JSLMPR cf8ec76430 fix(vision): Qwen2.5-VL llama.cpp backend — verified to break the moondream ceiling ... Provisioned and tested the stronger Tier-2 judge on this Intel Mac (CPU-only): - vision_caption_llamacpp.py: force CPU (-ngl 0 --no-mmproj-offload) because the integrated GPU times out on the vision encoder (Metal command-buffer timeout); extract the final assistant turn from the chat-templated output. - docs/LOCAL-MODELS.md: the VERIFIED build+run recipe, incl. two real gotchas -> Command Line Tools libc++ mismatch (add -isystem <SDK>/usr/include/c++/v1 to the cmake flags, else ggml-base fails on <array>), and the Intel-GPU Metal timeout. Verified result: where moondream described the bowling celebration as "standing in a bowling alley", Qwen2.5-VL-3B says "raising their arms in a celebratory gesture" (highlight-worthiness 1.0) and rates turn-around/anticipation low (0.2 / 0.15). End to end, the director's judge now chooses time=10.5s score=1.0 (the celebration) vs moondream's 13.0s score=0.15 (the turn-away). Model weights are gitignored under models/ (Apache-2.0, provisioned offline). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>		2026-07-25 23:50:12 +02:00
..
LOCAL-MODELS.md	fix(vision): Qwen2.5-VL llama.cpp backend — verified to break the moondream ceiling	2026-07-25 23:50:12 +02:00
RUNBOOK-highlight-e2e.md	docs: portable e2e runbook + local-models reference (any-AI usable)	2026-07-24 15:47:45 +02:00
alerts.md	Added the next operational-readiness artifact in docs/alerts.md and marked Alerts complete in docs/video-clipping-service-implementation-plan.md:722.	2026-07-10 00:20:31 +02:00
autoscaling-policies.md	• The next step is done and the plan is updated. Autoscaling policies is now marked complete in docs/video-clipping-service-implementation-plan.md:717.	2026-07-09 19:00:31 +02:00
backup-and-restore-validation.md	• Added the backup/restore validation artifact at docs/backup-and-restore-validation.md and marked Backup and restore validation complete in docs/video-clipping-service-implementation-plan.md:724.	2026-07-10 00:27:26 +02:00
cinematic-editing-runbook.md	- Implement end-to-end cinematic highlight rendering pipeline	2026-07-12 01:20:18 +02:00
cinematic-highlight-acceptance-review.md	docs: add P3.6 acceptance review (frozen thresholds + rubric); Gate A measured PASS	2026-07-22 11:59:06 +02:00
cinematic-highlight-operator-checklist.md	add asset generation	2026-07-21 10:03:27 +02:00
cinematic-highlight-poc-plan.md	Automatic two-tier highlight director + source-adaptive rendering	2026-07-23 22:49:05 +02:00
cinematic-quality-rules.md	refactor(director): decisive moment = measurement proposes, vision model judges	2026-07-25 18:37:38 +02:00
cinematic-video-editing-service-plan.md	• working version but not cinematic	2026-07-11 00:32:46 +02:00
cost-model-by-video-minute.md	I added a reproducible cost-model harness in src/test/java/org/example/videoclips/perf/CostModelBenchmarkHarness.java:1 that turns the checked-in FFmpeg and object-storage benchmark baselines into per-source-minute estimates	2026-07-09 19:04:23 +02:00
current_state.md	feat(cinematic): R10–R14 grade/cadence/edit rules + 420p compat fix	2026-07-25 12:36:18 +02:00
dashboards.md	Updated the prompt in docs/grafana-dashboard-creation-prompt.md so the VPS AI now has the full next-step instructions: generate the dashboard JSON, save it to dashboards/video-clipping-overview-grafana.json, and verify the	2026-07-10 00:17:51 +02:00
dlq-redrive-procedure.md	• Added the DLQ redrive artifact at docs/dlq-redrive-procedure.md and marked DLQ redrive procedure complete in docs/video-clipping-service-implementation-plan.md:725.	2026-07-10 00:29:26 +02:00
ffmpeg-preset-benchmark-baseline.md	• The next plan item is done and the plan is updated. Object-storage bandwidth tests is now marked complete in docs/video-clipping-service-implementation-plan.md:713.	2026-07-09 01:33:24 +02:00
gate-b-review-bowling.md	feat(cinematic): R10–R14 grade/cadence/edit rules + 420p compat fix	2026-07-25 12:36:18 +02:00
grafana-dashboard-creation-prompt.md	Updated the prompt in docs/grafana-dashboard-creation-prompt.md so the VPS AI now has the full next-step instructions: generate the dashboard JSON, save it to dashboards/video-clipping-overview-grafana.json, and verify the	2026-07-10 00:17:51 +02:00
highlight-rendering-gap-closure-plan.md	Add highlight rendering gap closure plan	2026-07-11 17:26:36 +02:00
input-folder-scheduler-plan.md	Configure folder scheduler in YAML	2026-07-10 12:24:16 +02:00
load-test-signoff.md	• Added the signoff artifact at docs/load-test-signoff.md and marked Load-test signoff complete in docs/video-clipping-service-implementation-plan.md:726.	2026-07-10 00:30:56 +02:00
local-cinematic-asset-worker-prompt.md	add asset generation	2026-07-21 10:03:27 +02:00
local-cv-visual-analysis-guide.md	Add local CV worker readiness checks	2026-07-11 17:46:40 +02:00
object-storage-bandwidth-benchmark-baseline.md	• The next plan item is done and the plan is updated. Object-storage bandwidth tests is now marked complete in docs/video-clipping-service-implementation-plan.md:713.	2026-07-09 01:33:24 +02:00
production-cinematic-ai-director-prompt.md	- Implement end-to-end cinematic highlight rendering pipeline	2026-07-12 01:20:18 +02:00
production-cinematic-highlight-editing-plan.md	- Implement end-to-end cinematic highlight rendering pipeline	2026-07-12 01:20:18 +02:00
queue-visibility-timeout-tuning.md	• I completed the next plan item and updated the checklist. Queue visibility-timeout tuning is now marked done in docs/video-clipping-service-implementation-plan.md:716.	2026-07-09 18:37:57 +02:00
runbooks.md	• Added the runbook artifact at docs/runbooks.md and marked Runbooks complete in docs/video-clipping-service-implementation-plan.md:723.	2026-07-10 00:22:35 +02:00
video-clipping-service-implementation-plan.md	• clip video locally plan	2026-07-10 00:49:29 +02:00
video-clipping-service-planning-prompt.md	initial	2026-07-08 17:00:58 +02:00
worker-cpu-disk-benchmark-baseline.md	I continued with the next plan step and updated the plan.	2026-07-09 01:25:04 +02:00