Change Log
Complete version history of this specification. Every published version has been verified in real systems — 1.0 and 1.1 ran across Windows, macOS and iOS applications, over LAN and over an internet relay. 2.0 re-derives the specification from the implementation as it now runs, after the record model advanced and the text did not follow. Errata that change no requirement may be noted here without a version bump.
| Version | Date | Changes |
|---|---|---|
| 2.0 | 2026-08-09 | Re-derived and corrected as an executable contract. 1.0 and 1.1 were
implemented and interoperating across Windows, macOS and iOS, over LAN and relay. What
changed is that the runtime’s record model advanced — a two-section record, a
Merkle-derived address, a new signing payload — while the published text continued to
describe the earlier shape. Every normative clause below was re-checked against the
public specification, schemas, vectors and running implementations. Where they disagreed,
v2.0 now rules the construction explicitly and publishes executable evidence; no runtime
is treated as normative merely because it shipped first.
§8.8 Record Model — corrected v2.0 contract. The record is two-section: categories
carries CAT7, metadata carries the address, author,
timestamp, audience and descent. 1.x described a flat record with
createdAt at the top level; the runtime has never
emitted that shape. §8.8.2 gives the byte-exact
cognition address (per-category keys under a domain tag, combined by a
promote-odd Merkle root — never duplicate-the-last, which silently yields a
different address) and §8.8.4 gives the byte-exact signature payload (uniformly
length-prefixed, audience-bound, with per-category descent committed alongside the root
rather than inside it, so identical observations still collapse to one address).
Address form corrected. 1.x declared a version-tagged address prefix as normative. The runtime rejects that form: a key is valid iff it is cmb- plus exactly
64 lowercase hex. An implementation built to the 1.x text would have minted keys every
deployed node refuses — interoperation failure from following the specification.
Admission wording. The receiver evaluates each of the seven categories against its own anchors and then decides locally what to accept. 1.x called this “per-field admission” and described an admission outcome per category; measured against the runtime, the per-category step is evaluation evidence and the receiver makes one whole-record admission decision. Receiver autonomy is unchanged and unqualified — no sender and no coordinator can force admission of anything. CAT7 terminology. The seven members are categories throughout, replacing mixed use of “dimensions” and “fields” in prose. A category is the semantic member; a dimension is the length of the vector encoding it. The wire key is categories too — prose and wire now use one
word. It never enters an address or signature preimage, so the rename moved no address and
invalidated no signature.
Single-file artifacts are now /spec/mmp-v2.0.md and
.html.
mmp-v1.0.* remain published, frozen, for existing
citations.
v2.0 alignment errata (2026-08-13; no version bump). The current corpus now enforces direct-parent-only wire lineage and derives transitive provenance by traversing locally verified parent records; CAT7 embeddings are explicitly receiver-local rather than wire fields; cmb-encrypted is the single canonical sealed-frame
name; and every active core and relay frame is mapped to a closed JSON Schema through a
machine-readable, schema-validated registry. The authenticated v2 handshake now publishes
its exact transcript-hash session identifier, HKDF salt, role-specific finished-key labels,
directional traffic-key labels, proof payload and confirmation payload. It is the only
extension-negotiation contract, and structured CMB extension bytes use the assertion-bound
metadata.application container. Executable gates now reject
schema drift, stale v1 claims in current pages, unregistered artifacts, broken local links,
duplicate rendered IDs and unsigned metadata extension siblings.
