CONTROLLER/METHODS EVIDENCE NOTES — 25 October2026

Fragment: analysis/controller_methods.tex. No source files or main manuscript edited.
This fragment is deliberately source-qualified. It must not be made into an assertion
that the uploaded source is the exact historical compiled implementation.

SOURCE ROOT
paperB_final_20261025/recovered/source/
GA=Module/GeneticAlgorithm_05.h; ES=Module/EvolutionarySelector_02.h;
SM=Module/SurrogateModel.h; SC=Module/Scheduler.h;
AWM=Module/AdaptiveWeightManager.h; TG=Module/ThermalGovernor.h;
CM=Module/ConfigManager.h; H=SampleTestIntegrationCode_v35.cpp.

ARCHITECTURE / CALL ORDER
ES602–603 owns Scheduler/GA; GA864–902 owns population/RNG/generation/AWM/surrogate.
ES711,747–815: snapshot read/fresh frame/workload gates.
ES816–845: collector outcome/score/admission/context BEFORE evolve847.
GA524–527: stability gate INSIDE evolve after collector hooks.
GA547–556: AWM due update; GA1085–1101: pending previous-request observation.
GA1105–1250: prediction/gating/objectives; GA1968–2134: rank/crowding.
GA2311–2387: feasible rankzero scalar choice, feasible and leastviolation fallbacks.
GA617/627: selected copy before scheduler apply; GA694–727: reproduction/increment/export.
Selected-action accessor depends on successful CSV export (GA2639–2692), a further
instrumentation limitation retained in detailed source review, not a validation guarantee.

GENES / MODEL
GA970–1051,1860–1903: mode,targetFPS,powerbudget,GPUpreference,CPUcap,split and
derived model concurrency. Actual Scheduler mappings can differ (SC811–830,999–1090).
GA1257–1322: four minimized objectives. FPS highperformance exception matters.
GA2140–2189: scalar/age/exploration/stall computation.
SM145–178: workloadbank switching;203–239,464–478: EMA updates/shared count.
SM249–346: exact/KNN/prior;499–560: key quantization;349–359: exploration.
GA1140–1154: FPS support/cap gates. Support is shared across targets; there is no
per-target evidence confidence. Do not describe measured context columns repeated
across candidate rows as candidate outcomes.
ES1125–1126: latency learning target is pipeline E2E; ES838 excludes score against
processing-latency target. No sensor-to-output claim.

AWM / THERMAL
AWM178–239,392–489: generation cadence/debounce/EMA/regimepriority/weights.
TG198–213,354–359: temperature predicate; SC543–556: scaling/GPU permission consumed.
SC ignores allow_inference/frame_skip_ratio/power_budget_reduction; shouldShutdown
is one-shot (TG244–253). Recovery/invalid-data paths can bypass normal thermal policy.
These are reason to omit hard safety guarantees, not proof every old run hit failures.
H317–396: profile assignments; current config merge can override some fields.

SWITCH
CM479–565; ES378–479; GA446–501; Modules algorithm replacement:
prepare replacement before oldworker stop/join, drain afterjoin; epochpublish/start;
context commit restores bank, canonicalizes genes, clears GAcredit, resumes.
Pilot prior ticket is separate and is NOT reset by GAcreditclear in this snapshot.
The counter/record proof does not establish whole-memory identity or bounded join.

RECORD-BASED CORROBORATION
full_evaluation/ERL_Full_Evaluation_20261023/pareto_analysis/FINDINGS.txt:
1,293,784 candidate rows =8*161,723 generations. One selected row per generation;
selected feasible rankzero highestloggedfitness, all66 comparedfields agree.
Thus population8 and selection rule have direct log corroboration.
Same directory candidate exports: obj_energy=pred_power_w/10;
obj_latency=pred_latency_ms/50 and obj_temp=max(0,pred_temp_c−45)/20.
Those objective normalizations are source/log concordance, not predictor validity.

tmp/paperB_review/CONTROLLER_SPECIFICATION.md is the preserved earlier banner
transcription for10–12October: nominal800ms,8/4/2population/min/elites,0.70crossover,
0.15mutation, seed42,AWMenabled/update5/smoothing0.30,frequencygenelevels.
It is not a fresh direct read of all six original console logs. Avoid asserting
seed42 or everythreshold/knob for all6 from this transcription alone. Current
fragment uses nominal interval and source mechanics, corroborated population.

STATIC PROVENANCE
full_evaluation/.../provenance/FINDINGS.txt and *_evidence.json retain exact console
line excerpts for all12statics. 1800s*12phases;warmup30s;CSI1280x720RG10 atconfigured60fps;
eightbuffers;320x240YUYV-likeoutput. Static pins4/1/2algorithmworkers. Common pool2
line is not algorithmconcurrency. Staticbanner explicitly noGA/AWM/Scheduler.
Shared strictwarningGPUgate threshold85inactive/50aggressive, not same fullERLpolicy.
All12manifests declare binary9d05f40d293e716f0002b7eac569ec3ea62a40d4777162600e3a6fc1b987687e,
config7e3ecb0619815da48f79b03688e5170160f98e310ac9b0843f3b60cef71eed42;
sourcehash unset-configure-time-hash, buildflagsunset. ERLdeclaresd72351786b61b1211d4eb3829257665d91ecbc85383472f55e6ecbc2deae962e.
These are record declarations, not newly hashed original binaries. Exact static
launch command/environment and binary/source-diff identity are not established.
GPU_STATIC aggressive REP1 console is truncated around80min while independent
sixhour CSV/eventstreams sealed. Do not claim fullrunconsole coverage for it.

EVIDENCE LIMITS
Equalwork: source_review/analysis/pipeline_review.txt and fullsource report establish
Sobel operator mismatch, Histogram constant-input mismatch, Gaussian halo/definition
risk; device output equivalence unverified, source-binary correspondence unresolved.
Retention: source architecture + records support state continuity; not causalbenefit.
Stats: independentunit is mission,2rep/cell; moderate hasno staticcell.
Prior18runMD has older overclaims before its reconciliation section. Prefer preserved
numeric CSVs and final source/recording audit qualifications over its summary adjectives.
