the method with POST plus Idempotency-Key, conditional PUT, WebDAV, and Prefer, and discusses security, caching, intermediaries, browser behavior, and deployment.