Document streaming vs non-streaming; point to metadata.csv

README.md

@@ -45,12 +45,65 @@ A directory name encodes four dimensions:
 | **Variant** | `diffusiondb_{100,405,2k}` | Sample size drawn from the DiffusionDB dataset. |
 | | `gen1` / `gen2` | Repeated capture batches of the same config. |
 | | `unblocked` | Run without rate-limiting/throttling. |
+| | `streamed` / `stream` | Streaming run. Inconsistently applied — see **Streaming vs non-streaming** below; use `metadata.csv`, not the name. |
 
 **BFCL workloads** come from the Berkeley Function-Calling Leaderboard v4 categories:
 `simple_python` (single function call), `parallel` (multiple calls from one prompt),
 `live_multiple` (real-world prompt, pick among many tools), and
 `live_parallel_multiple` (real-world prompt, multiple parallel calls among many tools).
 
+## Streaming vs non-streaming
+
+**Do not infer this from the directory name.** Use `metadata.csv` at the repo root,
+which carries an explicit `streaming` column for all 61 capture directories:
+
+```python
+import pandas as pd
+meta = pd.read_csv("metadata.csv")
+streamed = meta[meta.streaming == "true"].directory
+```
+
+| `streaming` | Dirs | Meaning |
+|-------------|-----:|---------|
+| `true`    | 33 | Response was delivered incrementally (SSE / chunked token stream). |
+| `false`   | 19 | Whole response body delivered in one shot. |
+| `unknown` |  6 | Provenance lost; the flag used at capture time is not recoverable. |
+| `n/a`     |  3 | Non-LLM baseline (video); API streaming does not apply. |
+
+Every row also carries `streaming_evidence`, naming the harness and code path the
+label came from, so the classification is auditable rather than asserted.
+
+### Why the directory name is not reliable
+
+The name encodes streaming inconsistently, in four spellings and positions
+(`_streamed_`, `_stream_`, `streamingbfcl`, and none at all). Two traps in particular:
+
+- **The 16 `openrouter_*` directories are streaming but carry no stream token.**
+  `openrouter_testing/benchmark.py` sets `stream=True` unconditionally in
+  `run_openrouter_request()`. A name-based filter misses all 16.
+- **`ethernet_baseline_video_stream` is *not* a streaming LLM run.** It is YouTube
+  video streaming — a workload, not a delivery mode. A `grep stream` over directory
+  names picks it up wrongly.
+
+### Measured columns
+
+`measured_median_downlink_gap_ms` and `measured_frac_gaps_over_5ms` are observed from
+the pcaps (median over ~8 mid-size captures per directory): the inter-arrival gap
+between downlink payload packets from the server.
+
+These are **supporting evidence, not the definition of the label**. They separate
+token streaming cleanly (a slow model such as `nemotron_3_super_120b` shows ~420 ms
+gaps; non-streamed bulk delivery shows sub-millisecond gaps), but two confounders make
+them unusable as a standalone classifier:
+
+- **Image generation.** A streamed `text_to_image` run delivers large partial-image
+  chunks at line rate, so it looks like bulk transfer (~0.1 ms gaps) despite streaming.
+- **Link RTT.** Mobile and tethered captures inflate gaps independently of streaming.
+
+The column is blank for 16 directories whose captures have fewer than 10 downlink
+payload packets — too small to measure, which is itself consistent with a single
+non-streamed response body.
+
 ## Datasets
 
 | Directory | Capture env | Model / Provider | Workload |