Skip to content

docs: ADR 0002 (accepted) — re-baseline L1 against a state-scale censor - #5

Open
croc100 wants to merge 2 commits into
mainfrom
docs/adr-0002-state-scale-rebaseline
Open

croc100 wants to merge 2 commits into
mainfrom
docs/adr-0002-state-scale-rebaseline

Conversation

@croc100

@croc100 croc100 commented Oct 6, 2026 •

Copy link
Copy Markdown
Owner

Docs only. Adds ADR 0002: re-baseline L1 against a state-scale censor, status Accepted (2026-10-07) with stricter terms. DESIGN, ROADMAP, README and CHANGELOG are rewritten to match.

Why

Slice 1's gate asks "can a prober tell this relay from the site it borrows?" After H3–H8 the answer is no. A national censor asks a different question: "is this flow plausible, given DNS, IP ownership, address-space type and every other flow at this address?"

The default posture fails that question before Warren's protocol matters:

  • A residential relay presents a name it doesn't own. DESIGN §14 already rejects SNI spoofing for exactly this reason.
  • It acts as an HTTPS server for many strangers.

The gate can't see either, because it only compares a relay with its own cover.

What's in it

  • Adversary model by capability (C1–C6):

    • line-rate per-flow features
    • cross-source joins
    • induced faults
    • unlimited clients
    • policy-scale action
    • coercion

    It argues for TPR at ~10⁻⁴ FPR, not AUC.

  • Ten findings, ranked by cost to the censor:

    • name/address ownership
    • residential fan-in
    • server TCP/IP stack and port surface
    • tunnelled-TLS shape (Xue et al. 2024)
    • flow shape
    • fault reactions, including NTP skew as a censor-operated off switch for the tag window
    • the relay's own traffic to its cover
    • in-country view of the cover
    • "censor runs a client" no longer acceptable
    • the gate checks self-consistency, not survival
  • Six proposed decisions:

    • D1 stop borrowing names: self-hosted cover / shared front; tls-borrow only when co-located in a datacenter
    • D2 residential relays leave the server role (WebRTC, outbound-only hops)
    • D3 handshake-phase shaping on by default
    • D4 flow budgets
    • D5 specified fault behaviour, ticket-based freshness instead of absolute time, a credential slot in the tag
    • D6 five new gates (join test, classifier, stack parity, fault response, outside eyes)
  • A table of what happens to current code. Nothing is deleted: the borrow/warm-pool/H8 machinery is demoted to the co-located variant.

Other changes

  • DESIGN is rewritten to match:
    • §6.3: new transportsself-cover, front, outbound-hop; tls-borrow restricted to co-location
    • §6.4: a minimal shaping floor and flow budgets
    • §6.5: the ladder re-ordered
    • §7.1: rewritten around household-shaped relays
    • §13: new threat rows
    • §14: six superseded decisions
    • §16: Slice 1 reopened
  • ROADMAP: Slice 1 is reopened, with gates G1–G5 ahead of mechanisms M1–M9. The earlier work becomes Slice 1a, and Slices 2 and 3 inherit the gates.
  • README: the security table is prefaced with what it doesn't measure.
  • CHANGELOG: a new entry.

See the ADR's Accepted with these terms section for the tightening. No code changes.

🤖 Generated with Claude Code

croc100 and others added 2 commits October 7, 2026 00:44
Slice 1's gate asks whether a prober can tell a relay from its borrowed site.
A national censor asks whether a flow is plausible given DNS, IP ownership,
address-space type and the other flows at the same address, and the default
posture fails that before the protocol matters: a residential relay presents a
name it does not own and serves long sessions to strangers.

The ADR models the censor by capability (C1-C6), ranks ten distinguishers by
what they cost to exploit, and proposes six decisions: stop borrowing names,
take residential relays out of the server role, shape the handshake phase by
default, budget flows like visits, specify fault behaviour and drop absolute-
time authentication, and replace the gate with five measurements (join test,
classifier at 1e-4 FPR, stack parity, fault response, outside eyes).

ROADMAP and DESIGN point at it; Slice 1 reopens if it is accepted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…atch

The project accepted ADR 0002 and asked for it to be applied harder: no waivers
for failing gates, a canary per finding, bucket shaping by default for
interactive traffic, a mandatory client credential, continuously running gates,
and the residential server role removed rather than deprecated.

DESIGN: new transports (self-cover, front, outbound-hop) with tls-borrow
restricted to co-location; minimal shaping floor and flow budgets; a re-ordered
ladder; §7.1 rewritten around household-shaped relays; new threat rows; six
superseded decisions; Slice 1 reopened.

ROADMAP: Slice 1 reopened with gates G1-G5 before mechanisms M1-M9; the earlier
work kept as Slice 1a; Slices 2 and 3 inherit the gates. README and CHANGELOG
updated. No code changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@croc100 croc100 changed the title docs(adr): 0002, re-baseline L1 against a state-scale censor (proposed) docs: ADR 0002 (accepted) — re-baseline L1 against a state-scale censor Oct 6, 2026
@croc100

croc100 commented Oct 6, 2026

Copy link
Copy Markdown
Owner Author

Accepted with stricter terms (ADR 0002 §Accepted with these terms). DESIGN, ROADMAP, README and CHANGELOG are now rewritten to match, not just pointed at the ADR. No code changes. Expect a trivial conflict at the top of docs/CHANGELOG.md with #3/#4, whichever merges second.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant