📝 DMS-NX__ISE_v9.6__Checkpoint_2026-09-25.mdTheme:

DMS/NX — Chapter 2 Session Checkpoint

Session: 2026-09-25 | Andrew.human & Claude.ai
Continued from: ISE_Checkpoint_2026-09-24.pdf (Chapter 1 close-out)

What this session covered

Continuation of the ISE/DMS-NX project. Five main threads:

DMS/NX Chapter 2 Master Post — new forum post written
Classification_Progress_Report v2.3 analysis — 925 unexplained store IDs root-caused
Audio classification pipeline — CLAP feasibility tested; AudioProcessor design started
Clone Python sync — scripts written, not yet run (mirror first)
Dev Python sync — script written, not yet run

Standing policies (carried forward, unchanged)

No manual edits to SMF code/files ever — fixes go through the Package Manager/install-script system only.
Deployment/hotfix timing on production systems is Andrew's call, not something to be directed.
No sycophancy — claims checked against real source before being agreed with, including Andrew's own.
Never reuse an old name for an updated transcript — each update gets the base name with a number appended (e.g. <name>_2, <name>_3); never overwrite an earlier file.

DMS/NX rebrand

DMS/NX = Document Management System / NeXt generation. This is a rebrand of ISE only — nothing replaced, nothing removed. Same engine, same indexes.

Five search sources:

Source	Searches	Notes
ISE (Posts)	Forum post text	Original engine
PDF	Text inside PDF attachments
Misc	Text inside flat-file attachments	Any plain-text: .json, .txt, .py, etc. Excludes .md, .html
ISEmega	Posts + PDFs + Misc combined	Soft cap 180 results
ISEmedia	Images and video	CLIP-based; keyword gate required before sort

Chapter 2 Master Post written and delivered: DMS-NX_Chapter2_Master_Post.md
Chapter 1 link left as placeholder — Andrew handles it.

Classification_Progress_Report v2.3 — root cause analysis

Output state when Andrew stopped: 98.5% complete, classifier stopped, 925 unexplained store IDs.

Root cause — mismatch between inventory and driver:

Category	Count	Cause
No post / avatar	351	`id_msg=0` rows — avatars excluded by `build_inventory()` but driver processes them
Thumbnails	349	`attachment_type=3` rows — thumbnails excluded by inventory but not driver
`ext=webp`	223	`IMAGE_EXT` in CPR v2.3 = `{jpg,jpeg,png,gif}` — webp missing
`ext=3gp`	2	`VIDEO_EXT` in CPR v2.3 = `{mp4}` — 3gp missing
Total	925

Fix design (not yet implemented):

Shared extension map in core_utils.py (so CPR and driver always agree)
build_inventory() must add webp and 3gp to its extension sets
Driver must exclude avatars (id_msg=0) and thumbnails (attachment_type=3) at index time, matching inventory

Audio classification — CLAP pipeline

Lay of the land

49 audio files on the forum attachment store (confirmed via audio_manifest.tsv)
44 × MP3 (644.9 MB) + 7 × WAV (173.1 MB)
Almost all music

Model choice

CLAP — laion/larger_clap_music (music-tuned; fallback: laion/clap-htsat-unfused)
Same 512-dim output vectors as CLIP — but must go in a separate store (audio_store.json) because CLIP and CLAP vectors are not comparable (different model families — mixing them gives meaningless cosine similarities).

Feasibility test on Live

clap_smoketest.py run on Live (/root/venv-audio-test/bin/python):

Peak RSS: 1484 MB on a 2 GB box with 813 MB swap already in use
Time per 10s window: 23.4 s
Decision: do NOT run AudioProcessor on Live

Plan: run on Clone, push back to Live

Clone has 6 GB RAM and ffprobe. Workflow:

Run AudioProcessor on Clone
rsync push audio_store.json back to Live
MediaSearch on Live reads it alongside existing stores

CLAP on Clone — setup (done on Live, to be replicated on Clone)

# torch stub (source-built torch has no dist-info — transformers can't detect it)
D="/root/venv-audio-test/lib/python3.8/site-packages/torch-1.10.0a0+git71f889c.dist-info"
mkdir -p "$D"
printf 'Metadata-Version: 2.1\nName: torch\nVersion: 1.10.0a0+git71f889c\n' > "$D/METADATA"

# tokenizers: use manylinux2010 binary wheel (avoids Rust build on glibc 2.12)
pip install --only-binary tokenizers "tokenizers<0.14"   # picks up 0.13.0

Verified on Live: torch seen: True | ClapModel from: transformers.models.clap.modeling_clap

Audio manifest

Built on Live:

/var/www/html/ISE_Data/audio_manifest.tsv   (49 rows)
fields: id_attach, id_folder, file_hash, filename, fileext, id_msg

49 (not 51) — 2 disk files excluded (non-post-attached or wrong extension).

