# Evidence used for this offline guide

Board root: `/home/aionic/work/vlm/`. Read on 2026-10-23.

- `smolvlm_camera_v124/README.md`, `METHODOLOGY.md`, `run_demo.sh`,
  `reports/counts_comparison.json`: selected patch/attention paths, validation,
  PSS engine counts. SDK/trained assets depend on kept `resident_vreg_v120`.
- `smolvlm_camera_v125/README.md`: persistent owner, host snapshots, shared
  arena, phase-switch copying, saved-frame comparison to v124.
- `smolvlm2_demo_v139/README.md`: matching pinned checkpoint, compact LM head,
  lossless storage versus runtime expansion, memory, captions, timing and counts.
- `smollm_chat_v143/README.md`, `VOICE_AND_PERFORMANCE.md`, `chat_voice.py`:
  v144 voice extension, English ASR limits, active launchers, compact kernel.
- `smollm_voice_v144/`: PocketSphinx private runtime, not a separate decoder.
- `llmjp_chat_v145/README.md`, `METHODOLOGY.md`, `VOICE.md`, `model/manifest.json`,
  `reports/program_validation.json`: FP32 vector KP decoder, N128 MMAP LM,
  top16 correction, pool allocation, tokenizer and original reference checks.

Version-specific facts should be rechecked against these sources after updates.
Diagrams show conceptual tensor flow; the 8×8 tile grid and dot animation are
illustrative, not exact operand address mapping, simulator output or a trace.
No universal accuracy or real-time throughput claim is made.

## Added benchmark evidence

See benchmark-methodology.md for external URLs, versions and metric definitions.
New board evidence: `llmjp_chat_v145/reports/cold_speech/` and
`reports/live_japanese/summary.json`, with original captured audio and replies
under the source `runs/voice_*` folders. Human question-wording confirmation for
the first round came directly from the user in this conversation. Four captures
are not a recognition benchmark population or a comparative hardware study.

## v150 additions — 23 October 2026

- User-supplied Acube `Logo1.2.png`, copied unchanged to `assets/acube-logo.png`.
- AiOnIc chip photograph: https://www.acub3.com/wp-content/uploads/2024/03/AiOnIc-Chichibu-1.png (Our Services, https://www.acub3.com/).
- Acube graphical tools: https://www.acub3.com/?page_id=600 ; screenshot https://www.acub3.com/wp-content/uploads/2023/04/AGTCAiOnIc-ToolsPlugins.png .
- Applications category: https://www.acub3.com/category/technology/ . Individual posts are linked on each application card. Descriptions are summaries of published posts, not new performance or current-deployment claims.
- Whisper: `/home/aionic/work/vlm/whisper_tiny_v149/RESULTS.json`, `README.md`, `live_y02.json`, `instructions_live.json`, `qkv_validation.json`, `strict_greeting_validation.log`, `cross_mask_validation.json`; original-checkpoint live reference and retained WAV. `whisper-v149-results.json` and `whisper-v149-methodology.md` are offline copies.
- v149 question-to-reply completion is 44.483 s after excluding the cue echo from onset timing. The 48.064 s v148 value is a same-recording replay estimate. Both exclude model loading and use a 32-token chat cap; the chat reply did not reach EOS. Recognition is correct on this live question and all nine original Whisper tokens match. This is not a general Japanese word-error-rate result.
- Keep Whisper's learned absolute positions distinct from the RoPE text/VLM decoders. Selected Whisper cached attention remains FP32 CPU math; MMAP cross-attention trials are retained but disabled because tokens changed.