Agent/provider portfolio study 06

.github/workflows/ci.yml

@@ -23,6 +23,8 @@ jobs:
       - run: python -m unittest discover -s tests -v
       - run: python -m mnel demo --workspace build/demo
       - run: python -m mnel ledger verify build/demo/evidence.jsonl
+      - run: python -m mnel provider-study-reference --workspace build/provider-study-reference
+      - run: python -m mnel ledger verify build/provider-study-reference/provider-portfolio-evidence.jsonl
       - run: git diff --check
 
   rust-provider-runtime:

CHANGELOG.md

@@ -12,6 +12,14 @@
   learned-provider artifact with deterministic serialization, reload, OOD abstention, and
   diagnostic-only observations.
 - Add the no-network `mnel distill-reference` study and distillation record schema.
+- Add a structurally distinct tabular nearest-centroid learned provider with bounded
+  feature scaling, explicit calibration data, artifact identity, deterministic reload,
+  OOD detection, and diagnostic-only observations.
+- Add `mnel provider-study-reference`: a no-network heterogeneous portfolio study with
+  random/heuristic/single/diversity routing, baseline comparisons, disagreement and
+  OOD/abstention metrics, cold/warm/resource measurements, optional energy reporting,
+  hidden-transfer visibility checks, and evidence-bound provider lifecycle records for
+  admission, quarantine, retirement, and rollback.
 
 ## 0.3.0a0 — unreleased
 

README.md

@@ -126,6 +126,11 @@ copy their authority or silently create substitute implementations.
 - `mnel distill-reference`, a no-network held-out study that writes an evidence ledger
   and exercises controls, hidden-transfer access, strategy transfer, and a tiny
   reloadable transition-frequency learned provider. Its observations remain diagnostic.
+- `mnel provider-study-reference`, a no-network heterogeneous portfolio study that trains
+  a transition-frequency provider and a structurally different tabular nearest-centroid
+  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.
 
 ## Install
 
@@ -211,6 +216,19 @@ or future-final records, freezes the transfer prediction before held-out evaluat
 and reports measurements rather than a success verdict. Its reference feature grouping
 is an explicit bounded baseline, not semantic understanding.
 
