vllm-openttt

TTTPS Proof-of-Time ASGI middleware for vLLM. Attaches a cryptographic audit-trail timestamp (a signed Proof-of-Time receipt) to every non-streaming /v1/chat/completions and /v1/completions response served by vLLM's OpenAI-compatible server, sealed against the public self-serve KPP Provenance API.

This attaches a cryptographic audit-trail timestamp and integrity hash. It does not certify legal or regulatory compliance (EU AI Act, FDA, and so on). Treat it as an audit-trail timestamp, not a compliance claim.

Install

pip install vllm-openttt

vLLM itself is not a dependency of this package. The middleware only needs the ASGI contract, so it installs next to whatever vLLM version you already run.

Usage

Mint a free key with POST https://kpp.kenosian.com/v1/keys, export it, and pass the middleware to vllm serve:

export KPP_API_KEY=...

vllm serve Qwen/Qwen2.5-0.5B-Instruct \
    --middleware vllm_openttt.TTTPSMiddleware

vLLM imports the dotted path and, because it resolves to a class, registers it with app.add_middleware(). Every sealed response then carries two extra headers:

X-TTTPS-Status: ok
X-TTTPS-Receipt: {"status":"ok","backend":"kpp","content_hash":"sha256:...",
                  "receipt_id":"...","receipt":"...","time":"...",
                  "time_source":"...","verify_url":"https://kpp.kenosian.com/v1/verify?receipt_id=...",
                  "overhead_ms":...}

Reading the receipt from a client:

import json, requests

r = requests.post(
    "http://127.0.0.1:8000/v1/chat/completions",
    json={"model": "Qwen/Qwen2.5-0.5B-Instruct",
          "messages": [{"role": "user", "content": "hi"}]},
)
receipt = json.loads(r.headers["X-TTTPS-Receipt"])
print(receipt["receipt_id"], receipt["verify_url"])

A receipt can be re-checked at any time via POST https://kpp.kenosian.com/v1/verify with {"receipt_id": ...}.

Configuration

Self-hosted backend

Setting TTTPS_BACKEND=openttt mints receipts on your own OpenTTT server instead of the public API, which additionally chains each event into that server's hash chain and keeps every anchor inside your own infrastructure.

OpenTelemetry

If a request carries a W3C traceparent header, the trace and span ids are parsed and echoed back in the receipt, so a trace and a receipt point at each other. On the self-hosted backend the ids are also written into the signed Proof-of-Time hash preimage and into the server's queryable event description.

receipt_to_span_attributes() maps a receipt onto span attribute names:

from vllm_openttt import receipt_to_span_attributes

span.set_attributes(receipt_to_span_attributes(receipt))
# {"tttps.receipt_id": "...", "tttps.content_hash": "...", "tttps.trace_id": "...", ...}

Parsing prefers vLLM's own vllm.tracing.extract_trace_context() when running inside a vLLM process, and falls back to a dependency-free W3C parser otherwise. No extra dependency is added either way.

Fail-open

A missing key, an unreachable or slow backend, a malformed traceparent, or any unexpected error degrades the receipt to {"status": "degraded", "reason": ...}. The response body, status code and media type are never altered, and no exception ever escapes the middleware.

Scope

Streaming responses (stream: true, server-sent events) are passed through untouched. Sealing a live event stream needs a different approach and is not attempted here.

License

MIT