Machine-Native Experimental Learning

Machine-Native Experimental Learning (MNEL) is an evidence-governed experimental learning framework in which autonomous investigators propose hypotheses, design bounded interventions, invoke deterministic tools and verifiers, and submit observed outcomes for causal attribution and verified-experience distillation.

MNEL does not treat model output as knowledge or authority.

investigators propose
        |
        v
learned micro-providers locate surprise and omitted-question candidates
        |
        v
Forge-compatible probes interrogate bounded claims
        |
        v
MNCS Fabric-compatible executors run identified experiments
        |
        v
immutable evaluators derive PASS / FAIL / UNKNOWN
        |
        v
causal attribution separates result from explanation
        |
        v
Verified Experience Distillation proposes principles and strategies
        |
        v
RAVEL receives separately governed candidate knowledge

The project explores a learning process built from persistent machine-readable experience, negative memory, causal attribution, transfer-gated principles, reusable strategies, and append-only candidate lineage rather than relying exclusively on conventional neural-weight training.

Current status: functional 0.4.0a0 study iteration. The repository now includes a backend-neutral accelerator placement policy, optional Torch/Accelerate adapter, process-local persistent Rust host, reusable identity-bound snapshots, bounded and normalized diagnostic results, failure quarantine, an executable Rust HMM baseline, deterministic runtime measurements, and bounded investigator context/workspace contracts, an executable local-harness/worktree path, and a validated Rust v1 dynamic provider loader, and an executable bounded Forge-oriented diagnostic lifecycle with compact snapshot views, a verifier registry, reference probes, witnesses, mutations, comparison, health, coverage, and deterministic skeptic candidates. It also includes 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, protected final custody, formal MNCS/MNCDS conformance, or automatic RAVEL promotion.

The repository also contains a dependency-aware MNCS-family integration reference: mnel family-integration-reference exposes a Forge Provider Protocol 0.1 adapter, validates pinned Forge/Fabric/Commons/RAVEL/Language compatibility shapes, executes a bounded local Fabric job through its public service, consumes the experimental MNCS receipt, and records 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.

The model or agent that creates a hypothesis must not also become the authority that accepts the resulting lesson. Evaluator identity, hard gates, partitions, resource ceilings, custody, and promotion remain outside the investigator and learned-provider surfaces.

Relationship to the project family

Project	MNEL role
MNEL-local-harness	Routes work across local models and exposes bounded tools to investigator roles
MNCS Forge	Supplies provider-neutral micro-verifiers, diagnostic probes, snapshots, and counterfactual witnesses
MNCS Fabric	Executes and reconciles content-addressed experiments across machines
RAVEL	Supplies the adaptive mechanism and consumes governed experience or candidate proposals
MNCDS	Records candidate generation, evidence eligibility, feedback boundaries, and lineage
MNCS	Evaluates frozen bounded claims; MNEL cannot issue conformance by itself

MNEL integrates with those systems through explicit records and adapters. It does not copy their authority or silently create substitute implementations.

