Venn

Decisions over stored embeddings: review assets#

The video and methods page links each beat to its source demo, numeric evidence and limitations. These assets are a private release preview. The owner confirmed clearances on November 7, 2026, and authorized uploading both MP4s to the private media-v1 release. Public launch remains on hold.

Recompose#

Use the checkout's existing dependencies; the arena extra provides Pillow and imageio-ffmpeg. Install Lato (the arena video's font) and DejaVu Sans Mono system fonts to reproduce the typography.

pip install -e ".[arena]"
python examples/same_memory_video/render.py --preview --output /tmp/same-memory-preview
python examples/same_memory_video/render.py --output /tmp/same-memory-render

The renderer reads docs/media/same-memory/*-eval.json, head-execution.json and the displayed function in head_demo.py. It makes no network or model calls, and writes both landscape and vertical cuts, posters, storyboards, SRT/VTT captions and a SHA-256 manifest. Temporary synthesized music WAVs are removed after encoding. Source speech is never played. The motion-led revision follows an inspection of the actual encoded Doom cut: continuous record flow, retrieval trails, data morphs and processing lanes, with rhythmic music and synchronized effects. Frames are rendered throughout each shot, not frozen after an entrance animation. There are no on-screen publication overlays or fine-print footers; those disclosures remain on the methods page. All text is at least 48 pixels at the final output scale. Frame rendering and encoding are CPU work; there is no GPU workload to distribute. Existing output files are overwritten. The landscape cut runs 2:30 and the vertical cut 1:00. As with the arena video, keep MP4s out of Git and publish approved binaries on GitHub Releases. Ignored copies under docs/media are for local review; do not upload these drafts until the owner explicitly requests publication. Source links use the intended release URLs, which are not available until upload. The site builder prefers local MP4s when present and otherwise uses those release URLs, just as a source-only deployment requires.

The opening sets up the comparison, explains Jev's question-and-choice interface, and then shows Venn with the identical example on a separate screen. Input, question, choices, reader, embeddings and head appear in order, with moving signals and a focus cue guiding the eye before the answer appears. The input and highlighted choice are authored illustrations, not recorded API responses. All frames are checked for overlapping text, padded answer-box and node containment, metric-column bounds and separation from the reserved narration area. Answer choices are stacked in both formats. Narration uses one cool-blue regular-weight caption block, visually distinct from graphic labels, without sentence-ending periods. Connected title lines share the same font and color; redundant explanation below the architecture is removed. Site caption tracks are opt-in, not a second default overlay over the video's existing captions. Score error is labeled in points with "Lower is better", without an on-screen metric acronym. The three examples then show why the difference matters: historical outcomes, learned label conventions and reusable custom heads. The audio beat and its closing claims are removed from both cuts. Its 20 landscape seconds and 16 portrait seconds are redistributed to reading time without lengthening either cut. A subsequent four-second landscape trim removes the summary animation and shortens the end card, not the explanations. Architecture stages now unfold over eight seconds; the code highlight makes a ten-second pass and holds the completed registration for four seconds. The filing gallery reveals companies one at a time, then connects them to the shared recorded Business Services label. Each new dataset and target is introduced before its comparison. Complaint label grouping precedes outcome retrieval; its wipe affects only that lookup, and the vertical cut restores it before closing. SEC fitting precedes its separately labeled test sets, and the company gallery does not attribute a linear-head prediction to retrieved neighbors. Wine preparation is explicitly a separate historical comparison, and price shows its distinct cached vectors.

The code scene is taken verbatim from register_head in head_demo.py. It was executed on the original fitting notes using their existing SentenceSimilarity cache, not a mock API. The execution record includes the source hash and head checksum; the renderer refuses a stale record. The replay is compressed and does not display a real-time clock. To execute it with your own local fitting notes and prepared embedding cache:

python examples/same_memory_video/head_demo.py \
  --examples /path/to/fitting-examples.json \
  --model-dir /path/to/venn-head \
  --cache /path/to/cache.sqlite \
  --output /tmp/head-execution.json

Writing to docs/media/same-memory/head-execution.json would replace the tracked record that the renderer uses.

The examples file is a JSON list of {"state": "note text", "points": 88} records. The cache must already contain their pinned SentenceSimilarity embeddings; cache misses fail instead of starting another reader run. This only fits the head. It does not rerun or replace the held-out studies. Do not treat a run on different examples as reproduction of the video results.

The three active notebooks can be opened in Jupyter from this checkout, or their code cells can be executed without Jupyter:

python - <<'PY'
import json
from pathlib import Path

for beat in ("history", "sec", "wine"):
    notebook = json.loads(Path(f"examples/same_memory_video/{beat}.ipynb").read_text())
    namespace = {}
    for cell in notebook["cells"]:
        if cell["cell_type"] == "code":
            exec(compile("".join(cell["source"]), beat, "exec"), namespace)
PY

These notebooks verify exported point estimates, not end-to-end study reproduction, original wall-clock performance or bootstrap intervals. Numeric exports carry original source hashes. The SEC notebook reports the failed primary gate rather than promoting confirmation. The archived sound notebook and eval file retain the earlier draft's provenance but are not used by the current renderer. No new fit is run during rendering.

The public export deliberately excludes narratives, tasting notes, source audio, raw vectors, checkpoints and credentials. Waveform envelopes retain CREMA-D's ODbL/DbCL terms and credit, as documented on the methods page. Keep the methods page's local-review and publication-hold notices until release is authorized.

View this page on GitHub

On this page