Guard - Leveling - Present (low-latency) - Summary

Convergence tuning events is the sampled subset size drawn from the preset's tuning-event pool for each convergence run.

Tail Default Convergence

Tuning events	Top winner	Unique winners	Tolerance window spread	Tail percentile spread	Median gap	Assessment
16	tail-w150-p55 (1/5)	5	150–1500 ms	21.0–55.0 %	0.042	Noisy
24	tail-w1500-p30 (2/5)	4	188–1500 ms	21.0–51.0 %	0.007	Noisy
32	tail-w1500-p35 (2/5)	4	1250–1500 ms	21.0–35.0 %	0.002	Noisy
48	tail-w1500-p27p5 (2/5)	3	1500 ms	27.5–40.0 %	0.232	Noisy
64	tail-w1500-p30 (1/1)	1	1500 ms	30.0 %	0.361	Stable

Scenario Settings

Baseline preset	Leveling - Present (low-latency)
Target loudness	-12.0 dB
Target tolerance	0.0 dB
Selection rejects on low need	187
Learned tail default	tail-w1500-p30
Learned tail tolerance window	1500 ms
Learned tail percentile	30.0 %

Tuning Search

Tolerance window coarse values	100, 150, 225, 350, 500, 750, 1000, 1500, 2250, 3000, 4500, 6000 ms
Tunable tail-percentile coarse values	1.0, 2.5, 5.0, 7.5, 10.0, 15.0, 20.0, 25.0, 35.0, 50.0 %
Rounds	Coarse coordinate search + single refinement round
Primary objective	Favor real attenuation / contrast first, then preserve target-seeking and dry restoration
Fast mode	disabled
Exhaustive default search	disabled
Report events	16 total (8 per split target)
Tuning events	64 total (32 per split target)
Minimum upward leveling need	4.0 dB at pre-event anchor
Target-tracking min work	1.0 dB guard-off movement
Under-leveling threshold	seek floor < 0.35
Min tunable tail percentile	1.0 % (0% kept for guard-off reference only)
Convergence study	5 repeats @ 16, 24, 32, 48, 64 tuning events
Tail Clip-tuned unique configs	13

Tail Configs

Learned default slug	tail-w1500-p30
Learned default tolerance window	1500 ms
Learned default tail percentile	30.0 %
Clip-tuned unique configs	13
Minimum upward leveling need	4.0 dB
Median selected upward need	5.87 dB
Clip-tuned best p25 ratio	0.75
Default p25 ratio	0.70
Default median seek floor	0.95
Default under-leveling	0.0%

Dry-relative tuning uses the raw input breath-to-anchor delta as the ideal target while retaining guard-off vs guarded attenuation and contrast metrics for continuity. Tuning candidates now require at least 4.0 dB of upward pre-event leveling need, the search is now effectiveness-first (attenuation / contrast before target-tracking tie-breakers), and 0% remains reserved for the explicit guard-off reference rather than the tunable search space. Clip-tuned best means the per-clip objective winner under the current search ordering; it is an upper-bound proxy, not a ground-truth optimum. Dry-anchor fallback was needed for 0 of 16 clips. The machine-readable tuning artifact is site/apu.software/analysis/Guard/tuning-summary.json.

Tail Guard Aggregate Metrics

Metric	Learned Default	Clip-tuned best
Median upward leveling need	5.87 dB	5.87 dB
Median guard-off pre-event target seek	5.66 dB	5.66 dB
P25 guard-off pre-event target seek	5.02 dB	5.02 dB
Median target-seek floor ratio	0.95	0.91
P25 target-seek floor ratio	0.89	0.77
Median target error increase	0.00 dB	0.00 dB
P90 target error increase	0.00 dB	0.00 dB
Under-leveling window rate	0.0%	0.0%
Applicable target-tracking windows	47	47
Median dry restoration ratio	0.76	0.78
P25 dry restoration ratio	0.70	0.75
Median dry ideal error	1.57 dB	0.96 dB
Median attenuation	5.02 dB	4.54 dB
Median contrast improvement	4.30 dB	4.36 dB
Median anchor drift	0.25 dB	0.50 dB
P90 anchor drift	1.33 dB	1.46 dB
Success rate	93.8%	100.0%
Median recovery drift	0.00 dB	0.00 dB
Median far-window drift	0.00 dB	0.00 dB

Per-clip Best-Found Settings

Clip	Best-found config
dev-clean:3752_4944_000100_000013:3	tail-w425-p52p5
dev-clean:2086_149220_000036_000001:1	tail-w1500-p50
dev-clean:5694_64029_000008_000001:1	tail-w875-p52p5
dev-clean:2428_83699_000015_000001:2	tail-w1500-p55
dev-clean:2803_154328_000007_000001:4	tail-w100-p35
dev-clean:3853_163249_000064_000000:1	tail-w1500-p36
dev-clean:1988_148538_000007_000000:3	tail-w1500-p52p5
dev-clean:3853_163249_000160_000000:2	tail-w1500-p52p5
test-clean:2300_131720_000003_000000:1	tail-w1500-p30
test-clean:4077_13754_000011_000006:1	tail-w1500-p55
test-clean:8463_294825_000024_000000:4	tail-w1500-p55
test-clean:3575_170457_000023_000002:1	tail-w1500-p51
test-clean:7176_88083_000003_000001:1	tail-w125-p51
test-clean:4077_13754_000002_000000:1	tail-w350-p30
test-clean:4970_29095_000049_000001:2	tail-w1250-p55
test-clean:4446_2275_000023_000000:2	tail-w287p5-p55