|
| 1.1.0 registry note | 2026-07-13 | §16.4 registry: two Draft Candidate Extensions added — no change to any core requirement, no version bump (extensions are the §16 growth path of a final specification). Error Handling v0.1.0 (failure as a first-class cognition event: evidence-carrying corrective requests with lineage-borne parentage, receiver-autonomous volunteering, one-level repair, separate grounding of failure and fix) and CMB Trust Horizon v0.1.0 (validator-attested, knowledge-scoped trust-weight invariants; grants ride content-bound CAT7 text; Canon-retention separation). Both application-layer CMB conventions over MMP v1.1; both Draft — a reference deployment reports an experimental implementation; independent interoperability not yet established. |
| 1.1.0 | 2026-07-05 upd. 2026-07-07 | Soundness & completeness update (2026-07-07), from the formalization of the deployed
mechanism. The mesh-cognition formalization re-derived this specification’s claims and
the amendments are folded into 1.1.0 in place: §9.2.1 pins the redundancy invariants to the
nearest-anchor basis (δfnear = 1 − maxa cos —
the fused attention readout provably cannot satisfy them: a block identical to a stored anchor can score
δ = 0.127 once other anchors pull the readout); §9.2 discloses that admission is
evaluation-time-dependent with the derived flip window (aggregated category drift in
(0.286, 0.714) at defaults admits fresh, rejects late); §9.2.1’s cold-start
bootstrap-admit now discloses its security consequence (new §18.4
cold-start-capture threat row); §6.7 adds repeat verification (a recognised
grounding is never refused solely for redundancy — the redundancy band provably
self-quenches the outcome stream otherwise), the failure channel is load-bearing
clause (positive-only grounding provably locks onto stale favourites; observed failures must not be
selectively suppressed), and an informative note on consuming the outcome stream (decay-half-life
theory); §15.8 specifies the lineage tether — the root-anchored drift bound
that closes grounding-inheritance laundering; §18.3.1 disclosed
the enforcement scope of strict signature mode. Wire-compatible throughout: no new
frames, no new categories.
The Work Layer — grounding cognition in reality. Through 1.0.x, the mesh could observe, admit, remix, and validate — it could establish what its members believe. 1.1.0 adds the missing half: a way to record what held up in practice, and to make the cognition that survives both judgment and reality the durable substrate real work builds on. Everything below is normative as of this release; new sections are marked “New in 1.1.0” in place. §6.7 Grounding — outcomes carried by lineage. A grounding CMB ( intent: "ground",
commitment verified: / failed:,
parents = the cognition it grounds) records a real-world result — tests passed, work shipped, a
prediction resolved — as the evidence-based sibling of §6.4’s judgment-based validation.
An outcome is an attestation, never a fact: its weight follows the author’s earned
authority (§6.5–§6.6), groundedness is receiver-relative (only
attestations a node’s own SVAF admitted count), conflicting observations resolve
latest-wins on receiver-local time (a regression un-grounds; a backdated timestamp
cannot game the ordering), and a grounding CMB never advances lifecycle by itself —
elevation to the Canon is an explicit, accountable act under validator-or-above authority.
§6.3 The Canon tier — committed cognition persists. validated and canonical
CMBs are exempt from age-based retention while they hold that lifecycle, so a mesh’s earned
knowledge compounds across sessions instead of evaporating with the retention window. Protection is
from purge, not from demotion — inactive validated cognition may still decay to archived
(§6.4, §19) — and the store stays bounded.
§14.12 Work Sessions as Mesh Members — the capture profile that closes the loop: a work session joins as an ordinary member; its charter is the intent root, its decisions chain by lineage, and completion emits an artifact grounded by the session’s real outcome. Day two of a mesh starts ahead of day one because day one’s work is in the Canon. §14.11 remains reserved for Commissions (planned for 1.2.0). Supporting sections: §8.3.1 well-known intent values (informative, extensible registry — charter, decision,
artifact, ground; unknown
intents remain ordinary content and confer nothing); §15.7.2 outcomes are observations (observing a
real outcome IS new domain data, so grounding remixes satisfy §15.7 with no intent-keyed exemption
— the anti-echo invariant is untouched); §18.4 gains the fake-outcome-attestation threat
model (fabricated verified: steering, low-authority
failed: griefing) with its mitigation chain; §17.5 lists the
draft conformance requirements.
Incorporates the 2026-07-05 coherence errata — a full-corpus adversarial review (41 findings) folded into this release: §10 state blending re-grounded in CMB-admission influence, completing the 1.0.2 supersession (the deprecated hidden-vector blend’s coefficients now bound per-admission influence; §13.4’s formula corrected to match); §11.4 feedback authority resolved through the signed grant chain rather than self-declared handshake roles; group isolation re-derived as endpoint-enforced via §18.3.1 audience binding (the relay is a dumb pipe); the §7.1 frame-type registry completed (mood, relay frames); handshake schema reconciled (§20.1 group optional, lifecycleRole
sender-MUST); §17 conformance refreshed with testable requirements; plus editorial corrections
across citations, examples, and terminology.