Implemented foundation

• canonical JSON and SHA-256 identities;
• append-only hash-chained JSONL evidence ledger;
• explicit experiment lifecycle and state transitions;
• investigator, skeptic, replicator, synthesizer, and auditor role contracts;
• immutable authority and resource-policy checks;
• development, selection-observed, transfer-hidden, and future-final visibility labels;
• independent hard-gate evaluation with PASS, FAIL, and UNKNOWN;
• causal-attribution records distinct from evaluator verdicts;
• Verified Experience Distillation (VED) principle and strategy proposals;
• transfer gating, source lineage, falsifiers, counterexamples, and negative memory;
• provider-neutral adapters for the local harness, Forge, MNCS Fabric, and RAVEL;
• typed learned micro-provider declarations and diagnostic observations;
• deterministic capability matching, cost filtering, and diversity-aware selection;
• an initial 12-family architecture catalog with declared advantages and limitations;
• accepted Rust-first runtime architecture decision and versioned C ABI;
• safe Rust provider SDK, host admission policy, reusable snapshot cache, and runtime manifest validation;
• backend-neutral CPU/full-CUDA/sequential-CPU-offload policy with reserve/cap/workspace accounting, real-probe requirements, precision checks, and bounded AUTO OOM recovery;
• explicit distinction between persistent provider lifetime and physical weight placement;
• process-local Rust host lifecycle, snapshot reuse, output-buffer limits, normalized status handling, measurement collection, and deterministic quarantine;
• executable mnel-provider-classical HMM diagnostic provider and host integration tests;
• eligible-context packing, read-only/proposal workspace models, identity envelopes, candidate transactions, quarantine queues, and deterministic morning-report records;
• bounded local-harness JSON-line execution with timeout/output ceilings, authority rejection, deterministic observations, and detached Git proposal worktrees;
• SHA-256 artifact-bound Rust dynamic loading for mnel_provider_entry_v1, descriptor and pointer/length validation, host-owned output copying, clean unload, and quarantine integration through the existing provider host;
• initial immutable transition, tabular, and pair diagnostic snapshot producers with compact binary payloads and dependency-bound content identities;
• bounded transition, pair, tabular, trace, graph, and composite snapshot views backed by one identity-keyed immutable snapshot store;
• explicit diagnostic verifier declarations and registry matching, bounded preconditions, proposal-bound probe requests, diagnostic-only witnesses, and deterministic reference transition/tabular/pair/trace/graph verifiers;
• registered-only mutation operators, independent witness comparison, verifier health and quarantine state, coverage records, learned-provider observation normalization, and proposal-only omitted-question candidates;
• deterministic reference workflow, JSON schemas, mutation-oriented tests, and CI.
• bounded skeptic-driven omitted-question discovery with evidence lineage, duplicate suppression, coverage/health/disagreement gaps, and proposal-only authority;
• source-preserving distillation groups, provisional principles/strategies, transfer prediction freezing, negative-memory-aware retrieval, deterministic study arms and ablations, and retrieval/calibration metrics;
• 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.
• 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

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

MNEL currently requires Python 3.11 or newer and has no Python runtime dependencies outside the standard library. Building the provider runtime contracts additionally requires Rust 1.79 or newer.

Quick start

Inspect the environment and built-in role contracts:

mnel doctor
mnel investigator list

Inspect the learned micro-provider catalog:

mnel learned-provider list
mnel learned-provider describe latent.transition-jepa
mnel learned-provider match \
  --uncertainty unexpected-transition \
  --artifact candidate-transition \
  --snapshot transition-feature-snapshot \
  --max-cost low

Request a heterogeneous panel rather than the highest-scoring correlated providers:

mnel learned-provider match \
  --uncertainty temporal-anomaly \
  --uncertainty unexpected-transition \
  --uncertainty metric-inconsistency \
  --diverse \
  --limit 4

The catalog commands do not download or train models. The Rust reference provider can be measured locally with:

cargo run -p mnel-provider-host --example provider-benchmark

The benchmark reports one local observation of host admission and warm invocation timing; hardware-independent placement policy tests use fake accelerator diagnostics.

Run the deterministic reference lifecycle:

mnel demo --workspace build/demo

The demo preregisters a bounded experiment, records an observation, evaluates hard gates, attributes the intervention, and creates a provisional principle proposal. It does not call a model or modify RAVEL.

Run the bounded 0.3 diagnostic lifecycle, including snapshot production, two independent reference witnesses, a mutation, comparison, health/coverage records, and a next-question candidate:

mnel forge-reference --workspace build/forge-reference

Run the deterministic 0.4 distillation study:

mnel distill-reference --workspace build/distill-reference
mnel ledger verify build/distill-reference/distill-evidence.jsonl

The study preserves development records, prevents development code from reading hidden or future-final records, freezes the transfer prediction before held-out evaluation, 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:

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.

