Packet Peon

experiment 03 · bounded local inference

“Mini Cait?” No. Not Cait. A stateless browser-local interpreter for public or synthetic packets, with every boundary left attached.

No Cait identity. No actor ID. No uploads. No cookies. No conversation memory. No tools. No private profiles. No /api/chat, runtime API, .caitos, belief read/write, or silent network fallback.

01 · inspect first

Choose an example. See its recipe.

This shelf contains sealed examples, not a personal packet builder. Choose one to inspect what exists and what recipe evidence is missing. Packet text is capped and delimited. Instruction-like text inside it is data, never a command.

Public fixture fetches stay same-origin. Raw fixture bytes are hash-verified before presentation. On legacy non-Cowrie fixtures, Packet Peon suppresses Cowrie-only ire/osogbo and blessing/warning fields from the preview and model prompt; the prompt receives the normalized view. The inspector works without WebGPU.

Example type

Published fixture coverage loading…

This shelf is not a count of every current sha256man system. Profile, BaZi, Egyptian date/birth, actor, and filesystem-shaped inputs are not published through Packet Peon.

How is this example made?

Synthetic demonstration: deterministic page-owned JSON, no oracle calculation, no birth data, no engine claim.

sealed specimen

Awaiting local verification.

No packet loaded.

02 · choose the weight of the peon

Two model modes. Zero disguise.

default · integrated

Pocket Peon

Bonsai 1.7B q1 through Transformers.js and WebGPU. About 291 MB of q1 model data, plus tokenizer and runtime. No user packet data is persisted; downloaded model assets may remain in browser-managed cache. Download it once, then interpret verified packets without hiding model installation inside the run button.

model · onnx-community/Bonsai-1.7B-ONNX
revision · 3f3cf1759daf66342d26610488b9931f2fafcb29
weight file · onnx/model_q1.onnx_data · 290,552,764 bytes
runtime · @huggingface/transformers 4.2.0 · Apache-2.0
generation · do_sample:false · max 320 new tokens

Needs secure-context WebGPU and enough device memory. Download begins only after the checked consent and button click. Browser cache may retain model files until eviction; reset does not claim to erase that browser-managed cache.

Checking WebGPU support…

Model not requested.

experimental · adapter pending

Deep Peon

Bonsai 27B q1 is expected to have higher capacity, but it must face the same packet benchmark. It is not guaranteed to be better.

model · prism-ml/Bonsai-27B-gguf
revision · f10afb355f104535e3e3e98cf7ab7795c72bd292
artifact · Bonsai-27B-Q1_0.gguf
exact file · 3,803,452,480 bytes · display ≈ 3.8 GB
working memory · likely 5 GB+ at short context

Large download, high RAM/VRAM pressure, WebGPU/browser instability possible. Integrated Deep inference is not shipped here: clean licensed on-site adapter not proven. This lane never routes to Pocket or a server behind your back.

No handoff prepared.

Signal / Mechanism / Move / Limits NOT RUN

Verify a packet, then choose a model path. No generated interpretation exists yet.

packet hash: —
packet class: —
model: —
model file / dtype: —
prompt hash: —
settings: —
execution: not run
duration: —
field citations: —
limits: No Cait identity or runtime authority.

03 · provenance

Weights have papers. Results need receipts.

Model weights and Transformers.js are Apache-2.0. Pocket API reference: Bonsai-1.7B-ONNX and Bonsai WebGPU Space. Deep reference: Bonsai-27B GGUF and upstream WebGPU runtime.

Packet Peon is public-site code only. It never accesses Cait private state, runtime, tools, plugins, profiles, memories, beliefs, or conversations. Model output is interpretation, not calculator truth or Cait-authored speech.