Agent/distributed mnel fabric 08

.github/workflows/ci.yml

@@ -27,6 +27,8 @@ jobs:
       - run: python -m mnel ledger verify build/provider-study-reference/provider-portfolio-evidence.jsonl
       - run: python -m mnel family-integration-reference --workspace build/family-integration-reference
       - run: python -m mnel ledger verify build/family-integration-reference/family-integration-evidence.jsonl
+      - run: python -m mnel fabric-reference --workspace build/fabric-reference
+      - run: python -m mnel ledger verify build/fabric-reference/fabric-evidence.jsonl
       - run: git diff --check
 
   rust-provider-runtime:

CHANGELOG.md

@@ -2,6 +2,17 @@
 
 ## 0.4.0a0 — unreleased
 
+- Add bounded distributed MNEL workloads and workload graphs on the current public MNCS
+  Fabric boundary. `mnel fabric-reference` runs heterogeneous reference providers on
+  two logical Fabric workers, preserves worker/receipt identities, exercises replicated
+  dispatch, and performs deterministic transition-frequency shard aggregation. Hidden
+  data, duplicate/overlapping shards, mismatched artifacts, and incomplete results fail
+  closed or remain UNKNOWN. Optional network configuration uses TLS transports and
+  explicitly requires operator-pre-staged bundles; no SSH or implicit transfer is added.
+- Extend the Forge Provider Protocol 0.1 adapter with distributed workload, worker
+  capability, shard provenance, and reconciliation diagnostics. Protocol version remains
+  0.1 and all new responses remain diagnostic-only.
+
 - Add a strict MNCS Forge Provider Protocol 0.1 adapter and project-scoped Forge config;
   capabilities and analysis responses are bounded, one-line, diagnostic-only records.
 - Add pinned MNCS-family compatibility snapshots and a `family-integration-reference`

README.md

@@ -49,7 +49,7 @@ conventional neural-weight training.
 > source-preserving distillation groups, lineage-bound strategies, negative-memory
 > retrieval, equal-budget study controls, retrieval/calibration metrics, and a tiny
 > reloadable CPU transition-frequency diagnostic provider. It still does not provide process isolation,
-> unattended model execution, distributed scheduling, protected final custody, formal
+> unattended model execution, protected final custody, formal
 > MNCS/MNCDS conformance, or automatic RAVEL promotion.
 
 The repository also contains a dependency-aware MNCS-family integration reference:
@@ -59,6 +59,18 @@ Fabric job through its public service, consumes the experimental MNCS receipt, a
 the result as diagnostic evidence. Live sibling availability is reported explicitly; no
 external checkout is required for ordinary MNEL CI.
 
+The `mnel fabric-reference` command is a network-free distributed execution study. It
+uses Fabric's public local controller/worker boundary to run the transition-frequency
+and tabular centroid providers on two logical workers, records replicated execution
+observations, and aggregates identified transition-training shards. The logical workers
+share one process and host, so this is execution evidence rather than independent
+evaluation. Live TLS workers require explicit operator configuration and pre-staged
+bundles; MNEL does not copy artifacts over SSH or use hidden remote paths.
+
+The operator-only network entrypoint is `mnel fabric-run --config ... --plan ...
+--manifest ...`; it accepts only Fabric's bounded fixed-argv plan/manifest pair and
+fails closed when trust material or the declared pre-staged bundle identity is absent.
+
 ## Core rule
 
 **Investigators and learned providers may propose knowledge. They may not declare it true.**
@@ -138,6 +150,10 @@ copy their authority or silently create substitute implementations.
   provider, calibrates and reloads both, compares seeded-random and heuristic controls,
   measures disagreement/OOD/abstention/cold-warm resource behavior, and records explicit
   admission, quarantine, retirement, and rollback evidence. It does not issue a verdict.
+- `mnel fabric-reference`, a no-network multi-worker Fabric study that distributes
+  reference expert inference and transition-frequency sufficient-statistic shards,
+  exercises replication and missing-shard semantics, and verifies deterministic model
+  equivalence without granting execution records evaluator authority.
 
 ## Install
 
@@ -236,6 +252,13 @@ disagreement, calibration, and transfer measurements are evidence, not consensus
 evaluator authority. The reference HMM is not silently applied to incompatible tabular
 inputs.
 
