v1_local_baseline

v1_local_baseline — Frozen Local Baseline for Phase 3 LAN Comparison

This directory contains the frozen local-only baseline for Relativist, produced on the tagged binary v0.10.0-bench. It is the reference the upcoming Phase 3 LAN campaign will subtract from to isolate network cost.

Do not modify any file in this directory. To reproduce the campaign on a different machine, run scripts/reproduce_local_baseline.sh — wall-clock values will diverge (different hardware), but row counts, correctness flags, and round structure must match.

Contents

File	Description
manifest.md	Full provenance: commit SHA, tag, hardware, timestamps, sha256 checksums, campaign knobs
phase1_lenient_detail.csv	Phase 1 (in-process) per-repetition rows, lenient BSP (default)
phase1_lenient_rounds.csv	Phase 1 per-round rows, lenient BSP
phase1_lenient_summary.csv	Phase 1 aggregate stats (mean, std, median, CV), lenient BSP
phase1_strict_detail.csv	Phase 1 per-repetition rows, strict BSP (cascade_cross, dual_tree subset)
phase1_strict_rounds.csv	Phase 1 per-round rows, strict BSP — multi-round data for Phase 3
phase1_strict_summary.csv	Phase 1 aggregate stats, strict BSP
phase2_detail.csv	Phase 2 (Docker / TcpLocalhost) per-repetition rows
phase2_rounds.csv	Phase 2 per-round rows
phase2_summary.csv	Phase 2 aggregate stats
cv_triage.md	CV > 0.15 datapoints flagged with keep/rerun/exclude disposition
raw/phase1/	Per-benchmark stdout/stderr logs and raw CSV fragments
raw/phase2/	Per-run metrics.json snapshots from Docker runs

Scope note

• Correctness method: Full G1 isomorphism for all benchmarks except condup_expansion at sizes 10k and 50k, where the --skip-g1 weak check (agent count equality) is used by default (abordagem A, see USAGE_GUIDE.md Section 11.5). The optional full-G1 overnight verification (abordagem B) is documented there for voluntary replication.
• Strict BSP data: Only cascade_cross (all default sizes) and dual_tree (sizes 6, 10, 14) have strict-mode rows; all other Phase 1 benchmarks run in the default lenient mode.
• Phase 2 L6: Configs that were previously blocked by the 256 MiB frame cap (dual_tree=22 at w=1, ep_annihilation_con=5M at w∈{1,2,4}) now complete within the 1800s timeout thanks to the CompactSubnet fix shipped with v0.10.0-bench.