ON: wait for wall power to settle back to the idle floor before each next pass (active-probe, the /rag /llm compare technique). OFF: use the fixed rest periods below. Variance calibration always keeps its fixed protocol.
ON
fixed rest between CPU and GPU runs (used when wait-for-idle is OFF; also the fallback after an idle-wait timeout)
10s
fixed pause between LLM batch / compare runs (used when wait-for-idle is OFF; also the timeout fallback)
10s
idle-wait: settle when a reading is within this of the captured floor
3W
idle-wait: consecutive in-band reads needed to confirm settle
3polls
idle-wait: cap before timeout → dialog (Lab) or fixed fallback
120s
idle-wait timeout dialog: auto-apply the fallback if no operator answer within this
75s
Show the live idle-wait readout (⏳ waited · current W · target) in the progress widget on every page during cooldowns. Display only — cooldown behaviour is unchanged.
ON
Decode rig
meter sampling cadence for decode rows (device and screen meters)
1s
fixed settle after device prepare, before idle guard / baseline
5s
baseline samples per row (× cadence = baseline window)
20samples
ON: wait for the device's draw to stabilise before each baseline — the same settle loop as the GoS1 pre-job guard (idle_wait.py), in self-stability mode; rows stamp protocol v3. OFF: fixed settle only (July/v2 protocol).
ON
idle guard: the last N readings must span ≤ this
0.5W
idle guard: N consecutive readings that must agree
4polls
idle guard: cap before proceeding unsettled (recorded in the row)
30s
screen rows: skip after playback start (absorbs the 1080p mode re-sync)
5s
Tapo P110 acting as the strip's master SWITCH (wall → this plug →
Shelly meter → strip). Empty = no switchable master (installed Shelly is metering-only).
The /decode Rig on/off button appears automatically when set.
Idle auto-off. The rig is off by default. Activity = a Lab control op or
/decode visit, an OWL decode job, a bench.py row (CLI campaigns touch /tmp/owl-rig-hold),
or a box powered from outside the UI. After the idle window every powered box is gracefully
stopped, then the master if one is switchable. Countdown shows on /decode.
ON: auto power-off the rig after the idle window below. OFF: boxes stay on until someone turns them off (how Ben found the whole rig powered after a week away, 2026-08).
ON
idle window before the rig auto-powers off
4h
ON: the idle auto-off also cuts the shared screen's plug (Lab-E). OFF by default — that panel doubles as the household TV / Mac extension, not just the bench monitor.
OFF
HDMI inputs. The shared screen has 4
HDMI sockets and the rig has more external devices than that — pick which device is cabled to each.
A device on no socket is headless-only: no Claim screen, no screen-mode rows. Takes effect
within ~10 s (poller); the C2 is the screen itself and never appears here.
Dummy sinks. A box on no HDMI socket above is not automatically
sink-less: an HDMI dummy/EDID plug is a real sink, and a box with no sink at all measures in a
different regime (Fire TV plays 0.77 W lower, Gen 2 0.42 W — JOURNAL S73). Tick the boxes
wearing a dummy plug so their rows are stamped sink: dummy instead of
none. Fitted 2026-09-05; the dummies negotiate 1080p60 where the panel gives 4K60,
so dummy rows still group separately from panel rows.
✓
✓
—
—
—
—
✓
✓
✓
Staging
auto-lower /tmp/owl-maintenance after this much Lab inactivity (CR-015 watchdog)
30min
Encoding targets
ABR target bitrate applied to both CPU and GPU presets for each codec — ensures apples-to-apples energy comparison. Custom ffmpeg commands on the video page override these.
H.264 target bitrate (libx264 + h264_nvenc)
4000kbps
H.265 target bitrate (libx265 + hevc_nvenc)
2000kbps
AV1 target bitrate (libsvtav1 + av1_nvenc)
1500kbps
operator quality target — the VMAF an encode should hit while minimising energy. Anchors the /video/budget calculator and the encode-parity/calibration study; display/analysis anchor only, does not change what /video encodes per-run. ⚠ calibrated in vmaf_v0.6.1 terms — live scoring moved to VMAF v1 (vmaf_model setting) 2026-07; revisit at the next re-calibration.
