Reproducibility Package for What Information Can Be Discarded

Abstract

Release purpose This archive is the release-grade executable evidence and verification system for “What Information Can Be Discarded?” It freezes the proof-support artifacts, experiment suites, source code, public or generated inputs, protocols, environment locks, sealed results, statistical analyses, and paper-to-result indexes required to reconstruct the publication’s quantitative claims. Scope Release 1.1.0, freeze identity minimal-1.1.0-20260821, contains only evidence used in the paper. The ZIP contains 1,954 members. PACKAGE_MANIFEST.json tracks the other 1,953 files. The dependency-free root audit verifies 1,191 JSON files, 19 JSONL files with 90,120 rows, 178 Python files, 12 nested ZIP archives, and 19 GZIP archives. Thirty-seven experiment-suite directories and seven registered first-party audit entry points are included. Historical drafts, rejected-result caches, credentials, private data, machine-specific absolute paths, virtual environments, compiled caches, and experiments not reported in the paper are excluded. Nine official FDA FAERS quarterly archives are acquired from their exact official URLs and verified byte counts and SHA-256 identities rather than duplicated in the package. Verification architecture The root README is the primary navigation layer. CURRENT_EVIDENCE_INDEX.json binds the release identity, paper-to-package paths, current suite navigation, and the exact commands for the seven registered first-party audits. PACKAGE_MANIFEST.json records the byte size and SHA-256 digest of every other tracked package file. AUDIT.py performs dependency-free package-integrity verification after extraction. The proofs directory contains machine-readable proof checks and theorem-support artifacts. The experiments directory contains the frozen suite protocols, source locks, public or generated inputs, formal results, statistics, and suite-specific verification programs. Suite-specific environment locks take precedence over the consolidated dependency pins. E4 and E5 seal independent dynamic-programming truth under explicit IEEE-754 binary64 semantics. Each truth value is stored as a JSON number and canonical float.hex text, and the verifier requires bit-for-bit equality. This removes platform-dependent extended-precision recomputation without changing the publication correctness gate, comparison class, retained result, or reported number. No retained suite requires local training or execution of a large language model. Public model records and detector resources are used only where explicitly identified by the paper and suite registry. Reproduction contract Run python AUDIT.py at the package root after extraction. Then follow the suite-specific README or frozen protocol for the claim being reproduced. Formal results constitute publication evidence. Development partitions are included only where they belong to a retained two-stage design. Exact outputs, frozen-tolerance outputs, classifications, endpoint values, multiplicity corrections, confidence bounds, and result hashes are checked under the suite contracts. Timing reruns retain the suite’s thread count, randomized arm order, preprocessing boundary, and hardware-accounting rule. Cross-machine timing is newly measured while exact outputs and inferential conclusions remain subject to the frozen verification gates. Evidence navigation Independent evidence paths cover: • the master theorem, local branch factorization, behavioral quotient, and exact and approximate certificates; • quantum channels, quantum instruments, quantum-comb interfaces, and generalized-Weyl operational defects; • probabilistic programs, probabilistic model checking, stochastic games, exact MDPs, and fixed-controller POMDP evaluation; • query-conditioned KV reduction, neural-structure coalescing, lifted probabilistic inference, and causal selection; • telemetry anomaly detection, network-epidemic dynamics, prospective pharmacovigilance, employment transitions, and public-process prediction; • exact optimization and verification quotients, differential-privacy workloads, process conformance, phylogenetic likelihood, and dependency resolution; • scientific model-order reduction, structural mechanics, molecular kinetics, reaction networks, signal and pixel tasks, and cross-domain trajectory prediction; • logic and circuit semantics, side-channel noninterference, graph sparsification, genome graphs, provenance, unlearning, and SynthID-Text audits; • exact ties, low-redundancy boundaries, irreducible cases, and nonidentifiability constructions; • fixed-seed computational proof audits and independent result recomputation. Release integrity The external manifest.json records the publication PDF and standalone reproducibility-package ZIP with their exact names, sizes, and SHA-256 digests. SHA256SUMS.txt separately lists those two artifact digests. The package-internal PACKAGE_MANIFEST.json records every other tracked package file. The associated publication and this archive are linked through reciprocal Zenodo DOI relations. The checksum index is covered by the trusted timestamp certificate identified in the Technical info section.