02 — Capture
No mode depends on another. Modes A, B, and C flow via the same signed protocol into multiple independent mirrors — none of them require a return channel to the end user. Mode D stands separately: it evaluates exclusively public registry data and runs mirror-internally.
Data flow
Modes A, B, and C initiate every connection themselves — egress-only. Mode D runs completely mirror-internally on public sources.
Dashed nodes are planned, not in operation. In Phase 1, exactly one bootstrap mirror runs.
The four modes
An SDK in the OSS library itself reports runtime usage. Exactly four
event types: install, session-start,
runtime-tick every five minutes, and feature-flag.
The library decides whether to include it — no one else.
Public manifests from npm, PyPI, Maven Central, crates.io, and Go Modules, plus a GitHub manifest crawl. Runs mirror-internally without signature roundtrip — the PII check still applies as defense-in-depth.
An agent in the host software reports which embedded libraries are actually running. Since this measures closer to the end user, a significantly stricter k-minimum applies.
A local aggregator bundles multiple sources before anything leaves the machine. This increases cohort size and further reduces the re-identification risk.
Common denominator
Whatever these three modes capture — it only leaves the machine in one form: canonically serialized, signed, and only after both PII gates. Mode D works differently: The scraper operates mirror-internally on purely public sources, without signature roundtrip and without k-anon guard — which would be meaningless there because a public registry has no cohort. The schema-reject-PII gate still applies for D, as defense-in-depth.
RFC 8785 (JCS) — the same input produces byte-identical output in every language. Without this, no signature would be cross-language verifiable.
Signing is done directly over the canonical bytes, without prehash. The
submission_hash is separate and serves as an idempotency key.
Always a Package URL — pkg:npm/…, pkg:pypi/…,
pkg:cargo/…. No free text, no name ambiguities.
Modes A, B, and C initiate every connection themselves. There is no return channel through which a mirror could address an end device.