TEN-SCENARIO FINITE-ENERGY BENCHMARK SUITE
October 6, 2026 — NUMERICAL EVIDENCE ONLY

This is a reproducible benchmark of ten parameter settings of ONE model.
It is not ten different products, ten validated specimens, or a device proven
to split rock or diamond. It extends the numerical evidence without changing
the model to make a desired outcome appear. The physical equations, energy
accounting, controls and limitations are described in efficiency/README.txt.

MATRIX FIXED BEFORE EXECUTION

Every entry uses the same quick search: 48 candidates, three training cases,
four unused sensitivity cases, and timestep refinement of both finalists.
Except for the chosen variable, each row keeps the reference settings:
10 mJ initial kinetic energy, target mode 1, modal damping ratio 0.020,
contact stiffness 30,000 N/m, frequency scale 1.0.

  B01  Reference settings, target mode 1
  B02  Target mode 2
  B03  Target mode 3
  B04  Lower modal damping ratio: 0.005
  B05  Higher modal damping ratio: 0.050
  B06  Lower contact stiffness: 10,000 N/m
  B07  Higher contact stiffness: 80,000 N/m
  B08  Modal frequency scale: 1.15
  B09  Lower initial kinetic energy: 0.001 J
  B10  Higher initial kinetic energy: 0.020 J

One source always has 20 g moving mass. Two sources each have 10 g, with the
same total initial kinetic energy and initial speed as the one-source case.
Thus source count never silently doubles the energy budget. This does not
equalize contact area, aggregate contact stiffness, cost, or reset power.

When the best pair hits simultaneously at the same model position, the suite
retains the matched-aggregate-contact single-source control. That control can
explain an apparent pair advantage through the changed contact law rather
than timing. The controlling comparison is not simply the biggest amplitude.

RUN LOCALLY

From the custom-poc directory, with Python and NumPy available:

  python efficiency/benchmark_suite.py
  python efficiency/benchmark_suite.py --no-reuse
  python -m unittest discover -s efficiency -p test_benchmark_suite.py -v

The first command may reuse explicitly checked existing evidence. The second
reruns every grid, holdout, refinement and applicable aggregate-contact control
from current source. No paid service, GPU, cloud compute or hardware device is
required. Runtime depends on CPU/load. The whole suite has a 570-second stop
condition; each study retains its own 120-second deadline. Intermediate
summaries preserve completed/failed/cancelled status after every scenario.

PROGRAMMATIC INTERFACE

  from efficiency.benchmark_suite import run_benchmark_suite
  summary = run_benchmark_suite(progress=print, cancel=lambda: False)

Optional output_dir selects another results directory. reuse_existing=False
forces full recomputation. progress receives scenario_index, scenario_count,
fraction and stage. No parameter outside the fixed matrix is silently added.

OUTPUTS

  results/NUMERICAL_benchmark_suite.json
    Compact per-scenario results, exact settings, source SHA-256 digests,
    Python/NumPy/platform metadata, execution times and evidence boundaries.

  results/benchmarks/B01.json through B10.json
    Full numerical reports: grid ranking, constraints (including rejected
    candidates), energy histories, per-source work, holdouts and refinements.

  results/NUMERICAL_benchmark_test_report.json
  results/NUMERICAL_benchmark_test_log.txt
    Saved suite-validation result and exact test output when generated.

Top-level summary fields include:
  schema_version, suite_version, evidence_status, physical_validation,
  fracture_validation, hardware_control, source_hashes, environment,
  started_utc, completed_utc, elapsed_s, scenario_count, completed_count,
  all_scenarios_completed, all_numerical_checks_passed,
  fresh_study_count, reused_study_count, scenarios, limitations.

Each completed scenario includes:
  id, label, settings, status, provenance, elapsed_s, report_path,
  report_sha256, candidate_count, feasible_candidate_count,
  best_single_fraction, best_dual_fraction, nominal_dual_over_single_ratio,
  holdout_min_ratio, holdout_max_ratio, all_holdouts_feasible,
  refinement_passed, max_relative_energy_error,
  matched_aggregate_control_applicable,
  matched_aggregate_dual_over_single_ratio,
  matched_aggregate_control_passed, fixed_initial_energy_check_passed,
  numerical_checks_passed, winner_positions_fraction,
  winner_requested_second_delay_s, interpretation.

