Port experiments

A test bench for group authority, key agreement, replication and delivery. The same scenarios run against several candidate stacks; from what each candidate needs from the app we derive the ports of the Real Life Trust Protocol. The findings per candidate live here in src/candidates/*/NOTES.md; the plan and its reasoning in the private workshop repository (rltp/design/port-experimente-2026-10.md).

Results

• The conformance matrix, interactive: 53 promises across four ports (group candidates in the authority, key and replication ports, multi-device candidates in the multi-device port), every cell with its scenarios and the recorded result: https://real-life.org/port-experimente/matrix.html. The page and the Markdown matrix (matrix.md) are both written by REPORT=1 pnpm vitest run report (src/lab/matrix-page.ts; the English text of ports, promises, scenarios and candidates in src/lab/matrix-en.ts, a test fails when one is missing); never edit public/matrix.html by hand. The browser page renders it at the end of a run.
• How to read it. A promise is kept only if every scenario mapped to it passes. "Not applicable" means the candidate has no capability for the scenario, or the carrier fault never touched a frame of that kind. That is a gap in coverage, never a pass.
• Every branch. A cell is kept only if it holds in every branch. Where a candidate decides by chance (two concurrent MLS commits: the smaller hash wins), the bench forces each side of every such decision and runs the scenario in all branches (src/lab/zweige.ts). The enumeration covers only that sibling-commit tiebreak (force_tiebreak in rust/mls-wasm); other random-dependent orders, such as the topological sort over nonce-dependent operation ids in src/candidates/rltp-beekem/authority.ts, are not enumerated (no changing outcome observed in 2 × 132 cells). Where a result depends on delivery time and the stack gives no delivery bound (the Subduction server in keyhive-ark, S6 and S6b), the cell is not kept, always: it reads "abhängig von: Zustellzeit …". The bench still runs the scenario several times, checks every run strictly (no exception, the expected membership) and reports the distribution within an observation window in the details (src/lab/zeit.ts); the window is the bench's waiting time, not a bound. A cell that holds only sometimes is not kept, and its result says what it depends on ("abhängig von …"). There is no "nondeterministic" marker and no oracle is widened to make a run pass.
• The bench itself speaks German: scenario titles, status lines and the recorded results are German strings in the code. The notes and this README are English.

Candidates

Id	What	Transport class
gen2	WoT generation 2 (@web_of_trust/core 0.6.0, Yjs, own relay in process)	own server
p2panda	p2panda-spaces 0.7.1 (WebAssembly)	bench relay
keyhive	Keyhive 0.3.0-alpha.1 (WebAssembly, content in Yjs)	bench relay
keyhive-ark	Keyhive over ARK 0.6.0-alpha.1 (Automerge, Subduction server)	own server, Node only
rltp-beekem	authority log by the conflict matrix, keys from BeeKEM 0.4.0	bench relay
rltp-beekem-sichten, -logreplik	the same with an enforcing service at the relay (class view or class log)	bench relay, enforcing
rltp-beekem-pws	the same on Portable Web Spaces (was-teaching-server 0.42) as blind storage	own server, Node only
rltp-mls, rltp-mls-sichten	the same candidate code, keys from MLS (OpenMLS 0.9.1)	bench relay
consensual-mls-any-staged, -owner	MLS without an authority log, the commit decides membership (room model)	bench relay
klartext	null candidate without encryption; shows that the bench detects faults	bench relay

Run times are comparable only within a transport class: with an own server, S9 is almost all waiting.

• mehrgeraet-k0: E15, the multi-device holder state. K0 is today's cut, one seed and no merge, the baseline that has to fail X1, X2, X3 and X7. Further candidates (journal with pre-assigned rights, device sets with a delegation profile, soft limits) follow.

