satsignal-otel 0.1.0

OpenTelemetry SpanProcessor that anchors selected GenAI spans to BSV via Satsignal.

satsignal-otel

OpenTelemetry SpanProcessor that anchors selected GenAI spans to BSV via Satsignal. One integration covers any observability stack already speaking OTel — Langfuse, LangSmith, Arize, Datadog, Honeycomb — and adds a tamper-evident receipt for the spans that matter.

Your observability stack shows the run. Satsignal proves the run record hasn't been edited since.

pip install satsignal-otel

What it does

Drop the processor into your TracerProvider. Only spans carrying the attribute satsignal.anchor=true are anchored — everything else flows through to your existing exporters untouched. Matching spans are batched and posted as a single manifest-mode anchor on BSV, binding all leaves under one Merkle root. The on-chain anchor is the receipt your auditor uses to prove the span record has not been edited since that block was mined.

By default the SDK is opt-in per span, batched (1-minute window), sha-only (your span bytes never leave the process), and fail-open (anchor failures log + drop; your app keeps running).

Failed-eval auto-anchor (headline)

Wire one line into your eval pipeline. When a scorer drops below threshold, mark the span — Satsignal anchors the failure with a per- span receipt so the timing claim ("we knew at 14:32 UTC") is provable.

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from satsignal_otel import SatsignalSpanProcessor, auto_anchor_on_eval_fail

provider = TracerProvider()
provider.add_span_processor(SatsignalSpanProcessor(
    api_key=os.environ["SATSIGNAL_API_KEY"],
    matter_slug="otel-evals",
))
trace.set_tracer_provider(provider)

tracer = trace.get_tracer(__name__)

with tracer.start_as_current_span("eval.scorer") as span:
    score = run_scorer(prompt, response)
    span.set_attribute("gen_ai.eval.score", score)
    auto_anchor_on_eval_fail(span, threshold=0.7)

The helper sets satsignal.anchor=true only when score < threshold; under threshold spans are anchored individually (mode=single) so the receipt timing is per-event, not amortized across a batch.

Release-gate anchor (secondary)

At deploy time, mark the release-manifest span. One receipt per release — what shipped, with the eval evidence around it.

from satsignal_otel import mark_for_anchor

with tracer.start_as_current_span("release.gate") as span:
    span.set_attribute("gen_ai.system", "anthropic")
    span.set_attribute("gen_ai.model", "claude-opus-4-7")
    span.set_attribute("prompt.version", PROMPT_VERSION)
    span.set_attribute("eval.pass_rate", PASS_RATE)
    span.set_attribute("config.hash", CONFIG_HASH)
    mark_for_anchor(span, mode="single", label=f"release-{GIT_SHA}")

Configuration

SatsignalSpanProcessor(
    api_key,                    # required: SATSIGNAL_API_KEY
    matter_slug,                # required: workspace matter for receipts
    base_url="https://app.satsignal.cloud",
    flush_interval=60.0,        # seconds between manifest flushes
    max_batch_size=500,         # force-flush when queue hits this size
    daily_anchor_cap=1000,      # client-side guard against runaway anchoring
    fail_open=True,             # log + drop on anchor failure
    transport=None,             # inject a callable for tests (see below)
)

Attributes the processor reads

Attribute	Type	Effect
satsignal.anchor	bool	Required to anchor. True → span is queued.
satsignal.anchor.mode	str	"manifest" (default, batched) or "single" (per-span anchor).
satsignal.anchor.label	str	Optional display label on the receipt (truncated at 256 chars).
satsignal.anchor.force_new	bool	True bypasses server-side dedup (single mode only).

What gets anchored

The sha256 of a deterministic canonical-JSON encoding of the span's {name, kind, status, start_time, end_time, attributes, events, links, resource}. The bytes never leave your process — only the hash. The OTel trace_id and span_id ride along as an off-chain session_id so a verifier can correlate the receipt back to your existing trace data.

This is chain-of-custody for the span record, not the underlying prompt + completion bytes. Your trace store still owns the bytes; the anchor proves they have not been edited since.

Threat model

• Span attributes are attacker-controllable when prompts include user-provided strings. The label is treated as untrusted display text server-side (length-capped, harness-string-rejected). The sha256 is bytes-are-bytes.
• An anchor proves anchorer-knowledge, not world-existence. The on-chain receipt commits the anchorer's knowledge of the canonical bytes at the anchored time; it does NOT prove the span existed before then. For end-to-end provenance pair this with a commit- reveal flow over the underlying prompt + completion.
• Semantic-convention churn. The GenAI OTel semantic conventions are still in development. auto_anchor_on_eval_fail reads gen_ai.eval.score, gen_ai.evaluation.score, and eval.score — pass score_attribute= if your stack names it something else, or pass score= directly.

Sister packages

• satsignal-cli — verify receipts and bundles on disk.
• satsignal-mcp — agent-callable anchoring + lookup tools via Model Context Protocol.
• langchain-satsignal — LangChain callback that anchors policy snapshots + decisions + evidence manifests.
• Steleet/satsignal-action — GitHub Action that anchors build artifacts as a workflow step.

Testing without the network

The processor accepts a transport= callable matching urllib's shape:

def fake(method, url, headers, body, timeout):
    return 200, b'{"bundle_id": "abc", "txid": "deadbeef", ...}'

processor = SatsignalSpanProcessor(
    api_key="sk_test", matter_slug="m", transport=fake,
)

The full test suite exercises the processor against a mock transport; see tests/test_processor.py for canned-response patterns.

License

MIT.