FINAL NUMERICAL RESULTS: VERIFIED TEXT AND CAPTIONS
25 October 2026

Scope and provenance
These summaries were recomputed from the preserved, raw-derived 23 October
18-run analysis. They are not a fresh replay of every original raw stream.
The source tables, hashes and exact arithmetic are recorded in
final_results_recompute.py and final_results_verification.json. Independent
inventory checks verified all276 packaged member hashes and38 surviving raw
files against their recorded input hashes;24 surviving raw static run/workload
rows exactly reproduce the corresponding pooled medians. Those validation
results are in full_recording_inventory_validation.json (separate reviewer).

The primary estimand is a median over all trace samples for one workload within
one run, pooling its three visits. The centre of a policy/profile/workload cell
is the arithmetic mean of its TWO independent run medians. It is not a pooled
median over both runs, a median of phase medians, or an average of frame samples
treated as independent observations. Replicate numbers are chronological labels,
not randomized or matched cross-arm block identifiers. The median-of-three-phase-
medians alternative is supplied only as a sensitivity, and can differ from the
primary run median by up to2.3083FPS,9.01675ms,0.637785W or1.5C. Do not silently mix
these estimands or substitute values between tables.

Manuscript-ready Results text

The comparison comprises six ERL missions and twelve static missions, each
nominally six hours with twelve1800-s phases. ERL has two chronological repetitions
under each of the inactive, moderate and aggressive profiles; the static policies
have two repetitions under inactive and aggressive only. The recorded observation
duration totals107.9996h. The six ERL streams contain161,723 controller-action
records and66 workload transitions. The commanded-cap/readback check matches in
all161,723 records; the separate reported-current-state check flags3,701 records
(2.2885%). Thus the actuation audit has separate pass and failure findings rather
than a blanket pass. There are198 workload transitions and8,077,130 retained
per-frame metric rows across all18 runs, of which2,585,429 rows belong to ERL.
These retained records do not establish lossless processing or physical frequency.

ERL is not the highest-throughput policy in this comparison. Under the inactive
profile, its two-run centres are44.90,44.97,37.18 and33.95FPS for Histogram, Sobel,
Gaussian and Median, respectively. Relative to the largest static throughput
centre in each workload cell, these are3.33%,1.68%,19.01% and21.99% lower. Static
Histogram and Sobel centres differ by only0.4061 and0.5552FPS, respectively, below
the manuscript's2-FPS discrimination floor. The selected largest centre is a
descriptive reference; these small separations do not identify a meaningful
universal winner. Inactive Gaussian shows substantial ERL repetition spread:
35.01–39.35FPS and6.19–23.49ms processing latency, compared with45.79–46.02FPS and
approximately5.58ms for GPU static. This is an observed implementation-path
contrast, conditional on the unresolved output-equivalence and environmental
differences, and does not by itself identify a causal surrogate failure.

Under the aggressive profile, ERL's throughput centres remain below the highest
static centre in all four workloads. Its sampled-power centre is nevertheless
0.560–1.206W lower than each corresponding static policy centre. This lower-power
ordering also holds for every ERL-versus-static combination of the two observed
run medians in each workload. The data therefore show a recorded power/throughput
trade-off, not an isolated learning-caused energy-efficiency effect. Temperature
and ambient conditions, backend gating, CPU caps, worker counts and uncontrolled
input content remain parts of the observed policy bundles.

Aggressive Median provides the clearest favourable observed trade-off. ERL's
centre is20.73FPS,40.03ms processing latency and5.51W, compared with17.31FPS,
54.20ms and6.44W for GPU static and16.87FPS,54.46ms and6.57W for Balanced. The
throughput differences are+19.79% and+22.87%; both ERL run medians exceed both
GPU-static and Balanced run medians. CPU static, however, retains a higher
throughput centre of26.27FPS, with a wide21.16–31.38FPS repetition range, and a
processing-latency centre of33.80ms. All CPU-static repetitions are retained.
This favourable case therefore does not establish superiority over every static
configuration, nor equivalence of useful outputs across execution paths.

Whole-mission channel-0 energy is integrated from i0*v0 against the host
system_time_us clock, with linear interpolation at mission endpoints and across
disclosed gaps. It is distinct from sampled trace-power medians. ERL aggressive
missions integrate115.673 and120.274kJ, while aggressive static missions span
131.412–139.266kJ. Inactive ERL integrates133.831 and136.211kJ; moderate ERL
integrates133.920 and131.087kJ. No static moderate comparison exists. Across all
missions the largest host-clock gap is8.1978s; intervals no longer than3s cover
at least99.8291% of the manifest interval. Omitting longer intervals changes
the estimated mean power by at most0.00145W. The large10,721-s meter-device clock
discontinuity is not a host acquisition outage. Equal mission duration permits
an energy comparison, but differing delivered work and unverified output
equivalence prohibit interpreting these totals as energy per useful output.

