ONLINE RESOURCE 2 — ERL EVALUATION EVIDENCE AND REPRODUCTION
26 October 2026
Article: Stateful Runtime Adaptation under Workload and Thermal Change:
An Audited Evaluation on a Heterogeneous SoC
Authors: Antonio Souto Rodriguez; José Reginaldo Hughes Carvalho
Journal: The Journal of Supercomputing
Affiliation: Institute of Computing, Federal University of Amazonas, Manaus, Brazil
Corresponding author: antonio.rodriguez@icomp.ufam.edu.br

SCOPE
This package contains preserved raw-derived analysis for 18 missions, revised
summaries/forecast diagnostics, analysis programs and a reference source snapshot.
It DOES NOT contain every original raw recording. Reproducing the full raw-stream
audit or executing hardware experiments requires the separate raw corpus/device.
Reference source identity is not verified against the recorded executable hashes.

PRIMARY RESULTS
analysis/final_results_* is the current result set. Run-level workload medians
pool three visits within a run; a cell centre averages the TWO run medians.
analysis/final_forecasts_* supersedes the historical forecast/persistence/recurrence
results under recovered/. It excludes 68 boundary/time inconsistencies before
constructing baseline histories: 161649 retained pairs and 161625 common targets.
The forecast comparison is exploratory, conditional on recorded timestamps and
assigned windows; it is not causal action credit or proof of memory benefit.
No frame-independent p-values or equivalence claims are provided.

RAW CORPUS
The complete 18-mission raw corpus (mission archives, analysis driver, full-corpus replay
log, checksums) is version 2 of Zenodo record 10.5281/zenodo.23350634: https://doi.org/10.5281/zenodo.23676764

CONTENTS
analysis/: current tables, figures, revised scripts and validation records.
recovered/full_evaluation/ERL_Full_Evaluation_20261023/: preserved data and audit
  programs from the preceding full recording evaluation. Old narrative PDFs and
  manuscript text are omitted. Historical unfiltered forecast summaries remain
  for provenance only; use the revised analysis/final_forecasts_* for this paper.
reference_source/: 90 redistributed members of the reviewed 91-file snapshot,
  byte-identical to the per-file reviewed manifest. The unused Xilinx-notice
  Stage_01/Others/LKof_defines.h is omitted. Historical source-review inventories
  still describe the complete original snapshot. Third-party notices and GPL
  text accompany the distribution; LICENSE_CODE.txt scopes the authors' MIT terms.
reference_source_review/: focused source probes and recorded review results.
  These are reference-snapshot checks, not historical binary execution results.
MANIFEST_SHA256.txt: hashes for the actual files in this package.
PACKAGE_PROVENANCE.json: source-archive hashes and selected-file inventory.
verify_reproduction.py: hash check plus isolated reproduction of primary summaries.

REPRODUCE THE REVISED CALCULATIONS
Requires Python 3 with pandas, numpy and matplotlib. No network or Jetson is used.
From this directory:
  python3 verify_reproduction.py --workdir /absolute/path/to/new/reproduction_dir

The verifier checks package hashes, copies the preserved inputs to the new work
directory, runs the revised results script, runs the revised forecast script, and
compares regenerated numeric CSV outputs with the supplied outputs. It does not
overwrite packaged evidence. The work directory must not already exist.

Direct scripts (run on a copy if preserving the packaged manifest):
  python3 analysis/final_results_recompute.py
  python3 analysis/final_forecasts_recompute.py --input recovered/full_evaluation/ERL_Full_Evaluation_20261023/erl_events --output /absolute/output/directory

PROVENANCE AND LIMITS
The original full-analysis archive was verified against all 276 recorded member
hashes before assembly. The current raw-subset check matched 38 canonical files
to prior input hashes and reproduced 24 static run/workload summary rows.
This validates preserved analysis and the surviving subset; it does not constitute
a new full replay of all raw streams. Source probes may contain legacy absolute
paths, GPU/toolchain dependencies, or unavailable raw-input requirements. Only
the two revised analysis programs above are covered by the portable reproduction
check. Hardware output equivalence, randomized conditions, source/build binding,
and mechanism ablations remain outside the supplied evidence.

RUN IDENTIFIERS
Use authoritative_18run_inventory.csv. Chronological replicate labels are not
randomized or matched cross-arm block identifiers. There is no moderate static arm.
Preserve the original run names and recorded hashes when linking raw archives.

PUBLIC SOURCE CORRESPONDENCE
The cited original archive hash identifies the 91-file reviewed input, not a
byte-identical public archive and not a verified compiled-binary source input.
reference_source/REVIEWED_SNAPSHOT_FILE_MANIFEST.sha256 preserves all 91 hashes;
90 source members are distributed unchanged and the omitted path is declared
in PACKAGE_PROVENANCE.json. Source-review findings about the original snapshot
remain historical records, including file counts and source-line evidence.
