miraIRIDE: a Covered Progressive Index at Sixteen Bytes per Key

Abstract

"The best pointer is the one that does not exist." miraIRIDE is a covered exact-lookup index for immutable data whose entire structure — projections, references, values, chaining — costs 16.3 bytes per key, measured, while answering batched point queries at roughly twice the speed of ankerl::unordered_dense v4.4.0 (the fastest general-purpose open-source hash map) on the same dual-Xeon hardware, ingesting at 10 million keys per second, and remaining fully queryable at every stage of its own construction. Coverage is not promised but counted: 0 violations across 1,200,000 audited candidates in two independent full runs. Four bit-level ideas carry the result: a positional implicit directory (zero bytes — the hash cell is the block address); 40-bit packed references; 13-bit projections probed branchlessly four-per-ALU-operation via SWAR, with the 13th bit living in the unused "attic" of the same machine word; and a printed coverage counter that turns a probabilistic guarantee into a number. Equally central is the method: every figure comes from an adversarial self-benchmarking tribunal — the state-of-the-art competitor compiled in-process, environment self-declared by the binary, page faults included for all parties, min..max spreads, known limits printed verbatim, and a public prediction ledger that retains the author's falsified predictions as first-class results. A second measured environment puts five index families (SQLite covered B-tree, std::map, std::unordered_map, ankerl, miraIRIDE) under one judge over one workload with five matching checksums: against the most deployed covered B-tree on the planet, the ratios are 43x ingestion, 18x lookup, 3.2x memory. miraIRIDE is the index layer of the mira* research line, companion to the miraNET Digital Runtime: where miraNET lets data speak C, miraIRIDE lets projections speak bits. This record contains the paper — 10 pages, 5 figures, 6 tables — in PDF and Markdown editions. License CC BY-NC 4.0.