The median measured GA computation time is2.04–3.24ms per run, or0.255–0.405%
of the nominal800-ms period. The corresponding95th percentile spans22.95–112.48ms
(2.868–14.060% of the period). Summed GA time occupies0.752–3.479% of elapsed
mission time; this is a different aggregation and is not the cost of the entire
controller, instrumentation or pipeline. Seven recorded GA computations exceed
800ms across all161,723 action records. ERL median algorithm-switch downtime
ranges from10.90 to24.72ms across runs; the maximum observed switch downtime is
662.179ms. These observations quantify the executed campaign, not a hard
real-time bound or exhaustive lifecycle guarantee.

Supplementary normalized-score text

With the original four-metric weights and floors, the strict normalized reference
gap is defined in only aggressive Sobel, giving1.591 and1.038 for the two ERL
runs. It is undefined in the other seven profile/workload cells. Even this admitted
cell changes with preference weights: equal weights give0.492/0.087, power-emphasis
-1.046/-1.666, and throughput-emphasis2.753/2.196. These are weighted positions
relative to the three measured static centres, not optimality certificates.

A separate post-hoc throughput/processing-latency score on the inactive profile,
with weights5/6 and1/6, gives Gaussian0.51385/1.01513 and Median0.40351/0.79951.
These values reproduce the earlier0.51/1.02 and0.40/0.80 summaries. They answer
a narrower question than the protocol's four-metric score and should remain
supplementary. Output-equivalence, environmental and source/binary limitations
remain; renormalizing weights does not repair them. Strict-gap values above one
or below zero only describe extrapolation beyond the measured reference score
range under chosen weights.

Main figure captions

final_throughput_eight_cells.pdf
Reported throughput across the two profiles with static references. Each symbol
is one run-level workload median, pooling three visits within that run; circles
and triangles denote chronological repetitions1 and2. Lines connect the two
observed values and are not confidence intervals or matched-pair effects. The
rows show inactive and aggressive profiles; each row uses a common vertical
scale. CPU, GPU and BAL denote the three static policy configurations. Under
aggressive conditions, GPU-labelled policies can execute CPU work after GPU
permission is withdrawn. All repetitions are retained. Output equivalence is
unverified, so the figure compares recorded implementation paths.

final_aggressive_median_tradeoff.pdf
Observed throughput, processing latency and sampled-power trade-offs for Median
under the aggressive profile. Points and connecting ranges have the same
run-level meaning as the throughput figure. Both ERL runs exceed the GPU-static
and Balanced throughput observations while using lower sampled power, but the
CPU-static centre remains faster and has substantial repetition spread. The
comparison does not establish causal learning gains or energy per useful output.

Optional supplementary figure captions

final_performance_by_run.pdf
Reported throughput and processing-latency run medians under inactive and
aggressive profiles. Individual runs are shown; connecting lines denote observed
repetition spread. Kernel/processing latency and sampled pipeline throughput
remain different estimands. No frame-level inferential intervals are shown.

final_power_temperature_by_run.pdf
Sampled-power and CPU-temperature run medians. Points show two runs per policy
cell; connecting lines are observed ranges. Power is the trace estimator, not
whole-mission integrated energy. Ambient temperature and source freshness are
not controlled by this representation; lower temperatures do not demonstrate
an enforced thermal guarantee.

final_mission_energy_by_run.pdf
Channel-0 energy integrated on the host clock over each nominal six-hour mission.
Chronological run numbers are not cross-arm pairing identifiers. Moderate static
cells were not run. Energy totals include all observed work and pipeline activity,
with disclosed interpolation across meter gaps; they are not energy per useful
output or isolated processor energy. See the meter-interval sensitivity.

Table captions

final_results_throughput_table.tex
ERL throughput relative to the largest observed static two-run centre in each
profile/workload cell. Entries are the arithmetic mean of two run medians with
their observed minimum/maximum in brackets. Difference is100*(ERL/reference-1).
The reference selection is descriptive, not a global or output-equivalent oracle.
Inactive Histogram/Sobel static spreads fall below the2-FPS discrimination floor.

final_results_aggressive_median_table.tex
Aggressive-profile Median trade-offs. Entries are two-run centres and observed
ranges of run-level workload medians. FPS, processing latency and trace power
are distinct recorded channels. These ranges are not confidence intervals.

final_results_energy_table.tex
Host-integrated channel-0 mission energy for all18 runs. Mean power is the mean
of the two per-run energy/duration values. Replicate labels denote chronological
order only. Moderate has no static arm. The comparison does not normalize for
equivalent completed outputs.

final_results_erl_operation_table.tex
ERL recording integrity and measured computation/switch timing. C1b flags compare
reported current policy state with the command, not measured physical frequency.
All action records pass commanded-cap readback equality. GA duty is summed
measured GA computation divided by elapsed mission time, not total-system cost.

Data/use boundaries
- No across-run p-values or confidence intervals are estimated from n=2 per cell.
- Four cross-run ERL/static combinations are sensitivity ranges, not fourreplicates.
- CPU_STATIC aggressive repetition1 remains in every primary calculation.
- The full comparison is schedule-matched; identical inputs, ambient conditions,
  runtime source revision and output equivalence are not established.
- All centre calculations match the preserved independent comparative analysis.
- Primary and sensitivity CSVs carry full precision; printed text rounds only.
