Field notes · Gary Engineering Lab
Gary Engineering Lab #003: Finding Ranch Moments in 28 GB of Camera Footage — and Catching AI False Positives
Gary Engineering Lab #003 · October 11, 2026 · Media retrieval and computer-vision experiment · Status: completed first-pass scan; human review required
At Freedom’s Purchase Ranch, the most useful AI workflows are often the unglamorous ones: finding valuable moments in large collections of everyday footage, preserving their context, and making the results easy for a person to review. This experiment tested a local-first computer-vision pipeline on a removable camera card without uploading source media or interrupting our production AI inference.
The workload
The camera card contained 74 H.264 videos and 75 still photos, captured across October 9–11, 2026. Video alone occupied approximately 28 GiB. Most recordings were roughly ten-minute segments from a stationary interior camera recording real ranch work.
The question: Can we identify likely segments containing people or animals, copy useful surrounding footage into a private editing workspace, and build reviewable compilations without modifying the originals?
Local pipeline and reproducible controls
- Discover sources: enumerate camera originals read-only and inspect streams with FFprobe. The primary video stream is 2304×1296, with an additional lower-resolution video stream at 320×176.
- Sample efficiently: FFmpeg decodes the secondary stream and samples a frame approximately every 15 seconds. The model operates on scaled proxy frames to keep decoding and inference practical on CPU.
- Classify candidates: a local, COCO-trained YOLOv8n object detector flags possible people and several common animal classes. It is not a face-identification system.
- Preserve context: time-adjacent candidate hits are merged into intervals with lead-in and follow-through; FFmpeg stream-copies the original high-resolution camera footage into separate private clips.
- Enable review: the output includes a searchable local HTML gallery, video posters, original-source references, machine-readable detection receipts, checksums, and private rough-cut edits.
Detection ran on the CPU, leaving GPU model routing unchanged. No original media files were modified, and no private camera footage was published to the website or social media.
Measured first-pass outcome
| Metric | Observed |
|---|---|
| Video originals | 74 |
| Still-photo originals | 75 |
| Video frames sampled | 2,942 |
| Sampled frames with a person detection | 496 |
| Sampled frames with an animal-like detection | 16 (all unverified as animals) |
| Extracted review-clip intervals | 132 |
| Extracted review footage | Approximately 173 minutes / 7.05 GB |
| Still-photo candidates copied for private review | 58 |
| FFmpeg reported extraction errors | 0 |
| FFprobe-validated extracted clips | 132 of 132 |
The 132 outputs are review candidates, not 132 verified publishable highlights. Some intervals include empty moments as context, while others may include unintended bystanders, private conversations, or visible operational details.
A revealing failure: the “animal” that was not an animal
The local detector repeatedly confused human hair, a bent-over worker, and ordinary equipment with cats, dogs, birds, or other animals. High-resolution spot checks disproved several of these predictions. Thus, the honest current animal result is no animal independently confirmed from the reviewed samples, not “16 animals filmed.” The detector’s labels are hypotheses for human examination.
This is exactly why we are developing FPR Mechanic and an evidence-first publishing workflow. A confidence score is not ground truth, and a detector that appears certain can still be wrong. An autonomous content pipeline must preserve the difference between predicted, reviewed, and approved to publish.
Private editorial outputs
To make the results useful rather than merely statistical, the system generated a local gallery covering all 132 video intervals, a separate selection of 58 stills, and an eight-scene 63.5-second horizontal rough cut. A matching vertical version preserves the full scene against a blurred background so it can be evaluated as a potential short-form format without cropping workers out of the frame.
Both edits are private prototypes with recorded audio removed. Their footage, captions, pacing, subject releases, and consent must be reviewed before any public release.
Limitations and what we are improving
- Recall is not proven. Sampling every 15 seconds can miss a brief appearance; especially small animals may be missed in the low-resolution secondary stream.
- Precision is not established. A systematic human-labeled ground-truth sample has not yet been scored; some object classes generated conspicuous false positives.
- “Funny” cannot be inferred reliably from a person box. We need editorial assessment of motion, reactions, pacing, context, audio, dignity, and participant consent.
- Timing needs editorial trim. Stream-copy extraction preserves resolution but may seek to nearby keyframes; cut-in and cut-out should be checked by a human.
- Release is separate from extraction. Even technically successful output remains private until identity, location, audio, rights, and consent checks are complete.
Open-source design implications
A broadly useful FPR Mechanic media module should provide a removable-drive importer, multiple detector backends, immutable source manifests, configurable sampling policies, explicit ground-truth review, one-click extraction, downloadable editing manifests, and consent-based export gates. Lightweight installation and independent modules matter as much as AI accuracy. Detector model licenses and redistribution conditions also need review before packaging a public distribution.
What passed today: local inventory, reproducible first-pass detection, non-destructive extraction, 132/132 clip-format checks, and private preview generation. What did not pass: exhaustive person/animal discovery, confirmed wildlife identification, human rights clearance, or public posting authorization.
Follow Gary Engineering Lab for engineering records that include the false starts alongside the working systems. Our guiding rule remains: show the evidence, document the limits, and never call a prediction a verified result.