92VMAF
Confidence thresholds — CI model (CR-028 Phase 2)
🟢 min confidence_positive Φ(z) that task draws above idle
0.95P
🟡 min confidence_positive
0.8P
🟢 minimum task polls (both models)
9polls
🟡 minimum task polls (both models)
4polls
calibrated idle noise floor — feeds the CI model (SE_calibrated)
2.26 %
between-window drift CV — feeds SE_drift in the CI model
1.44 %
run-level repeatability CV — NOT used in the single-run flag (reserved for aggregate layer)
1.61 %
run-level repeatability CV — NOT used in the single-run flag (reserved for aggregate layer)
1.75 %
Confidence thresholds — legacy variance model (fallback for results without raw samples)
legacy composite variance — used only when a run has no raw samples
1.87%
🟢 (legacy) ΔW must exceed this multiple of noise floor
5× noise
🟡 (legacy) ΔW must exceed this multiple of noise floor
2× noise
Variance calibration
Runs H.264 CPU then H.265 GPU on Meridian N times, sampling raw P110 readings throughout.
Writes Variance Idle % (the idle noise floor) and Variance Idle Drift %
(between-window drift) — these feed the live CI confidence model (SE_calibrated + SE_drift).
Also records per-codec repeatability CVs (CPU/GPU ΔW), reserved for a future aggregate layer — not used in
the single-run flag. The composite Variance % (mean of the three) now feeds only the
legacy fallback for results saved without raw samples. Queue is blocked for the duration.
number of H264-CPU + H265-GPU run pairs · steps of 2 (a pair needs ≥2) · 0 disables calibration here and in the benchmark
20runs
cooldown between each run pair
50s
Mirrors the /video H.264 CPU preset · bitrate from h264_bitrate_kbps · {input}/{output} substituted at runtime
Mirrors the /video H.265 GPU preset · bitrate from h265_bitrate_kbps · {input}/{output} substituted at runtime
/usr/local/bin/ffmpeg-master -y -hwaccel cuda -hwaccel_output_format cuda -i '{input}' -vf scale_cuda=-2:1080 -c:v hevc_nvenc -b:v 2000k -g 120 -profile:v main -bf 2 -rc cbr -c:a aac -b:a 128k '{output}'
Calibration requires lab access.
Overnight benchmark
Full-pipeline benchmark (CR-061): variance calibration → video all-codecs × reps × sources → LLM/RAG/image compare panels.
Runs as one queue job that blocks other runs; follow it on /queue-status, view results at /benchmark.
⚠ calibration is ambient-sensitive — don't launch during a heat wave.
Which measures & sources run is set by bench_run_* / bench_sources in settings.json.
video all-codecs repeats per source
10reps
Benchmark requires lab access.
Tier limits
CR-001 part D — concurrent-job caps and upload-size caps per audience
tier. Anonymous keyed by IP, Member by email, Lab uncapped.
concurrent (queued + running) Anonymous jobs per IP
1jobs
concurrent jobs per Member email
4jobs
Anonymous /video/upload byte cap
124MB
Member /video/upload byte cap
1024MB
Member /enhance-run upload byte cap (Lab uncapped)
1024MB
Member /enhance-run clip-duration cap — runs are 1×-paced, so duration ≈ processing time (Lab uncapped)
180s
un-kept /enhance-run uploads are swept this long after their run (keeps the source comparable on the result card)
12h
Models (CR-050)
Model enable/disable matrix (LLM · RAG · Image) — expand
Auto-discovered from ollama list and the HuggingFace cache (60 s TTL). To add a new LLM: ollama pull <name> on the server. To add an image model: download into ~/.cache/huggingface. Reload this page to see new entries. Display-only — sign in as Lab to edit.
LLM models
Drives /llm and /llm/compare. Source: ollama list.
RAG models
Drives /rag and /rag/compare. Same Ollama catalog as LLM; can be a different subset.
OWL ← Home