Tail Default Worst Dry-Ideal Misses

Clip	Error
test-clean:4970_29095_000049_000001:2	3.04 dB
dev-clean:2428_83699_000015_000001:2	2.91 dB
dev-clean:3853_163249_000160_000000:2	2.59 dB
test-clean:3575_170457_000023_000002:1	2.31 dB
test-clean:4077_13754_000002_000000:1	2.02 dB

Tail Default Weakest Seek Floors

Clip	Seek Floor
test-clean:4077_13754_000002_000000:1	0.66
test-clean:2300_131720_000003_000000:1	0.76
dev-clean:2086_149220_000036_000001:1	0.78
dev-clean:3853_163249_000064_000000:1	0.85
test-clean:4077_13754_000011_000006:1	0.90

Tail Clip-tuned Gain Over Default

Clip	Ratio Gain
dev-clean:2803_154328_000007_000001:4	0.40
test-clean:4970_29095_000049_000001:2	0.22
test-clean:4446_2275_000023_000000:2	0.17
dev-clean:3853_163249_000160_000000:2	0.12
dev-clean:5694_64029_000008_000001:1	0.12

Assertion Log 31 entries

Status	Time	Condition / Message
PASS	18:34:54	Guard tail default search visited 28 configs; threshold-passing=21; winner=tail-w1500-p30
PASS	18:41:05	Guard tail learned default search ordering: (0, 0, 0.0, (-0.9375, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -5.018434610813049, -3.270758967859056, -4.295226422369192, -3.058746522646758, 0.2514419461415969, 1.3311852757747582, 0.0, -5.656379967939661, -5.024765649917775, -0.9521570575706861, -0.8895073856915869, -1.0, 0.0, 0.0, -0.7593730646268183, -0.6957679256471071, 1.5678346616866445, 0.0, 0.0)) (fixed baseline (1, 1, 0.012499999999999956, (-0.6875, 0.0, 0.14981021560366337, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, -1.8304178268376496, -0.8501897843963366, -1.5971318664585699, -0.7895008762896811, 0.017386187975300338, 0.3992131506453944, 0.0, -5.656379967939661, -5.024765649917775, -0.9963489856698392, -0.9409973489488069, -1.0, 0.0, 0.0, -0.3178961306059914, -0.14063316173680218, 4.224367149192268, 0.0, 0.0)))
PASS	18:41:05	Guard tail default median attenuation: 5.02 dB (expected >= 1.50 dB)
PASS	18:41:05	Guard tail default median contrast improvement: 4.30 dB (expected >= 1.00 dB)
PASS	18:41:05	Guard tail default median anchor drift: 0.25 dB (expected <= 0.75 dB)
PASS	18:41:05	Guard tail default p90 anchor drift: 1.33 dB (expected <= 1.50 dB)
PASS	18:41:05	Guard tail default success rate: 93.8% (expected >= 70.0%)
PASS	18:41:05	Guard tail default median recovery drift: 0.00 dB (expected <= 4.25 dB)
PASS	18:41:05	Guard tail default median guard-off pre-event target seek: 5.66 dB (expected >= 1.50 dB)
PASS	18:41:05	Guard tail default p25 guard-off pre-event target seek: 5.02 dB (expected >= 1.00 dB)
PASS	18:41:05	Guard tail default median target-seek floor ratio: 0.95 (expected >= 0.75)
PASS	18:41:05	Guard tail default p25 target-seek floor ratio: 0.89 (expected >= 0.50)
PASS	18:41:05	Guard tail default under-leveling window rate: 0.0% (expected <= 20.0%)
PASS	18:41:05	Guard tail default median target-error increase: 0.00 dB (expected <= 0.75 dB)
PASS	18:41:05	Guard tail default p90 target-error increase: 0.00 dB (expected <= 1.50 dB)
PASS	18:41:05	Guard tail clip-tuned median attenuation: 4.54 dB (expected >= 1.50 dB)
PASS	18:41:05	Guard tail clip-tuned median contrast improvement: 4.36 dB (expected >= 1.00 dB)
PASS	18:41:05	Guard tail clip-tuned median anchor drift: 0.50 dB (expected <= 0.75 dB)
PASS	18:41:05	Guard tail clip-tuned p90 anchor drift: 1.46 dB (expected <= 1.50 dB)
PASS	18:41:05	Guard tail clip-tuned success rate: 100.0% (expected >= 70.0%)
PASS	18:41:05	Guard tail clip-tuned median recovery drift: 0.00 dB (expected <= 4.25 dB)
PASS	18:41:05	Guard tail clip-tuned median guard-off pre-event target seek: 5.66 dB (expected >= 1.50 dB)
PASS	18:41:05	Guard tail clip-tuned p25 guard-off pre-event target seek: 5.02 dB (expected >= 1.00 dB)
PASS	18:41:05	Guard tail clip-tuned median target-seek floor ratio: 0.91 (expected >= 0.75)
PASS	18:41:05	Guard tail clip-tuned p25 target-seek floor ratio: 0.77 (expected >= 0.50)
PASS	18:41:05	Guard tail clip-tuned under-leveling window rate: 0.0% (expected <= 20.0%)
PASS	18:41:05	Guard tail clip-tuned median target-error increase: 0.00 dB (expected <= 0.75 dB)
PASS	18:41:05	Guard tail clip-tuned p90 target-error increase: 0.00 dB (expected <= 1.50 dB)
PASS	18:41:05	Guard tail learned default dry metrics: median ratio=0.76, p25 ratio=0.70, median error=1.57 dB, seek floor=0.95, under-leveling=0.0%
PASS	18:41:05	Guard tail clip-tuned dry metrics: median ratio=0.78, p25 ratio=0.75, median error=0.96 dB, seek floor=0.91, under-leveling=0.0%
PASS	18:41:05	Guard Leveling - Present (low-latency) defaults: tail=tail-w1500-p30