AudioProcessor.py — PENDING

Not yet written. Still waiting on:

MediaIndexer.py source (not yet supplied)
Decision: whole-track embedding vs. sliding-window moment-in-mix search

MediaSearch.py changes needed (v1.4.0 baseline)

VIEWER_AUDIO_EXTS already defined (mp3, wav, flac, ogg, m4a, aac, wma) — audio playback in results supported ✓
TYPE_EMOJI needs an audio entry (not yet added)
Audio store keys must not use "_t" suffix (that's the video-thumbnail convention; is_video check keys off it)

ffprobe missing on Live

MediaProcessor.py's get_video_duration() silently falls back to 60 s when ffprobe is absent — wrong frame timestamps for short videos. Clone has ffprobe. Pending fix: copy from Clone; update MediaProcessor to report failure unconditionally (not silently).

Python environment — three machines

Interpreters

Interpreter	Live	Clone	Dev
python3 symlink	→ 3.6 ✓	→ 3.8 (wrong; fixed by clone_py_sync.sh)	→ 3.6
python3.6	/usr/local/bin ✓	/usr/local/bin	/usr/local/bin
python3.8	/usr/local/bin ✓	/usr/local/bin	—
python3.12	/opt/python-3.12.11	(to be synced)	—
torch	1.10.0a0+git71f889c (source build)	(to be synced)	NOT copied
venv-audio-test	✓ (CLAP installed)	(to be synced)	NOT copied

Dev is i686 (32-bit) — cannot copy Live's compiled binaries. Uses pip --only-binary --no-deps instead.

Key glibc constraint

All three machines: glibc 2.12. torch was source-built specifically to avoid the glibc 2.14 requirement of pip-installable torch wheels.

Scripts delivered (all v1.0.1)

Script	Location	Status
`py_inventory.sh`	`/mnt/user-data/outputs/`	Ready; run on each machine
`clone_py_sync.sh`	`/mnt/user-data/outputs/`	Ready; NOT yet run
`dev_py_sync.sh`	`/mnt/user-data/outputs/`	Ready; NOT yet run
`clap_smoketest.py`	`/mnt/user-data/outputs/`	Run on Live (1484 MB); pending on Clone
`DMS-NX_Chapter2_Master_Post.md`	`/mnt/user-data/outputs/`	Delivered

v1.0.1 changes (both sync scripts):

alias step removed — Andrew maintains aliases manually on each machine
--info=stats1 → --stats (CentOS 6 rsync 3.0.x compat)
cd / before import tests (prevents stdlib shadowing by stray random.py)

Pending tasks (in order)

Clone: webroot + DB mirror (Clone pulls from Live) — do this FIRST, before Python sync
bash clone_py_sync.sh --dry-run on Clone → review output → run live
bash py_inventory.sh on Clone → diff against /root/py_inventory_forum.txt
Clone CLAP smoke test: clap_smoketest.py on Clone to get real time-per-window
Dev: webroot + DB mirror then bash dev_py_sync.sh → review /root/dev_py_sync_report.txt
AudioProcessor.py — design + write (needs MediaIndexer.py source first)
Fix 925-ID self-check — shared extension map in core_utils; add webp/3gp; exclude avatars/thumbnails at index time
ffprobe on Live — copy from Clone; fix MediaProcessor silent fallback
Conversation transcript (.md + .pdf) — not yet written

Key file paths (Live, unless noted)

/var/www/html/ISE_Data/
    hf_cache/                   CLAP + CLIP model cache
    audio_manifest.tsv          49-row audio index (built this session)
    embeddings_store.json       CLIP image/video vectors (512-dim)
    captions_store.json         SmolVLM captions
    audio_store.json            CLAP audio vectors (512-dim) — TO BE BUILT ON CLONE

/root/venv-audio-test/          Python 3.8 venv with CLAP stack
/root/models/smolvlm/           SmolVLM model files

Live IP: 128.199.200.202

DMS/NX Chapter 2 session checkpoint — 2026-09-25