token_break — LLM streaming traffic captures

Packet captures (*.pcap) and paired client-side streaming logs (*.jsonl) from commercial LLM streaming APIs, collected to study which side channels survive provider-side obfuscation (TLS record padding). Companion to the private analysis repo token_break (see its FINDINGS.md / HANDOFF.md).

No secrets and no prompt text are included. The jsonl logs record only per-chunk metadata (character/token counts and timing), never the token strings.

Files

pcap	paired jsonl	provider / model	notes
orcap_oai4o.pcap	stream_oai4o.jsonl	OpenAI native, gpt-4o	primary — used for the length/count/uplink results
orcap_oai54.pcap	stream_oai54.jsonl	OpenAI Responses, gpt-5.4	primary — reasoning model, timing/TSval
orcap_gpt54.pcap	stream_gpt54.jsonl	OpenRouter → openai/gpt-5.4	exploratory
orcap.pcap, orcap_full.pcap	stream_or_pcap.jsonl, stream_or_pcap_full.jsonl	OpenRouter → openai/gpt-4o	earlier exploration
—	stream_or_gpt4o_n20.jsonl, stream_or_repeat.jsonl	OpenRouter → openai/gpt-4o	earlier runs (client logs only)

uc_prompts.jsonl — the exact prompt set used for every capture above (60 prompts sampled from UltraChat). One JSON object per line: {"prompt": "..."}. This is the file prompt_idx indexes into (0-based line number). You must use this copy — re-sampling UltraChat would misalign prompt_idx with the published pcaps/jsonl. Analysis scripts default to /tmp/uc_prompts.jsonl; pass this file's path instead (e.g. analyze_uplink_prompt.py <pcap> <jsonl> data/uc_prompts.jsonl).

Server→client TLS Application-Data record counts (from the VM): orcap_oai54 ≈ 5300, orcap_oai4o ≈ 3000 packets.

jsonl schema

One JSON object per request:

{
  "prompt_idx": 0, "rep": 0,
  "provider": "openai", "model": "gpt-4o",
  "wall_start": 1788833190.43, "wall_end": 1788833194.36,  // client CLOCK_REALTIME (shared with pcap)
  "n_chunks": 249,
  "total_time_s": 3.90, "ttft_s": 1.55,
  "total_tokens": 249, "total_chars": 1427,
  "chunks": [
    { "t": 0.12, "dt_ms": 8.3, "ntok": 1, "nchar": 4 }  // t = offset from wall_start; nchar = char length (NO text)
  ]
}

prompt_idx = 0-based line number in uc_prompts.jsonl (the prompt that produced this record). rep = repeat index when a prompt was sent more than once.

The client and tcpdump ran on the same VM, so wall_start + chunk.t share one clock with the pcap timestamps — this is what lets you align client tokens to server TLS records.

How to align pcap ↔ jsonl

Extract server→client Application-Data records with tshark:

tshark -r orcap_oai54.pcap \
  -Y "tcp.srcport==443 && tls.record.opaque_type==23" \
  -T fields -e frame.time_epoch -e tls.record.length \
  -e tcp.options.timestamp.tsval

then match record times against wall_start + chunk.t from the jsonl. See the analysis scripts in the code repo (scripts/analyze/).

Known caveats

• Captured on one VM close to the provider → sub-millisecond path jitter; timing results assume this clean path.
• chunks carry nchar only, not token text → structure/boundary fingerprinting needs re-collection with text logged.
• OpenRouter runs proxy to OpenAI models and may differ from native endpoints.