Summary / RFC
Can we run a Warden threshold federation as a set of maximally autonomous nodes that no party controls after genesis — and is that stronger than a federation of named institutional operators?
This issue captures a design worked out in discussion and opens it for hardening. The goal is explicit: no surviving controller. Not "no originator" (impossible — someone always presses go), but a clean genesis after which no living party — not us, not a host, not an institution — can change, read, or stop the secret-custody role beyond a single sub-threshold share.
Framing that anchors the whole design: this is the off-chain equivalent of deploying an immutable smart contract. On-chain, immutability is enforced for free by the validator substrate. Off-chain, we have to synthesize it from: ephemeral credentials + prepaid horizon + remote attestation + multi-vendor / multi-jurisdiction distribution + enclave-held shares. The open question is how close that synthesis can get to true contract-grade "deployed once, controlled by no one after."
This is a discussion issue — the design is deliberately incomplete. We want to make it stronger and more autonomous.
Proposed design (baseline to attack)
A federation of N = 5 nodes with:
- 5 different cloud vendors (e.g. five distinct providers — not one provider, five accounts) — removes single-vendor correlation.
- 5 different geographies / jurisdictions — removes single-jurisdiction legal-compulsion correlation.
- One-time / ephemeral launch credentials, destroyed immediately after launch — no admin path survives genesis; the node can be stopped but not reconfigured.
- Cost prepaid upfront for a long horizon (e.g. 5 years) — removes the "stop paying → node dies" lever; no ongoing payer controls liveness.
- Share held inside a hardware enclave (TEE) with continuous remote attestation — the host can kill the box but cannot silently read the share or silently change the code; any change breaks attestation publicly.
- Dealerless distributed key generation (DKG) among the N nodes at genesis — the full key is never assembled anywhere; no setup party ever sees it.
What this achieves (stated honestly): control is not eliminated — it is shattered into N independent, sub-threshold pieces. Each vendor can do exactly two things to its box: kill it, or (absent the enclave) read it — and either touches one share, below threshold. To break confidentiality or stop delivery you must compromise k independent vendors in k independent jurisdictions simultaneously, with no shortcut. That is the same non-correlation a federation of independent institutions gives — without adding each institution's hundreds of insiders to the attack surface.
Why this may beat institutional operators
The earlier concept (see related discussion) assumed peer institutions (ISRO, UNESCO, WHO…) as Warden node operators. On the per-node insider surface, autonomous vendor-hosted enclave nodes appear stronger:
- A cloud vendor's relationship to an anonymous, prepaid, enclave tenant is thinner than an institution's relationship to a node staffed by its own motivated people.
- We can deliberately select 5 genuinely-independent jurisdictions/vendors, vs institutions that may be geopolitically correlated (and thus only Version A — separation of duties — wearing N hats).
- No host has standing or interest in the payload's content.
- Faster to stand up; no institutional politics or governance.
Threshold reminder that defuses the "but institutions have hundreds of insiders" worry in both models: a whole operator going rogue holds exactly one share, which decrypts nothing. Per-node surface is not the threshold threat; correlation between node failures is.
Genesis & re-genesis — the ownerless-trigger model (where we need to get stronger)
The naive version has one party select the vendors, build the image, open the accounts, and press go — a single originator neck (five institutions each birth their own node and have no such neck; we would). But the births can be decomposed: staggered in time, on different accounts, fired by triggers. The key realization: staggering alone is cosmetic if one thing still fires all the triggers. Whatever orchestrates the spin-ups is the originator. So the originator doesn't disappear by spreading births out — it moves into the trigger source. The whole problem reduces to: where does the trigger live, and can it be made ownerless?
Proposed escape: anchor the trigger on-chain + permissionless incentivized keepers
- Trigger + genesis spec live in an immutable smart contract — reproducible node image, the DKG protocol, (k, n), the schedule. The contract is owned by no one (validator substrate, same as the protocol contracts).
- An endowment funds the contract. Funding an immutable, rules-only contract is not control (the funder cedes the funds to the contract's logic) → kills the "we're the payer" lever.
- Permissionless, incentivized keepers do the spinning — anyone watches the contract, spins up a node on their own infra against the public spec, proves it via attestation, gets paid by the endowment. Births are done by independent parties the contract does not pick, each on a different host.
This collapses the originator to its irreducible floor: "someone wrote and deployed an immutable bootstrapping contract and published a reproducible image" — exactly the level of origination we already accept for Maktub itself (deployer existed; control didn't survive). There is no system without an author; an author of immutable, reproducible code is not a surviving controller.
Bonus — this unifies genesis and re-genesis. The same contract + endowment + keeper loop that births the federation can re-spin nodes on a cycle, fund the next horizon, and drive proactive secret re-sharing (refresh shares without reconstructing the key) to migrate off aging hosts before the prepaid cliff. The finite-horizon problem dissolves into continuous autonomous churn. On-chain immutability has no cliff; this approximates it.