Compatibility: fully wire-compatible — no new frames, no new categories; a 1.0.x node interoperates unchanged and treats grounding CMBs as ordinary CMBs. Reference-implementation status: two §6.7-adjacent mechanisms are specified ahead of the reference implementation (the §15.7.1 convention): the §6.4 inactivity archiver, and elevation-authority resolution through the §6.6 grant chain — the shipping implementation performs elevation as an explicit operator act pending earned-authority activation. Both are runtime work, outside this final specification. |
| 1.0.6 | 2026-07-04 | §5.9–5.11 Gateway Federation (informative pattern) — introduces an informative pattern for composing meshes: a node is a membrane over an arbitrary interior (atom = one agent; gateway = a node whose interior is a sub-mesh, presenting a boundary to exterior gateways). A gateway participates in its interior group and exchanges a lossy CAT7 projection with configured peers over a dumb boundary transport (HTTP), each keeping its own store — no center. Invariants that hold: no-center-per-level, partition-tolerance, §3.2 one-agent-one-node. The reference implementation is a prototype (observe-and-summarize; admit-then-reproject, signed/attested projections, and cross-mesh echo-dedup are unbuilt), and a production security bar — signed cmb- projection, origin-authenticated origin, anti-replay, boundary-scoped credential — is a prerequisite, not a shipped guarantee. This is a topology pattern, not a normative cross-mesh wire; single-mesh conformance is unchanged. |
| 1.0.5 | 2026-07-03 | §14.10 Operator Directives — Steering the Mesh — specifies how a human operator injects intent into a running mesh: a directive is an ordinary signed CAT7 CMB (perspective: "operator") emitted through the control plane’s node. Normative: a broadcast directive carries no privileged authority — every node MUST evaluate it through SVAF (§9.2) like any peer CMB and MAY reject it; steering is receiver-autonomous, not command-and-control (no router, no bypass). An implementation MUST NOT grant a broadcast directive elevated admission weight for originating from the operator (elevated influence comes only from earned authority, §6.5, evaluated identically for human and agent emissions); a directive MAY be directed to one node (§4.4.4/§9.2.2, delivery not admission); the per-node verdict SHOULD be recorded in the admission audit. Backward-compatible addition (patch). |
| 1.0.4 | 2026-07-02 | §12.8–12.15 Collective Query: the Ask → Synthesis Path — specifies the query-initiated Layer 5 flow: a question posed to the mesh as a type: "question" CMB is answered by a single cited synthesis no one agent held. Adds the four-stage path SELF-SELECT → ADMIT → SYNTHESISE → CRYSTALLISE alongside the inbound §12.2 pipeline. Normative additions: self-selection is receiver-autonomous and computed only from an agent’s own store (no router, SELF_SELECT_THRESHOLD default 0.1); each contribution carries lineage to its grounding and MUST pass SVAF (§9) before it can be synthesised; the single synthesis at the asking node MUST cite specific CMB ids and MUST NOT assert beyond them; the answer is crystallised back as an immutable type: "synthesis" CMB whose parents are the question key plus every citation, so the mesh’s cognition compounds across Asks. Includes the five Ask invariants (I-Ask-1–5) and marks the local-store grounding breadth (§12.14) as an implementation limitation, not an architectural constraint. Backward-compatible addition (patch). |
| 1.0.3 | 2026-06-16 | §15.7.1 Source-Novel Forwarding — carve-out distinguishing forwarding from the remix-paraphrase §15.7 forbids. An agent MAY re-emit an admitted observation it did not natively produce, carrying the inherited lineage root, when and only when that observation is source-novel to the receiver — its lineage roots are not already present in the receiver’s admitted store. This is not the value-only echo §15.7 prevents: a forwarded observation carries a source the receiver has not yet seen even though the forwarder adds no new domain data. Forwarding MUST NOT mint a fresh root for content that already carries one, and MUST NOT re-emit a source the receiver already holds — the anti-echo guarantee is preserved exactly. Forwarding SHOULD be non-selective, so every observation reaches the agents whose understanding depends on it. In short: remix requires new domain data; forwarding requires a new source. Backward-compatible addition (patch). |