PROVENANCE AND REUSE

Source hashes cover model.py, study.py, and benchmark_suite.py. Hashes identify
the exact source files used for a fresh run. Full report files are hashed by
the compact summary, so a later change to a report is detectable.

The three preexisting mode studies were created before code hashes were added.
They can be reused after exact settings/grid/evidence checks and an independent
replay of both frozen finalists and their timestep refinement using current
source. Reuse records the old report hash and the replay checks. It does not
retroactively claim that all old grid rows have an original code fingerprint.
Run --no-reuse to replace that limited legacy provenance with a fully fresh run.

Later benchmark reports with matching source hashes and settings can be reused
directly. The summary says which reports were fresh and which were reused.
Fresh execution and reuse are never counted as the same thing. Reproducible
means others can run the included source/settings and compare results; this
package alone does not mean the evidence has been publicly published or
independently peer reviewed.

Fresh output hashes need not be identical across runs because timestamps,
elapsed runtimes and environment metadata change. Reproduction should compare
the fixed settings/source and numerical metrics with stated tolerances, while
report hashes verify that each saved artifact has not subsequently changed.

WHAT A PASS MEANS

The suite checks complete evidence, consistent settings, matched initial
energy, finite nonnegative accounting, global energy residual no larger than
0.5%, successful timestep refinement, and feasible frozen candidates in all
four unused scenarios. When applicable, the aggregate-contact control must
also match. These are numerical/model-consistency checks.

A pass does not validate the ideal clamp, material constants, three-mode
truncation, passive contact law, actual force/strain, hardware, crack paths,
controlled splitting, fragmentation energy, product durability, or commercial
throughput. No fracture model is present. Deterministic holdouts are scenario
brackets, not measured material distributions or statistical confidence.

BEST DOES NOT MEAN GLOBALLY OPTIMAL

Every winner is the best feasible result in the fixed finite grid for the
chosen reversible modal-energy objective. Changing the grid or calibrating
the physical model can change the winner. Ten scenarios are a coverage matrix,
not an exhaustive search of every mechanism, specimen, or possible variable.
All outcomes, including a pair losing to one source, remain in the evidence.

RECORDED RUN

The delivered final suite recomputed all ten studies from current source,
with zero reused studies, in 298.01 seconds on this PC. All numerical checks
passed. The saved source/report hashes were verified by 10 suite tests; the
underlying model passed 25 physics/accounting/input-validation tests. Passing
these 35 tests does not establish physical specimen or fracture validation.

The table gives peak target-mode energy for the best pair divided by the best
single source. Ratios above one indicate more of this modeled energy metric,
not better fracture performance. The holdout interval is the minimum/maximum
across four deterministic parameter cases, not a confidence interval.

  Scenario  Nominal pair/single   Holdout minimum   Holdout maximum
  B01             0.983680            0.982474          0.985405
  B02             1.112262            1.060825          1.182466
  B03             1.457872            1.326814          1.594527
  B04             1.062060            0.890798          0.992844
  B05             0.990461            0.987262          0.998700
  B06             0.975455            0.973493          0.978137
  B07             0.991125            0.990016          0.992856
  B08             0.982553            0.981365          0.985019
  B09             0.985451            0.984162          0.987350
  B10             0.983054            0.981830          0.984886

B04 is the useful timing sensitivity example: the low-damping optimum delays
the second arrival by one nominal mode period, about 36.74 ms. It improves the
nominal target-energy metric but loses to the frozen single-source comparison
in all four unused cases. Static phase tuning did not generalize here.

B02 and B03 choose simultaneous contacts at the same longitudinal position.
Their apparent advantage disappears when a single source has the matching
aggregate contact law. Every applicable aggregate-contact control in this
suite has pair/control ratio exactly 1.0. The recorded results therefore do
not establish a robust timing-based two-source improvement in this matrix.

An earlier pass was superseded after stricter input validation and a
NumPy-version-neutral test helper were added. The full rerun reproduced all
comparison ratios with zero numerical change. The provenance note records
both source fingerprints and the reason for recomputation; historical hashes
were not merely relabeled as if the old run used new code.
