Repository navigation
HashFrontier v0.5.0 — trace pairing infrastructure and high-load crash audit
HashFrontier v0.5.0 — Paired-workload instrumentation and honest crash reporting
Research update, 2026-10-22. Not a claim of a new superior hashing algorithm.
The primary result is an audit of v0.4.0 author-source multi-load evidence: 85% occupancy completed all four native variants twice (320 cycles each); at 95% occupancy RHM completed 320 cycles, TRHM segfaulted after 191 cycles (exit 139), and the remaining six matrix entries were not executed because the GitHub shell exited. Data: results/v0.5.0-v04-incident.csv, results/v0.5.0-v04-provenance.json, and figures/v0.5.0-v04-incident.svg; original run 39795819645 and uploaded raw Artifact 12432837986.
This release adds a new research capability designed to fix the earlier non-paired bias: scripts/patch_zombie_paired.py instruments only the author benchmark input generation in pinned saltsystemslab/ZombieHT commit 5064a381fd3b7a540594fadd6f7595fdcc748a6c. It replaces OpenSSL-sampled test-key material with seedable C++ PRNG bytes, seeds the benchmark rand(), and fingerprints every complete generated operation tape. The original hash-table algorithm files are unchanged. scripts/audit_zombie_paired.py verifies full phase logs, process exits, and per-cycle trace identity before admitting complete cohorts. An independently running GitHub Actions job executes 85%/95% occupancy with seeds 41/53/67 and 320 cycles, saving all raw logs and checksums even when a variant crashes. Regression tests include deliberate trace divergence.
Important: A successful v0.5.0 release does not mean the separately-triggered native paired benchmark is successful. Consult the "ZombieHT paired-seed author-source benchmark" GitHub Action and raw Artifact. It is a scaled q=13 instrumented comparison, not the full published paper's q=27 setting, nor original-algorithm independent correctness proof. Failed/incomplete runs cannot count as comparable endpoints. Our bounded-repair candidate remains falsified by v0.3.0 data.
Next: validate identical input traces and operator equivalence under oracle tests, run more independent paired seeds, diagnose the TRHM crash with sanitizers, and only then design invariant-preserving bounded tombstone relocation.