Protocol

Where the artifacts live

• Source: https://github.com/PeterPonyu/edit-federation-map
• Numerical archive (concept record): https://doi.org/10.5281/zenodo.21405273

Archive layout:

code/        experiment + analysis source (MIT)
results/     merge arrays and operating tables (CC BY 4.0)
figures/     R rebuild of the public plots

Frozen protocol documents ship in the same archive under CC BY 4.0. The concept DOI is unchanged (v1.0.0, 2026-07-17).

Environment used for the shipped numbers

Python 3.13.7, numpy 2.2.6, scipy 1.16.2, torch 2.12.1 (CUDA 13.0), transformers 5.12.1. Plots additionally need R with jsonlite, ggplot2, and tikzDevice.

Two-phase measurement

Each ROME cell is produced by the two-phase merge measurement in the public source:

• Phase 1 installs edits and writes per-edit keys, values, residuals, and a closed-form geometry table.
• Phase 2 merges groups at g ∈ {2, 3, 5, 10, 20, 50, 100}, writes the merge arrays and an operating-curve table. Qualification at a (g, seed) cell requires non-negligible damage, no argmax saturation, a coherence check, and partial ρ ≥ 0.15.

An editor-general analogue covers MEMIT and AlphaEdit; ROME is the fp64 equivalence anchor. Drivers expect a local model directory and local fact files. All analysis scripts below are CPU-only and read the shipped result arrays with no GPU.

Rebuild tables from the shipped arrays

From the archive root, make the result arrays visible to the analysis tree, then rebuild the operating map, signed reanalysis, cross-term alignment, matched-dose span, gain screen, gain holdout, and admission-benefit tables. Re-score one shipped merge bundle by pointing the two-phase measurement at that cell’s array directory. Public plots rebuild from the R figure pipeline.

ln -s ../results code/results
cd code
python3 experiments/rg_map_evidence_consolidate.py
python3 experiments/rg_signed_reanalysis.py
python3 experiments/rg_crossterm_alignment.py
python3 experiments/rg_matched_dose_spread.py
python3 experiments/rg_gain_law.py
python3 experiments/rg_gain_holdout.py
python3 experiments/rg_admission_benefit.py
cd ../figures
Rscript make_figures.R
make

Canonical numerical tables

• 22-cell operating map (65,868 merges)
• Gain screen and ordering Spearman (point estimate −0.82; frozen object ≤ −0.7)
• Constructive-regime sign decomposition (14B constructive fraction 81–88%)
• Geometry-ordered versus magnitude-only admission. The high-gain benefit +0.647 is measured as a member-target quantity under fixed composition
• Matched-dose response span

Phi-3.5 rows use a corrected tokenizer mapping (2026-07-30). Superseded rows are kept for audit alongside the freeze.

Scope of the claims

Stated once here, so the tables above can be read at face value. Gain works as a rank-level screen: it orders the two regimes, and geometry carries the within-regime size of the cross-term. The geometry-valid group size is a per-cell window — g ≤ 5 at the reference cell, wider at six further qualifying cells — and the per-cell windows ship in the archive. The +0.647 admission benefit is retrospective under fixed composition; the prospective group-formation run supports the primary geometry comparisons in all three seeds, and its utility-matched gate reads Mixed.