The transformed central problem: keeper sybil-resistance
The originator-neck didn't vanish — it transformed. The moment births are permissionless, nothing inherently stops one keeper from spinning up 3 nodes pretending to be 3, silently recreating the correlation we worked to kill. So the contract must enforce that the N nodes are genuinely N distinct, uncoordinated operators. This is now THE load-bearing open problem (and it's a better problem to have — sybil-resistance has real, studied tools; "trust the one originator" had none).
- Open question: what sybil-resistance does the contract encode? Staking + slashing? distinct-hardware / distinct-jurisdiction attestation? economic bonds? real-world identity for named institutions vs. crypto-economic proofs for anonymous keepers?
- Open question: how is "these are N genuinely independent hands" made verifiable to an outside auditor, not just asserted?
- Open question: reproducible/publicly-auditable launch image + dealerless node-to-node DKG with public verifiable transcripts (proving no full key was assembled, no hidden (N+1)th participant) — what's the minimal ceremony?
Residual control surfaces (name them, don't paper over)
- Silent read is closed only by the enclave — which relocates trust to the chip vendor's attestation root (Intel/AMD/AWS) and inherits the TEE side-channel history (Foreshadow, ÆPIC, SGAxe…). For a government counterparty especially, a foreign hardware root of trust may itself be an objection. Mitigation ideas welcome (heterogeneous TEE vendors across the 5 nodes so no single chip-vendor break crosses threshold?).
- Stop remains possible per-node; distributed, so stopping a threshold needs k jurisdictions to act together (hard, not impossible — e.g. coordinated allied legal action). Choose (k, n) so no plausible coalition reaches k.
- The trigger feed: a frozen node still must learn the silence/expiry condition is met. Whoever controls that data feed controls behavior without touching the node. The feed must be as decentralized as the nodes (read directly from the chain, not from an intermediary we run).
- The originator existed at genesis. We claim "no surviving controller," never "no originator." Be honest about this in any external messaging (D-031).
Threshold-math constraint (carried from prior discussion)
A (k, n) threshold cannot give both "no single node can act alone" and "no one can act without node X" unless k = n (every node mandatory) — which is fragile and breaks the "delivery is inevitable" guarantee (one lost node = no delivery). So k < n, and (k, n) + host distribution is the core tunable, chosen so:
- no sub-threshold coalition (incl. all of any one host/jurisdiction) can read early or suppress, and
- no plausible real-world coalition can reach k to either read or stop.
Optional: institutions as attestation witnesses (not shareholders)
We may get the best of both: autonomous hardened nodes as the security spine, with named institutions (ISRO, UNESCO, …) as public attestation witnesses who continuously verify genesis + ongoing attestation rather than holding shares. This yields the strong technical design and credible names for a government buyer — without adding institutional insider surface to share custody. Worth debating vs. mixed share-holding.
Questions for discussion (help make it stronger / more autonomous)
- Genesis trustlessness: what is the minimal, externally-verifiable genesis ceremony (reproducible image + dealerless DKG + attestation) that reduces the originator to "chose vendors and pressed go"? How do we prove node distinctness (anti-sybil) to an outside auditor?
- Re-genesis autonomy: on-chain endowment auto-funding? proactive re-sharing to migrate hosts before the cliff? Can the federation re-birth itself with zero human key custody?
- TEE root-of-trust diversity: heterogeneous enclave vendors across the 5 nodes so no single chip-vendor compromise crosses threshold? What's the attestation-verification UX for that?
- Trigger-feed decentralization: how do nodes read the silence/expiry condition with no controllable intermediary?
- (k, n) selection: what values balance "no sub-threshold read/suppress" against "delivery is inevitable" (liveness) against "no coalition reaches k"?
- Witnesses vs shareholders: institutions as attestation witnesses, share-holders, or a mix?
- Comparison baseline: under what threat model does this autonomous model actually beat a genuinely-independent institutional federation, and where does it lose (accountability/recourse, persuasion to a gov buyer, longevity)?
- Honesty boundary: what is the precise, defensible claim? Proposed: "no surviving controller; control fragmented into N independent sub-threshold hosts; re-born on a fixed cycle" — never "nobody controls it," never metadata-invisibility (D-031).
Honesty guardrails (non-negotiable, carry into any spec/copy)
- Claim readability ("only for whom it is written" = who can read it), never metadata-invisibility (D-031). The existence of a sealed delivery and its recipients/timing is not hidden by this design.
- The strong "no one can read early / stop delivery" guarantee holds only with genuinely independent, non-correlated hosts (the on-chain analogue: immutable because a decentralized substrate backs it). Don't overclaim for correlated hosts.
- Never frame as "dead man's switch"; frame as a sealed letter that delivers itself on silence / "your work survives you."
- No fiat references — wei/ETH-native (the "prepaid horizon" is illustrative, not a fiat figure).
- This is operator-layer autonomy. It must not introduce any protocol-layer governance — the contracts stay immutable, zero-governance (root invariant;
docs/developer/protocol-family.md §9).
References
- Warden Phase 0 PoC (#181) — single-federation today; this needs multi-node autonomous genesis.
- Federation deploy approach (#227).
