#Aims and reasoning
#Authority should survive a service outage
Participants need useful state and agency when a service disappears or a network partitions. Moving packets between peers solves only part of that problem. A system must also decide who may admit a device, disclose an object or change shared state, and what observers can learn from identifiers and proofs.
The September Matrix study distinguishes several ways to move responsibility toward endpoints. TalkCo should compare those responsibilities against its own promises, rather than accept or reject every design under a single “decentralized messaging” label.
#Where the projects align
Useful local state, replaceable delivery, endpoint validation and explicit partition behavior support TalkCo’s aims. So does separating application authority from possession of a decryption key. Reuse is valuable when an accepted implementation removes work without importing an incompatible identity, evidence or governance model.
A recurring conversation, a private file, a shared record and a finite storage lease serve different purposes. Choosing a group engine should not require turning all of them into rooms. Likewise, compatibility with another network should not silently change an audience, retention rule or authentication claim.
#Protocol and architecture
#Three architectures must remain distinct
The following diagrams redraw the 2026-09-24 study’s inspected and historical architectures. They do not report fresh package verification on 2026-10-24. Exact source pins, discrepancies and scope are retained in the protocol findings and validation ledger.
#Variation — Historical embedded Dendrite and Pinecone
The supplied Dendrite/Pinecone diagram belongs to this historical family, including browser/WASM experiments and version-specific low-bandwidth paths. It does not establish current maintenance, mobile readiness, present wire compatibility or offline custody. It must not be used as the architecture of the inspected Tachyon client or as a timeless description of Neutrino.
#Variation — Neutrino embedded homeserver and transport adapters
This is a smaller embedded-homeserver/state-resolution laboratory. Encrypted links do not establish application E2EE, unlinkable identity, ordinary-Matrix interoperability or production security. The inspected composition binds a persistent node identity to federation addressing and advertises identity with a display name during BLE discovery; TalkCo would need an explicit context-to-device-to-route policy.
At the inspected revision, source contains signing and verification even though the README says signatures are absent. trusted_network = false selects signed mode; the base configuration defaults to trusted-network mode. The existence of that code does not establish an Android binary’s configuration or an independent audit. The source ledger pins the finding.
#Variation — Tachyon client-owned rooms with DMLS epochs
The archived Aurora deployment moves virtual-room state resolution to the client while retaining Matrix account/device services and routing metadata. Stable user/device and room/epoch handles remain visible in the inspected wrappers. This is supported by the pinned Tachyon wire analysis, not a claim about every Matrix deployment.
The archived two-peer experiment demonstrates a branch in the message event graph. It does not demonstrate arbitrary cryptographic membership-epoch fork convergence. The observed 86,400-second retired-epoch message-ratchet setting is a specific retention policy, not a claim about every historical secret. DMLS also does not supply a completed TalkCo deniability construction.
#The finding that changes the comparison
A client-owned virtual room is structurally closer to the endpoint authority required by TalkCo’s A0/A1 profiles than conventional server-owned room state. The useful experiment is therefore a Tachyon-style engine at the recurring-group seam. Its total adaptation and maintenance cost must be compared with OpenMLS or another accepted group engine under the same context, history and evidence contract.
This does not require Tachyon to become the foundation of every profile sphere, channel, attachment, workspace record or custody operation. Nor does it require embedding Neutrino beneath it. An additional implementation must remove a named responsibility rather than duplicate an existing state machine.
#Where adaptation is structural rather than cosmetic
| Boundary | TalkCo requirement and adaptation question |
|---|---|
| Contextual identity | Can device, room/epoch and transport handles be scoped without publishing a universal owner link? Renaming a profile does not change protocol correlation. |
| Private evidence | Review invitations, identity/device bindings, MLS authentication, checkpoints and recovery together. Encrypting signed history does not make it non-transferable to a willing recipient. |
| Object meaning | Keep reading, editing, accepted authorship, publication consent and storage responsibility distinct. A room is not the default unit of every object. |
| Partition policy | Permit useful authorized local edits while defining freshness for new disclosure. Convergence cannot retract a key already released before removal is learned. |
| Custody | Add finite leases, redundancy, repair, expiry, quotas and recovery where needed. A homeserver, relay acknowledgment or live connection does not supply those contracts. |
| Recipient attention | A remote-server receipt must not be presented as a human read receipt or create unsolicited presence. The sender’s transport observation does not establish a recipient’s attention. |
These are protocol and trust-boundary questions. TalkCo need not inherit Matrix IDs as its person model, room power levels as its entire permission model, Matrix media storage as custody, or signed room history as the default evidence for private speech.
#New source findings that prevent misleading conclusions
The September source inspection found a concrete attention boundary: Neutrino can expose a remote-server delivery acknowledgment using m.read; its configuration comments distinguish that from actual reading. The inspected Iroh/BLE composition explicitly enabled it. A TalkCo adapter must preserve the narrower meaning or disable the signal where its context policy forbids it (source ledger).
README statements, available code paths, selected configuration, observed demonstrations and audited security are different evidence classes. The signing-mode discrepancy and the limited Tachyon branch demonstration above are reasons to preserve those distinctions, not evidence that an untested configuration is secure.
#Research disposition, not a stack replacement
| Independent choice | Candidates and bounded role |
|---|---|
| Core implementation ownership | Smaller A0-oriented core or integrated-workspace countercandidate, with one accepted owner for each state machine |
| Recurring-group engine | OpenMLS or the bounded Tachyon-style client-owned room/DMLS comparison |
| Delivery or interoperability | An opaque Matrix carrier where its metadata contract is acceptable; ordinary Matrix conversation/export as an explicit separately labeled boundary |
| Embedded-state research | Neutrino for federation/state-resolution and transport experiments where its machinery removes named work |
The group comparison must use the same object, authority, context, recovery and observer fixtures as the other candidates. Build the smallest experiments needed to decide those seams; do not build three full products or stack several authoritative logs for the same state. A retained Matrix integration needs a clear exit path and must not change a private context’s meaning silently.
#Evidence and open questions
#Read the evaluation
| Document | Owns |
|---|---|
| Protocol and implementation findings | Historical Dendrite/Pinecone, inspected Neutrino/Iroh source, Tachyon’s two graphs, transport/confidentiality boundaries and source discrepancies |
| Composition, conflicts and adapter boundaries | TalkCo compatibility, state-machine ownership, changes by layer, existing alternatives and unresolved architecture questions |
| Evidence, computational models and falsification plan | Source pins, evidence classes, modeled costs, partition schedules, reproducible commands and unexecuted integration tests |
The master research table remains the sole comprehensive inventory. This page owns the Matrix comparison, not a competing live package catalogue.
#Scope of this study
The 2026-09-24 study inspected TalkCo’s repository and the existing Tachyon archive, refreshed primary P2P Matrix/MLS sources, and inspected selected Neutrino and Neutrino/Iroh source files at recorded revisions. Its computational model is an analytical litmus test, not execution of either implementation. The Aurora site could not be freshly retrieved in that study; deployment findings refer to the archive captured on 2026-09-23. No live Matrix account, real-room traffic, mobile radio experiment or independent cryptographic audit was used.
The 2026-10-24 revision reorganizes that evidence and adds architecture diagrams. It does not report a new source refresh, runtime test or audit. The questions in #1–#13 remain open. Exact Tachyon source revisions, production build/configuration evidence, controlled homeserver/device fixtures and an agreed observer and authentication model are the highest-value additional inputs.