+Run the local distributed reference path:
+
+```bash
+mnel fabric-reference --workspace build/fabric-reference
+mnel ledger verify build/fabric-reference/fabric-evidence.jsonl
+```
+
 Verify and summarize the resulting ledger:
 
 ```bash

compat/mncs-family-compatibility-0.1.json

@@ -1,13 +1,13 @@
 {
   "schema": "mnel-family-compatibility/0.1",
-  "snapshot_version": "2026-08-08",
+  "snapshot_version": "2026-08-09",
   "projects": [
     {"repository": "mncs-forge-mcp", "commit": "7710ea606bd592e0be95957c96132e8732fbb955", "public_contract": {"version": "Provider Protocol 0.1", "types": ["capabilities", "analysis_request", "analysis_response"]}},
-    {"repository": "mncs-fabric", "commit": "fd6a1e1fe617b77402a3d40c278776dd8f159fb0", "public_contract": {"version": "mncs-fabric public service 0.1", "types": ["FabricService", "job-plan", "execution-record", "reconcile"]}},
-    {"repository": "machine-native-complexity-standard", "commit": "1f0a1f903379d740eeeb59fe8689aed398bfb47c", "public_contract": {"version": "execution receipt 0.1-experimental", "types": ["mncs-execution-receipt", "claim_boundary"]}},
-    {"repository": "MNCS-Commons", "commit": "5afff351efd3563febb7b3149d7de711d3342a44", "public_contract": {"version": "commons.mncs.dev/v0alpha1", "types": ["Observation", "Replication", "Advisory", "commons-bundle/v0alpha1"]}},
-    {"repository": "mncs-language", "commit": "26cd7f015cb857abe3f0601780de096e04dea7b4", "public_contract": {"version": "semantic identity boundary", "types": ["opaque semantic identity", "HIR identity"]}},
-    {"repository": "RAVEL", "commit": "4b7c3c5503ec6bd11a7ffb96cbb32599cd1f342c", "public_contract": {"version": "ravel-development-record/0.6-preregistration", "types": ["candidate", "development record"], "promotion_authorized": false}}
+    {"repository": "mncs-fabric", "commit": "740b4b3a2590f76aa6eeb5365d3bdd8a40e39964", "public_contract": {"version": "mncs-fabric public controller/service 0.2.0a0", "types": ["FabricService", "LocalController", "NetworkController", "TLSNetworkTransport", "job-plan", "execution-record", "execution-receipt", "challenge/replay", "persistent worker evidence", "reconcile"]}},
+    {"repository": "machine-native-complexity-standard", "commit": "80f08d312dce963265c7f69ac5b4bae8245bd692", "public_contract": {"version": "execution receipt 0.1-experimental", "types": ["mncs-execution-receipt", "immutable execution bundle", "claim_boundary"]}},
+    {"repository": "MNCS-Commons", "commit": "b1eb5a1081bbb63ee3a6284e8046035bd72a47bc", "public_contract": {"version": "commons.mncs.dev/v0alpha1", "types": ["Observation", "Replication", "Advisory", "commons-bundle/v0alpha1"]}},
+    {"repository": "mncs-language", "commit": "f234cc8079faa5895a38b7abce0c96031f7d2565", "public_contract": {"version": "semantic identity boundary", "types": ["opaque semantic identity", "HIR identity"]}},
+    {"repository": "RAVEL", "commit": "d572d68ab9c8eaf163425748d44729aaa8028e98", "public_contract": {"version": "ravel-development-record/0.6-preregistration", "types": ["candidate", "development record"], "promotion_authorized": false}}
   ],
   "authority": "diagnostic-only",
   "limitations": ["Pinned shapes are compatibility evidence, not live conformance claims."]

docs/ARCHITECTURE.md

@@ -101,6 +101,21 @@ MNCS Fabric is the intended distributed execution and reconciliation substrate.
 also supports a deterministic single-host reference path. Execution produces raw facts,
 not semantic acceptance.
 
+The distributed layer (`mnel.fabric_execution`) models bounded identified workloads and
+acyclic workload graphs. Fabric answers which compatible worker executes a workload;
+the provider placement policy separately answers CPU, full CUDA, or sequential offload
+on that worker. The local backend calls Fabric's public `LocalController`/`LocalWorker`
+boundary. The network backend calls `NetworkController` with `TLSNetworkTransport` and
+operator enrollment state, and rejects missing trust material or non-matching pre-staged
+bundle identities. It never turns remote completion into correctness, independence,
+conformance, or promotion.
+
+Coarse-grained training is supported for the reference transition-frequency and centroid
+providers. Workers produce identified partial sufficient statistics; the controller
+rejects duplicate, overlapping, missing, wrong-dataset, or hidden shards and merges in
+deterministic shard order. A missing required shard makes the study incomplete rather
+than silently producing a canonical full model.
+
 ### Evaluation plane
 
 An immutable evaluator derives every hard gate. Failed gates cannot be compensated by a
@@ -170,7 +185,9 @@ capability, authority, disclosure, partition access, or cost ceilings.
 ### MNCS-family integration plane
 
 Forge is the preferred external control/evidence plane. The `mnel.forge_provider` module
-implements only Provider Protocol 0.1 capabilities and bounded analysis responses. The