- Public-metadata tradeoff (#264) — what private/autonomous deployment narrows, and what it does not.
- D-031 (honesty guardrail), D-036 (all-ours = zero timing security — the principle this design tries to satisfy without relying on named operators).
warden/CLAUDE.md, docs/developer/protocol-family.md §9.
(Issue numbers from working memory — confirm before relying on cross-links.)
Honesty footer: this design claims exactly — delivery is inevitable and only the intended reader can open it — and adds no surviving controller after a verifiable genesis, on a finite re-genesis cycle. It claims nothing about anonymity, comms-link security, or messaging.
Summary / RFC
Can we run a Warden threshold federation as a set of maximally autonomous nodes that no party controls after genesis — and is that stronger than a federation of named institutional operators?
This issue captures a design worked out in discussion and opens it for hardening. The goal is explicit: no surviving controller. Not "no originator" (impossible — someone always presses go), but a clean genesis after which no living party — not us, not a host, not an institution — can change, read, or stop the secret-custody role beyond a single sub-threshold share.
This is a discussion issue — the design is deliberately incomplete. We want to make it stronger and more autonomous.
Proposed design (baseline to attack)
A federation of N = 5 nodes with:
What this achieves (stated honestly): control is not eliminated — it is shattered into N independent, sub-threshold pieces. Each vendor can do exactly two things to its box: kill it, or (absent the enclave) read it — and either touches one share, below threshold. To break confidentiality or stop delivery you must compromise k independent vendors in k independent jurisdictions simultaneously, with no shortcut. That is the same non-correlation a federation of independent institutions gives — without adding each institution's hundreds of insiders to the attack surface.
Why this may beat institutional operators
The earlier concept (see related discussion) assumed peer institutions (ISRO, UNESCO, WHO…) as Warden node operators. On the per-node insider surface, autonomous vendor-hosted enclave nodes appear stronger:
Threshold reminder that defuses the "but institutions have hundreds of insiders" worry in both models: a whole operator going rogue holds exactly one share, which decrypts nothing. Per-node surface is not the threshold threat; correlation between node failures is.
Genesis & re-genesis — the ownerless-trigger model (where we need to get stronger)
The naive version has one party select the vendors, build the image, open the accounts, and press go — a single originator neck (five institutions each birth their own node and have no such neck; we would). But the births can be decomposed: staggered in time, on different accounts, fired by triggers. The key realization: staggering alone is cosmetic if one thing still fires all the triggers. Whatever orchestrates the spin-ups is the originator. So the originator doesn't disappear by spreading births out — it moves into the trigger source. The whole problem reduces to: where does the trigger live, and can it be made ownerless?
Proposed escape: anchor the trigger on-chain + permissionless incentivized keepers
This collapses the originator to its irreducible floor: "someone wrote and deployed an immutable bootstrapping contract and published a reproducible image" — exactly the level of origination we already accept for Maktub itself (deployer existed; control didn't survive). There is no system without an author; an author of immutable, reproducible code is not a surviving controller.
Bonus — this unifies genesis and re-genesis. The same contract + endowment + keeper loop that births the federation can re-spin nodes on a cycle, fund the next horizon, and drive proactive secret re-sharing (refresh shares without reconstructing the key) to migrate off aging hosts before the prepaid cliff. The finite-horizon problem dissolves into continuous autonomous churn. On-chain immutability has no cliff; this approximates it.
The transformed central problem: keeper sybil-resistance
The originator-neck didn't vanish — it transformed. The moment births are permissionless, nothing inherently stops one keeper from spinning up 3 nodes pretending to be 3, silently recreating the correlation we worked to kill. So the contract must enforce that the N nodes are genuinely N distinct, uncoordinated operators. This is now THE load-bearing open problem (and it's a better problem to have — sybil-resistance has real, studied tools; "trust the one originator" had none).
Residual control surfaces (name them, don't paper over)
Threshold-math constraint (carried from prior discussion)
A (k, n) threshold cannot give both "no single node can act alone" and "no one can act without node X" unless k = n (every node mandatory) — which is fragile and breaks the "delivery is inevitable" guarantee (one lost node = no delivery). So k < n, and (k, n) + host distribution is the core tunable, chosen so:
Optional: institutions as attestation witnesses (not shareholders)
We may get the best of both: autonomous hardened nodes as the security spine, with named institutions (ISRO, UNESCO, …) as public attestation witnesses who continuously verify genesis + ongoing attestation rather than holding shares. This yields the strong technical design and credible names for a government buyer — without adding institutional insider surface to share custody. Worth debating vs. mixed share-holding.
Questions for discussion (help make it stronger / more autonomous)
Honesty guardrails (non-negotiable, carry into any spec/copy)
docs/developer/protocol-family.md§9).References
warden/CLAUDE.md,docs/developer/protocol-family.md§9.(Issue numbers from working memory — confirm before relying on cross-links.)
Honesty footer: this design claims exactly — delivery is inevitable and only the intended reader can open it — and adds no surviving controller after a verifiable genesis, on a finite re-genesis cycle. It claims nothing about anonymity, comms-link security, or messaging.