+Run the deterministic heterogeneous learned-provider portfolio study:
+
+```bash
+mnel provider-study-reference --workspace build/provider-study-reference
+mnel ledger verify build/provider-study-reference/provider-portfolio-evidence.jsonl
+```
+
+The transition-frequency and tabular nearest-centroid artifacts use separate snapshot
+families and feature/objective identities. Routing preserves each observation; agreement,
+disagreement, calibration, and transfer measurements are evidence, not consensus or
+evaluator authority. The reference HMM is not silently applied to incompatible tabular
+inputs.
+
 Verify and summarize the resulting ledger:
 
 ```bash

docs/ARCHITECTURE.md

@@ -114,6 +114,15 @@ and inconclusive experience. Records are content-identified and chained in order
 Learned observations retain declaration, weight, feature, query, snapshot, calibration,
 and limitation identities before they are eligible for later study.
 
+The 0.4 provider-study plane adds a small explicit portfolio contract on top of that
+ledger. `mnel.provider_study` binds provider artifacts to training/calibration dataset
+identities, supported snapshot families, artifact/model identities, and diagnostic-only
+metadata. It runs a transition-frequency provider beside a structurally distinct tabular
+nearest-centroid provider and explicit random/heuristic controls. Routing is recorded,
+not learned authority; every observation remains separate. Cold load, first inference,
+warm samples, artifact/model bytes, bounded Python allocation, calibration, disagreement,
+OOD, abstention, and optional energy availability are measurements.
+
 ### Verified distillation and study plane
 
 The 0.4 data plane makes reuse explicit without replacing evidence:
@@ -138,6 +147,13 @@ Study arms and ablations carry equal-budget declarations and deterministic ident
 The `distill-reference` command is a synthetic, no-network study surface, not an
 evaluator or a substitute for Forge, Fabric, MNCS, MNCDS, or RAVEL.
 
+Provider lifecycle records separately capture candidate, development admission,
+transfer-pending, transfer admission, quarantine, retirement, and rollback. Admission
+criteria identify artifact/calibration/resource/OOD/authority checks; a provider cannot
+transition on its own score, and quarantine preserves its artifact and evidence. The
+portfolio reference command is a bounded local study, not a production provider
+orchestrator and not a native ABI export.
+
 ### Distillation plane
 
 VED proposes compact principles and strategies while preserving source lineage,

docs/LEARNED_PROVIDER_RUNTIME.md

@@ -172,6 +172,18 @@ diagnostic provider. It is not a verifier and is not exported through ABI v1; na
 export remains a later task once the artifact contract has a measured compatible runtime
 boundary.
 
+The companion `mnel.provider_study` module adds a second, structurally different
+CPU reference provider: a bounded tabular nearest-centroid artifact with explicit
+normalization parameters, calibration dataset identity, OOD threshold, deterministic
+serialization, and reload equivalence. The portfolio harness compares both providers
+with seeded-random and explicit heuristic controls, while preserving separate provider
+observations and recording cold-load, first-inference, warm, model-size, artifact-size,
+allocation, calibration, disagreement, and OOD measurements. Its admission, quarantine,
+retirement, and rollback records are control-plane evidence; they do not change the
+stable C ABI or create evaluator authority. The existing ABI v1 query has no portable
+model-artifact initialization field, so this iteration deliberately does not smuggle a
+Python artifact path through global state or claim native export.
+
 ## Native-language exceptions
 
 A non-Rust provider may enter `native-trusted` only when its manifest includes:

docs/LEARNING_MODEL.md

@@ -83,3 +83,15 @@ equal-budget, and hidden-transfer controls. It computes retrieval and calibratio
 measurements and trains a tiny transition-frequency diagnostic provider that serializes
 and reloads deterministically. These are measurements from a bounded synthetic fixture,
 not evidence of general learning improvement or evaluator authority.
+
+The companion `mnel provider-study-reference` path extends this into a small controlled
+portfolio: a transition-frequency provider and a structurally different tabular
+nearest-centroid provider are trained from development-visible records, calibrated from
+separate identified examples, serialized/reloaded, and evaluated on a held-out fixture
+through random, heuristic, single-provider, and diversity routing policies. It records
+pairwise disagreement, bounded error correlation, abstention/OOD behavior, cold-load,
+first-inference, warm-latency, model/artifact size, and optional energy measurements.
+Provider admission, quarantine, retirement, and rollback are explicit diagnostic
+lifecycle records. The synthetic labels and transfer policy are study evidence, not
+MNEL/MNCS verdicts; the Rust HMM is retained as a separate native baseline because it is
+not input-compatible with the mixed compact fixture.

docs/RESEARCH_PROTOCOL.md

@@ -51,3 +51,13 @@ The current 0.4 reference study is deliberately synthetic and deterministic. Its
 transition-frequency provider is a diagnostic observation source with explicit artifact,
 training dataset, feature extractor, calibration, and reload identities. It is not a