+implements only Provider Protocol 0.1 capabilities and bounded analysis responses,
+including distributed workload, worker capability, shard provenance, and reconciliation
+diagnostics. The
 family reference path binds an MNEL study and provider artifact to a Fabric manifest, calls
 the public `FabricService`, consumes the typed experimental execution receipt, and stores a
 companion MNEL observation. Forge controls the declared workflow, Fabric records execution,

docs/INTEGRATIONS.md

@@ -64,6 +64,14 @@ normalizes the observation into the MNEL ledger. Repeated execution is labelled
 `local-in-process-replication`; it is not multi-host independence, authenticated worker
 enrollment, or protected custody.
 
+`mnel fabric-reference` uses the current public `LocalController` and `LocalWorker`
+interfaces for a two-logical-worker, network-free study. `NetworkFabricBackend` uses the
+documented `NetworkController`/`TLSNetworkTransport` pair only after validating explicit
+CA, client-certificate, key, trust-store, capability, and pre-staged bundle references.
+Bulk artifact transfer remains outside MNEL until Fabric exposes a public verified
+transfer profile. A worker is an execution location, not an expert identity; provider
+artifact and model identities remain bound in every workload and observation.
+
 The same path emits an inert Commons Observation-shaped record and a RAVEL 0.6 proposal
 context fixture. Neither is published or granted trust-domain, evaluator, freeze, selection,
 or promotion authority.

docs/LEARNED_PROVIDER_RUNTIME.md

@@ -205,7 +205,12 @@ artifact and does not establish a general language preference.
 4. Add Forge snapshot producers and reuse measurements.
 5. Export one Python-trained neural provider and compare it with the baseline.
 6. Add WASM quarantine only after native measurements establish the overhead budget.
-7. Integrate Fabric placement after single-host identity and replay behavior is stable.
+7. The local distributed reference now composes MNEL provider placement with MNCS Fabric
+   worker placement: Fabric selects a compatible machine while the provider runtime still
+   selects CPU, full CUDA, or sequential CPU offload on that machine. Expert artifact,
+   model, calibration, snapshot, and capability identities remain bound. Network dispatch
+   is operator-only and requires TLS trust material plus explicit pre-staging; Fabric
+   execution is observation evidence, not evaluator authority.
 
 Each stage must preserve the current diagnostic authority boundary and may terminate in
 `UNKNOWN` rather than silently widening capability.

docs/ROADMAP.md

@@ -97,15 +97,21 @@
 
 ## 0.5 — MNCS Fabric execution
 
-- **Started:** content-addressed local experiment material binding to Fabric manifests;
-- **Started:** capability-aware local dispatch through the public `FabricService` boundary;
-- **Started:** typed MNCS execution-receipt collection and MNEL normalization;
-- **Started:** local duplicate/replay and same-node reconciliation evidence;
-- authenticated worker enrollment;
-- replicated and sharded trial matrices;
-- remote node-loss handling, protected custody, and scaling measurements;
-- heterogeneous learned-provider placement by snapshot locality and node capability;
-- ABI compatibility and provider-artifact admission across Fabric nodes.
+- **Implemented (local reference):** content-addressed identified workloads and bounded
+  acyclic workload graphs;
+- **Implemented (local reference):** capability-aware dispatch through Fabric's public
+  `LocalController`/`LocalWorker` boundary with provider-artifact identity binding;
+- **Implemented (local reference):** typed Fabric execution/receipt observations retained
+  as MNEL evidence without evaluator authority;
+- **Implemented (local reference):** replicated expert runs and deterministic reconciliation
+  observations across distinct logical worker identities;
+- **Implemented (local reference):** transition-frequency and nearest-centroid sharded
+  sufficient-statistic training with duplicate/overlap/missing/visibility validation;
+- **Started:** optional authenticated remote dispatch through `NetworkController` and
+  `TLSNetworkTransport`; current network mode requires operator-pre-staged bundles;
+- **Started:** node-loss/UNKNOWN handling, locality hints, and scaling measurements;
+- **Started:** authenticated worker enrollment consumption and provider-artifact admission
+  across physical nodes; live remote execution remains operator-only and unverified here.
 
 ## 0.6 — RAVEL integration study
 

docs/THREAT_MODEL.md

@@ -107,6 +107,22 @@ unasserted. Commons records are inert and unpublished; RAVEL material is proposa
 only. Forge configuration and local sibling checkouts remain operator-controlled inputs and
 must be reviewed for substitution or drift.
 
+### Distributed Fabric boundary
+
+MNEL sends only bounded, identity-bound workloads to Fabric. A compromised or stale
