Datasets:

wesupport-ai
/

aics-protocol-examples

license: cc-by-4.0
language:
  - en
task_categories:
  - text-generation
tags:
  - ai-agents
  - tool-use
  - synthetic
  - cryptographic-verification
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: cryptographic_verification
        path: data/cryptographic_verification-*.jsonl
      - split: onboarding_gate
        path: data/onboarding_gate-*.jsonl
      - split: swarms_discovery
        path: data/swarms_discovery-*.jsonl
      - split: sandbox_troubleshooting
        path: data/sandbox_troubleshooting-*.jsonl

wesupport-ai/aics-protocol-examples

50 deterministic multi-turn examples in each of four scenario splits: onboarding_gate, swarms_discovery, cryptographic_verification, and sandbox_troubleshooting. These are scenario splits, not independent train/test partitions. Templates are shared and should not be used to claim held-out accuracy.

All network and shell outputs are simulated. No production service was contacted. Every signature, receipt, checkpoint, Merkle proof and sealed-box test vector is computed using aics_identity, aics_tlog and the server's entry builders. Verification does not establish live service acceptance, DNS trust, witness quorum or freshness. The restricted sandbox scenario defers local verification; the dataset builder still verifies its public read artifacts offline before writing them.

Schema

messages: chat objects with role, content, assistant tool_calls (id, type, function.name, JSON-string function.arguments) and paired tool outputs with tool_call_id and name. tools: OpenAI-compatible function JSON schemas. split: scenario family. scenario_id: stable family and ordinal. verification_artifacts: fixture flag, origin, log_kid, C2SP log_vkey, current and previous checkpoint notes, consistency proof, entries with canonical entry bytes, bodies, receipts and inclusion proofs, session-bound gate signatures, and sealed-box vectors. U64 fields are unpadded base64url; proof arrays and gate signature_b64 use standard base64. Timestamps are fixed historical UTC seconds.

Fixture signing keys and deterministic ephemeral keys derive from public test material; encryption recipient seeds are deliberately included for repeatable decryption tests. They protect no secrets and must never be used for live identities. Sealed vectors use crypto_box_seal framing and its XSalsa20-Poly1305 construction, and are opened by the real aics_identity.sealed implementation.

Verification

From the AICS repository, using its aics-server uv environment:

uv run --project packages/aics-server python3 tools/export_hf_datasets.py --verify PATH/aics-protocol-examples
uv run --project packages/aics-server pytest tools/test_export_hf_datasets.py

Library: verify_protocol_example(json.loads(line)) returns True on success and raises ValueError for malformed or invalid artifacts. It checks every entry signature at submission time, per-key chain, nonce uniqueness, receipt binding, checkpoint signature, inclusion and consistency proof, gate command/signature, body commitment, and sealed-box header/ciphertext/plaintext binding.

Generated code baseline: abb9396148cf02ac31a7923694d862940dbf2d02; source SHA-256: cf62840fa2121bac811aa46ad37b7924a62a1e2dc199e15bfe5a402f3f948008. Licensed under CC BY 4.0.

Dataset card corrected on 19 October 2026 to list the cryptographic verification split first, so Hugging Face infers the populated sealed-box schema before reading splits with empty sealed-box lists. JSONL payloads and their provenance are unchanged.