 verifier, evaluator, or promotion mechanism.
+
+The provider portfolio control surface is also deterministic and equal-budget by
+construction. `provider-study-reference` trains separate transition and tabular provider
+families, keeps calibration distinct from training, and applies seeded random, explicit
+heuristic, single-provider, and architecture/objective diversity routing. It preserves
+per-provider observations and reports disagreement, joint/overlapping failure only when
+reference labels exist, abstention, OOD, latency, memory scope, and energy availability.
+Admission and rollback are evidence-bound records, not automatic promotion. The fixture
+does not claim that two providers are independent evaluators or that a successful
+synthetic transfer generalizes.

docs/ROADMAP.md

@@ -68,15 +68,26 @@
   precision/recall/hit-rate/reuse/diversity metrics plus bounded calibration metrics;
 - **Implemented:** deterministic `mnel distill-reference` study and a tiny reloadable CPU
   transition-frequency learned provider with OOD abstention and diagnostic-only output;
-- **Started:** heterogeneous learned-provider portfolio training and calibration;
+- **Implemented:** a deterministic heterogeneous reference portfolio with a structurally
+  distinct tabular nearest-centroid provider, explicit calibration records, deterministic
+  serialization/reload, routing controls, disagreement/correlation/OOD/abstention/resource
+  measurements, and diagnostic-only evidence ledgers;
+- **Implemented:** provider candidate/admission, transfer-pending, quarantine, retirement,
+  and rollback records with explicit evidence checklists;
+- **Started:** broader provider portfolios and native export of Python-trained artifacts;
 - export Python-trained providers into the versioned native runtime boundary;
-- **Started:** compare learned providers against deterministic and classical baselines;
+- **Implemented:** compare the reference providers against seeded-random and explicit
+  heuristic controls; the Rust HMM remains a separate not-applicable native baseline for
+  this compact mixed snapshot fixture;
   the reference transition study records deterministic and heuristic controls; the
   existing Rust HMM baseline is not input-compatible with this compact fixture and is
   retained as a separate native runtime baseline;
-- random, heuristic, single-provider, and diversity-routed controls;
-- correlated-error, disagreement, abstention, and out-of-distribution studies;
-- useful confirmed probes per operation, latency, memory, energy, and cold-start metrics;
+- **Implemented:** random, heuristic, single-provider, and diversity-routed controls;
+- **Implemented:** correlated-error, disagreement, abstention, and out-of-distribution
+  measurements, plus cold-load/first-inference/warm-latency and bounded Python allocation
+  measurements;
+- **Started:** energy measurements; an optional reader records trusted joules when supplied,
+  otherwise the study records unavailable rather than estimating;
 - non-Rust native exception studies with benchmark and threat-review identities;
 - hidden-transfer admission, quarantine, retirement, and rollback workflows;
 - optional small proposer-model distillation from verified traces.

docs/THREAT_MODEL.md

@@ -81,6 +81,16 @@ heuristic limitation, so it cannot masquerade as semantic understanding. Distill
 retrieval, calibration, and learned-provider observations remain proposal or diagnostic
 records; none includes evaluator verdict authority.
 
+The provider portfolio adds further guards: provider training and calibration accept only
+explicit development views; the held-out transfer fixture is exposed through a separate
+transfer-evaluator identity after prediction freeze; provider artifacts are hash-bound and
+reload-tested; and lifecycle transitions require evidence identities and a policy identity.
+Random and diversity routing cannot expand snapshot compatibility, and pairwise
+disagreement is not treated as error without an identified reference outcome. Optional
+energy readings are recorded as unavailable when no trusted source exists rather than
+estimated from latency. A small tabular provider is intentionally bounded and CPU-only;
+native export remains outside this iteration rather than weakening the ABI boundary.
+
 ### Apparent independence
 
 Multiple local machines run the same operator-controlled stack. This is replication,

schemas/mnel-calibration.schema.json

@@ -0,0 +1,27 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://github.com/epi13/Machine-Native-Experimental-Learning/schemas/mnel-calibration.schema.json",
+  "title": "MNEL identified calibration records",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema",
+    "snapshot_identity",
+    "source_record_identity",
+    "expected_label",
+    "visibility",
+    "record_identity",
+    "authority",
+    "semantics"
+  ],
+  "properties": {
+    "schema": {"const": "mnel-calibration-record/0.4"},
+    "snapshot_identity": {"type": "string", "minLength": 1},
+    "source_record_identity": {"type": "string", "minLength": 1},
+    "expected_label": {"type": ["integer", "null"], "enum": [0, 1, null]},
+    "visibility": {"enum": ["development-visible", "selection-observed-not-repairable"]},
+    "record_identity": {"type": "string", "minLength": 1},
+    "authority": {"const": "diagnostic-only"},
+    "semantics": {"const": "calibration-example; not-a-verdict"}
+  }
+}