| 1.0.2 | 2026-06-14 | §2.7 Hidden State Locality — states the invariant that a node’s hidden state (its Layer 6 LNN vectors h₁/h₂) MUST remain strictly local and MUST NOT cross the wire; only Cognitive Memory Blocks cross. Defines hidden state (private machinery) vs. the remixed CMB (communicable understanding), and the four reasons hidden state must stay local: sovereignty, auditability, semantic incompatibility across agents, and privacy. Supersedes the state-sync model: the state-sync frame and any exchange of h₁/h₂ vectors are deprecated; the peer-drift and state-blending mechanisms described in §5, §7, §9.1, and §10 from exchanged hidden-state vectors are superseded — peer influence is mediated entirely by CMBs evaluated through SVAF (§9.2). Resolves a self-contradiction between the “hidden state never crosses the wire” claim and the state-sync sections. |
| 1.0.1 | 2026-06-12 | Layer 6 renamed “xMesh” → “Cognitive State” to disambiguate from the xMesh runtime (naming note §1, §13; wire identifiers incl. xmesh-insight unchanged; published papers retain the legacy “xMesh (L6)” label). Normative additions, backward-compatible with the v1.0 contracts: §9.2.1 specifies δf as an admission interface — anchors-only baseline (incoming block excluded), cold-start non-evaluable-category exclusion + bootstrap-admit — ruling out self-referential collapse and cold-start starvation. §9.2.2 specifies the directed (peer-bound) vs autonomous (group-bound) delivery contract, separating delivery from memory admission: directed CMBs (§4.4.4 to = receiver) surface unconditionally; rejected broadcasts do not surface (mood excepted, §9.3). §18.3.1 specifies CMB signature verification (Ed25519 author signature + content-address integrity; forged/tampered blocks rejected) as the end-to-end authenticity layer above transport identity. |
| 1.0 | 2026-04-27 | Public-stable-API release. Marks the v0.2.x development cadence as complete and the protocol surface as production-stable. Contracts unchanged from 0.2.3; v0.2.x → v1.0 is a maturity declaration, not a breaking change. Note: arXiv:2604.19540 cites v0.2.x as the version implemented at paper-publication time; v1.0 covers the same contracts. |
| 0.1–0.2.3 | 2026-03-27 → 2026-04-27 | The development cadence. 0.1 (27 March 2026) was the initial public draft — the 8-layer architecture, the CAT7 seven-category schema, SVAF per-category admission, content-hash lineage and remix, and DNS-SD discovery. The 0.2.x series stabilised the wire contracts (handshake, frame registry, TCP + WebSocket relay transports) in production use; contracts were frozen at 0.2.3 and declared stable, unchanged, as 1.0. arXiv:2604.19540 documents the protocol as implemented in this era. |
| 0.2.3 | 2026-04-17 | Section 13.9 — Compact Channel Best Practices: CMB envelope header convention (RECOMMENDED) for structured message headers with signal keywords and focus tags. Lazy-load channel pattern (RECOMMENDED) for MCP server implementations: compact header push with on-demand full-content retrieval via sym_fetch, reducing mesh-traffic context consumption by ~75%. Token-count hint RECOMMENDED. Rolling message store with RECOMMENDED default of 200 messages. Signal-keyword priority table (informational): HALT > DIRECTIVE > RESULT > ACK. |
| 0.2.2 | 2026-04-06 | Section 11 — Feedback Modulation: how collective intelligence becomes self-correcting. Validator-authority CMBs with per-category reasoning modulate SVAF coupling weights and CfC temporal adaptation through the existing mesh cognition loop. Neuroscience-grounded: dopaminergic prediction error model with per-category direction and τ-modulated adaptation rate. Directive feedback for standalone domain knowledge injection. Validator-origin anchor weight 2.0 with role-grant verification. CfC state persistence across restarts. ABNF wire format grammar. CMB forward compatibility. Multi-relay failover. All cognitive content MUST use cmb frames. |
| 0.2.1 | 2026-04-02 | Node model: every autonomous agent MUST be a full peer node with own identity, coupling engine, and memory store. SVAF band-pass evaluation: four-class model (redundant/aligned/guarded/rejected) with per-category redundancy detection. CMB lifecycle: observed/remixed/validated/canonical/archived with anchor weight progression. Node lifecycle roles (participant/validator/anchor) with identity-bound validation authority and earned role progression. Validation authority for CMB lifecycle transitions bound to cryptographic node identity, not content. Semantic encoder SHOULD for SVAF drift computation. Handshake adds version, extensions, and lifecycleRole fields. Error frame type. Role-grant frame type. |
| 0.2.0 | 2026-03-27 | Formal specification published. 8-layer architecture. CAT7 CMB schema with lineage (parents + ancestors). SVAF per-category evaluation. Wire format normatively specified. Error frame. Frame type registry. Extension mechanism. JSON Schema. Connection state machine. Wire examples. |
| 0.1.0 | 2025-08-01 | Initial protocol design (Consenix Labs Ltd). 4-layer architecture. Scalar drift evaluation. |