DATASHEET.md

# A2A-MetaTrace datasheet

A labeled, metadata-only corpus of agent-to-agent (A2A) workflow traces, captured
from official A2A sample agents composed into multi-agent workflows. Released as a
standalone artifact alongside the paper; this datasheet follows the Gebru et al.
"Datasheets for Datasets" structure.

## Motivation
The corpus evaluates metadata leakage on workloads that are not the authors' own
generator: the protocol path (Agent Cards,
discovery, `message/send`, SSE) and the agents are *officially authored*; only the
workflow composition and the class labels are ours.

## Composition
- **Instances.** One trace = one workflow execution. The corpus is **270
  traces**: 30 runs each of **9 workflow classes** over **6 capabilities**, each class
  realized by 3 compositional **variants** (see `corpus/classes.py`), runs split across a
  class's variants.
- **Per message (obs(m)).** `(src, dst, t, length, direction)` plus ground truth:
  `stage_idx`, `step_type`, `capability`, `task_class`, `n_stages`, and a `variant` tag
  for the leave-variant-out split. **Payloads are discarded**: the corpus is metadata
  only (the capture records lengths, not bodies), so it is exactly a passive TLS
  observer's view and carries no message content.
- **Labels.** `task_class` is the workflow class (the ground-truth the adversary
  recovers); `variant` is the specific composition (the generalization group). Labels
  and composition are ours; the protocol path is the real `a2a-sdk`.

## Collection process
Each capability is an official `a2a-samples` agent launched as a standalone A2A server in
its own env (`corpus/sample_agents.py`, `corpus/launcher.py`); an orchestrator composes
them per (class, variant) and a real stage runner (`corpus/runner_samples.py`) records
obs(m) via a real discovery round-trip + `message/send` + SSE streaming, the same A2A
wire the measured binding observes. The runner overrides each resolved Agent Card's
interface URL to the real launch port (some samples ship a stale hardcoded card URL).
Transport: HTTPS-direct (the protection ladder is applied analytically in the
accompanying analysis).

## Honest disclosures / limitations
- **Provenance.** The corpus is captured from **official `a2a-samples` agents**,
  each run unmodified as its own server: `travel_planner` (OpenAI), `adk_currency_agent`,
  `content_planner`, `adk_skills_agent` (Gemini via Vertex), `adk_expense_reimbursement`
  (LiteLLM→OpenAI), and `helloworld` (no-LLM echo). The protocol path (Agent Cards,
  discovery, `message/send`, SSE streaming) and the agent logic are official; **only the
  workflow composition and the class labels are ours.**
- **Model substitutions (disclosed; agent logic unchanged).** Three bindings differ from
  the samples' shipped defaults, with no behavioral edits:
  - `adk_currency_agent`, `adk_skills_agent`: `model=` changed `gemini-3-flash-preview`
    → `gemini-2.5-flash`.
  - `adk_expense_reimbursement`: `LITELLM_MODEL=openai/gpt-4o-mini` (env only).
  - `a2a_telemetry` is **excluded**: the sample hardcodes an ADK `google_search` tool with
    a LiteLLM model, which the current ADK rejects as-shipped. These edits are not committed
    into the `a2a-samples` clone; they are recorded here for reproducibility.
- **Observation model.** Records are per A2A application event. Because a network observer
  sees TLS-record bursts, not application deltas, the accompanying analysis also reports a
  **wire-faithful** view that aggregates each stage's streamed response into one
  observation; recovery is robust to this (see below).
- **Uniformity.** The sample agents are more uniform than a production deployment; a
  generalization limit, reported as such (leave-variant-out below).

## Recommended splits
Random k-fold; **leave-variant-out** (generalization: hold out whole compositions, by
grouping on `variant` so no composition spans the train/test boundary). Report both.

## Known result: real, artifact-robust, composition-specific, mitigable
On this corpus (270 traces, 9 classes, chance 0.111, 8 seeds; reproduced by the
accompanying analysis):
- **Random k-fold = 0.668 ± 0.02 (6.0× chance)**; **leave-variant-out = 0.18 (~chance)**.
  The label-blind adversary recognizes *specific, previously seen compositions*, not the
  abstract task intent, so a held-out composition of a known class is not recovered.
- **Not a streaming artifact.** Recovery is unchanged under the wire-faithful aggregation
  (per-delta 0.676 → aggregated 0.668; mean msgs/wf 376 → 11), so it rests on real workflow
  *structure*, not on the chattiest agent's per-delta volume.
- **Mitigable.** Under a metadata-protecting transport (wire-faithful, adversary vocabulary
  fixed from the undefended deployment): none 6.0× → metadata-min shim 4.2× → cover 0.26×
  → cover+unlinkability **1.00× = exactly chance**.

The corpus thus confirms the leakage claim on real official-SDK agent traffic, scopes it
(recurring/profiled workflows, not zero-shot novel compositions), and shows the defense
closes it.

## Distribution
Dataset (Parquet) + this datasheet + the capture/generation harness (`corpus/`). Load
with `datasets` / `pandas` (see `README.md`).