Scenarios

• S1 to S11 (src/lab/scenarios.ts): membership under conflict, devices per person, history, stolen state, a service at the relay.
• T1 to T12 (src/lab/traeger.ts): the carrier as adversary (drop, reorder, duplicate, tamper, withhold, freeze, carrier switch, what the carrier reads).
• T9, T10, Z1 to Z5 (src/lab/zustellung.ts): pair delivery over the wires plain envelope, TSP Rev 3 and DIDComm v2 (authcrypt, anoncrypt): lost receipt, device offline, what the carrier learns, carrier switch, multiple devices, foreign sender, wrong recipient key. E11b step 2 adds the direct send forms (b) tsp-direct-connection and (a) tsp-direct-egress-session, with sessions per relationship under a carrier that forwards only senders holding a session. E11b step 3 adds the routed form (c) tsp-routed (and tsp-routed-egress-session) with an intermediary that opens the routing layer and forwards the inner frame, under foreign senders, duplicates, tampering and rebinding.
• S9, S9b (src/lab/load.ts): load, 30 members.
• X1 to X12 (src/lab/mehrgeraet/scenarios.ts): several devices of one person against the target model of Trust 0.32 §8 and §6a: double consumption, budget, consent against decline, revocation against send, departure against replay, restore from backup, continuity across devices, enrollment, device revocation, long separation, delivery per device. Each runs with two devices and serially as the control.

Layout

• src/lab/types.ts: the driver interface of the bench. This is not the port we are looking for, only as much as the scenarios need.
• src/lab/world.ts: simulated world with one Y.Doc per device and a store-and-forward network with partitions and fault injection.
• src/lab/matrix.ts: promises per port with the mapping scenario to promise; generates the matrix.
• src/candidates/<id>/: one candidate per directory, with NOTES.md. rltp-beekem is the experiment on the port cut: the authority log (authority.ts) decides membership, the key port (keyport.ts) delivers the keys, BeeKEM or MLS; the service at the relay is in service.ts.
• src/candidates/zustellung/: the wires of the delivery bench.
• src/lab/mehrgeraet/: the multi-device bench (E15): target model, holder device, world, scenarios.
• rust/: WebAssembly wrappers (p2panda-wasm, beekem-wasm, mls-wasm), built by scripts/build-wasm.sh; the bindings land under src/candidates/*/pkg/. rust/mls-wasm/examples/vectors.rs exchanges KeyPackage, Welcome, Commit and message as raw bytes with other MLS implementations (cargo run --example vectors).
• Candidates marked nodeOnly run only in Node and in CI; the browser page states the reason.

Commands

pnpm install
pnpm wasm                       # build the Rust wrappers (Rust 1.96, wasm-bindgen 0.2.129)
pnpm server:fetch               # Subduction server (Rust binary) for Keyhive over ARK
pnpm test                       # bench tests
pnpm test:rust                  # native tests of the Rust wrappers
REPORT=1 pnpm vitest run report # result table of all candidates and matrix.md
pnpm dev                        # browser run, local

Live probe against the VTI mediator (E11b)

tests/zustellung-live.test.ts measures Rev 3 routed TSP against a live mediator. It needs the network, is skipped by default and never enters the report. Results: src/candidates/zustellung/NOTES.md, section "E11b live".

RLTP_LIVE_MEDIATOR=1 pnpm vitest run tests/zustellung-live.test.ts

Variable	Meaning	Default
RLTP_LIVE_MEDIATOR	1 runs the probe	off
MEDIATOR_DID	another mediator instance	instance of 27 September 2026
MEDIATOR_ALLOW_INSECURE	1 sets net policy allowInsecure and allowPrivate (local mediator)	off
RLTP_WIRE_DEBUG	1 prints row outcomes and error classes on stderr	off
E11B_FORMS	comma list of forms (D-c,D-es,D-e,R-cc,R-c,R-c',R-es,R-ce)	all
E11B_RUNS	runs per form pass (integer ≥ 1)	3
E11B_TIMEOUT_MS	delivery window and ack deadline	20000
E11B_PROTOCOL_OUT	path of the protocol JSON	.cache/e11b-live/<UTC-ISO>.json

The post-observation per form pass is the constant OBSERVE_MS (30 s), not a variable.

The protocol holds no keys, seeds, tokens, payloads or error texts: only shortened DIDs, digest prefixes, counts, hosts and values from fixed lists.

The browser page is published via GitHub Pages on every push to main. There, all runs can be started on a real device (S9, for example on GrapheneOS) and the result copied as JSON.

License

MIT. The vendored p2panda crates under rust/p2panda-wasm/vendor/ keep their own license, MIT OR Apache-2.0; the texts are in rust/p2panda-wasm/vendor/LICENSES/.