docs(archive): record transcription statistics and processing metrics

Author: just3ws

backlog/docs/retrospectives/transcription-stats-2026.md

@@ -0,0 +1,21 @@
+# Transcription Archive Statistics (May 2026)
+
+This document captures a snapshot of the archive's transcription metrics during the implementation of the asynchronous `zdots-ctx` + `whisper.cpp` pipeline.
+
+## ⏱️ Video Runtime vs. Processing Time (The Hardware)
+Based on logs from the background worker running the `max-accuracy` (Large v3) Whisper model on an Apple M4 GPU:
+*   **Sample Video Duration:** 865.6 seconds (~14.5 minutes).
+*   **Processing Time:** Finished transcription and diarization in 213.6 seconds (~3.5 minutes).
+*   **The Ratio:** The system transcribes audio at roughly **4x real-time speed**. For every 4 minutes of video, it takes 1 minute to generate a high-accuracy text record.
+
+## 📚 Video Duration vs. Transcript Length (The Content)
+Calculated across the `_data` models before the final batch of 40 videos completed:
+*   **Known Transcribed Video Duration:** 16 hours, 42 minutes (across the 94 videos that have exact durations logged).
+*   **Total Transcript Length:** **247,896 words** across 159 transcript files.
+*   **The Density Ratio:** On average, interviews generate **~247 words per minute of video**.
+
+*Perspective:* The archive is currently sitting at roughly the equivalent of **4 to 5 full-length non-fiction books** worth of high-density technical conversation. Once the final batch of 40 videos finishes processing, the archive is expected to cross the 300,000-word mark.
+
+## Future Milestones
+- Re-review previously transcribed interviews (legacy transcripts) for quality control and phonetic accuracy against the newer Whisper v3 baseline.
+- Verify the completion and ingestion of the SCMC and general archive videos.