+worker can return a malformed result, wrong artifact identity, poisoned sufficient
+statistics, replayed output, or omit a shard. Fabric protocol/TLS identity proves the
+configured peer and replay binding; it does not prove worker honesty, independent
+evaluation, or hardware attestation. MNEL validates dataset, shard, provider, model,
+feature-extractor, visibility, and bundle identities before aggregation, preserves raw
+Fabric dispatch/result/receipt identities, rejects duplicate/overlapping shards, and
+reports missing workers as UNKNOWN/incomplete. Multiple machines operated by one party
+are cross-node execution evidence, not independent validation.
+
+Remote bundle staging is an explicit operator precondition. MNEL does not add SSH,
+developer-directory mounts, network downloads, or ambient model paths to the semantic
+execution protocol. Certificate/private-key paths remain outside ledgers and source.
+
 ## Current residual risks
 
 The foundation now validates and loads identified native provider libraries through a

examples/mnel-fabric.toml

@@ -0,0 +1,16 @@
+# Operator-only example. Paths refer to external trust material and must not be
+# committed with real certificates, private keys, or worker addresses.
+controller_id = "mnel-controller-example"
+state_path = "build/fabric-controller.jsonl"
+# NetworkFabricBackend requires a verified, pre-staged bundle before dispatch.
+pre_staged_bundle_identity = "sha256:replace-with-the-expected-bundle-identity"
+
+[[workers]]
+worker_id = "worker-a"
+host = "replace-with-operator-endpoint"
+port = 4433
+capabilities = ["python", "os:linux"]
+ca_file = "operator-pki/ca.pem"
+client_cert = "operator-pki/controller.pem"
+client_key = "operator-pki/controller.key"
+trust_store = "build/fabric-trust.jsonl"

mncs-forge.toml

@@ -51,12 +51,12 @@ useful_benefit_objective = "README.md"
 id = "mnel-family-provider"
 name = "MNEL family diagnostic provider"
 identity = "mnel-family-provider-protocol-v1"
-version = "0.1"
+version = "0.2"
 command = ["python", "-m", "mnel.forge_provider"]
 transport = "stdio-jsonl"
 required = false
-capabilities = ["evidence_derivation", "mncs_bundle_validation", "provider_study_summary"]
-supported_constructs = ["identified-study-summary", "provider-artifact-binding", "execution-receipt-binding"]
+capabilities = ["evidence_derivation", "mncs_bundle_validation", "provider_study_summary", "distributed_workload_inspection", "fabric_worker_capability_summary", "distributed_training_provenance", "shard_completeness", "reconciliation_summary"]
+supported_constructs = ["identified-study-summary", "provider-artifact-binding", "execution-receipt-binding", "distributed-workload-identity-binding", "fabric-worker-capability-observation", "sharded-training-provenance", "cross-node-reconciliation-observation"]
 unsupported_constructs = ["hidden-transfer-content", "future-final-content", "evaluator-verdict", "ravel-promotion"]
 limitations = ["diagnostic-only provider; no external evaluator authority"]
 
@@ -79,3 +79,43 @@ provider_protocol = true
 provider_id = "mnel-family-provider"
 subject = "candidate"
 disclosure = "compact"
+
+[[workflows]]
+name = "mnel-fabric-reference"
+category = "evidence_derivation"
+mode = "development"
+command = ["python", "-m", "mnel.forge_provider"]
+provider_protocol = true
+provider_id = "mnel-family-provider"
+subject = "candidate"
+disclosure = "compact"
+
+[[workflows]]
+name = "mnel-fabric-reconcile"
+category = "evidence_derivation"
+mode = "development"
+command = ["python", "-m", "mnel.forge_provider"]
+provider_protocol = true
+provider_id = "mnel-family-provider"
+subject = "candidate"
+disclosure = "compact"
+
+[[workflows]]
+name = "mnel-fabric-plan"
+category = "evidence_derivation"
+mode = "development"
+command = ["python", "-m", "mnel.forge_provider"]
+provider_protocol = true
+provider_id = "mnel-family-provider"
+subject = "candidate"
+disclosure = "compact"
+
+[[workflows]]
+name = "mnel-fabric-train"
+category = "evidence_derivation"
+mode = "development"
+command = ["python", "-m", "mnel.forge_provider"]
+provider_protocol = true
+provider_id = "mnel-family-provider"
+subject = "candidate"
+disclosure = "compact"

schemas/mnel-distributed-study-matrix.schema.json

@@ -0,0 +1,15 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://mnel.dev/schemas/mnel-distributed-study-matrix.schema.json",
+  "title": "MNEL distributed study matrix",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["schema", "matrix_identity", "cells", "authority", "semantics"],
+  "properties": {
+    "schema": {"const": "mnel-distributed-study-matrix/0.5"},
+    "matrix_identity": {"type": "string", "pattern": "^sha256:[0-9a-f]{64}$"},
+    "cells": {"type": "array", "minItems": 1, "maxItems": 512, "items": {"type": "object"}},
+    "authority": {"const": "diagnostic-only"},
+    "semantics": {"const": "distributed-study-matrix; not-a-verdict"}
+  }
+}