Run the local distributed reference path:

mnel fabric-reference --workspace build/fabric-reference
mnel ledger verify build/fabric-reference/fabric-evidence.jsonl

Verify and summarize the resulting ledger:

mnel ledger verify build/demo/evidence.jsonl
mnel ledger summarize build/demo/evidence.jsonl

Experiment lifecycle

draft
  -> preregistered
  -> running
  -> observed
  -> evaluated
  -> attributed
  -> distilled proposal

Any stage may terminate as rejected or UNKNOWN. A successful task metric alone does
not establish causal understanding, transfer, conformance, independence, or promotion.

Each material record binds its source identities, visibility, experiment identity, parent or candidate identity where applicable, and the authority that produced it.

Learned micro-providers

The registry starts with transition and graph JEPA variants, a message-passing GNN, selective state-space and temporal convolution sequence models, a contrastive Siamese encoder, Deep SVDD, a denoising autoencoder, a tiny transformer, gradient-boosted trees, a hidden Markov model, and reservoir computing.

The portfolio is intentionally heterogeneous. Each architecture has a declared input view, purpose, objective, cost, size range, strengths, and limitations. Learned outputs remain diagnostic-only, carry verdict_semantics: not-a-verdict, and can only help an investigator decide which bounded Forge question to ask next.

See Learned micro-provider registry.

Provider runtime implementation policy

Rust is the reference and default production language for the persistent provider host, provider SDK, dispatch, budget enforcement, snapshot reuse, and CPU-first provider implementations. Python remains the training, calibration, experimentation, export, and high-level orchestration language.

The stable cross-language boundary is mnel-provider-c-abi/1. Native-trusted providers must be Rust unless an identified benchmark and threat review justify a specialized non-Rust implementation. WASM is reserved as a quarantine and portability tier. Admitted providers are persistent, weight-resident, and consume identity-bound compact binary snapshot views; process startup and JSON parsing are not part of the normal hot path.

See ADR 0001 and the learned-provider runtime contract.

Placement and residency

Provider admission is persistent, but GPU residency is a separate policy decision. resident-on-admission means the provider is loaded and reusable; it does not mean all weights must remain permanently on a GPU. With sequential CPU offload, weights remain in system RAM while individual modules execute temporarily on CUDA, trading VRAM for host memory and transfer overhead. Explicit CPU/CUDA/offload choices are honored or rejected; only auto may use bounded full-CUDA → sequential-offload → CPU recovery.

Investigator roles

• Investigator — proposes falsifiable hypotheses and bounded interventions.
• Skeptic — searches for alternative explanations, omitted assumptions, and verifier gaps.
• Replicator — repeats frozen experiments across seeds, nodes, or providers.
• Synthesizer — proposes principles and reusable strategies from eligible attributions.
• Auditor — checks lineage, contamination, budgets, identity, and authority boundaries.

These are contracts, not trusted personalities. A role label does not grant execution or acceptance authority.

Verified Experience Distillation

VED is the consolidation stage inside MNEL:

verified episodes
  -> evaluator-derived effects
  -> causal attribution
  -> provisional principle
  -> transfer tests
  -> supported strategy
  -> separately evaluated RAVEL candidate proposal

Distillation never deletes the raw evidence. A compact lesson retains references to supporting attributions, known counterexamples, falsifiers, declared scope, transfer state, and failure modes.

Repository map

src/mnel/                          Python control plane and executable foundation
crates/mnel-provider-api/          versioned provider ABI vocabulary
crates/mnel-provider-sdk/          safe Rust provider authoring surface
crates/mnel-provider-host/         persistent process-local host, policy, and snapshots
crates/mnel-provider-classical/    executable deterministic HMM diagnostic baseline
include/                           language-neutral provider ABI header
schemas/                           machine-readable record and runtime vocabulary
docs/                              architecture, decisions, method, boundaries, roadmap
examples/reference-study/          deterministic lifecycle example
examples/learned-providers/        architecture catalog and runtime manifest example
tests/                             lifecycle, integrity, registry, runtime, negative tests
.github/workflows/                 Python and Rust continuous verification

Run the checks

python -m compileall -q src tests
python -m unittest discover -s tests -v
python -m mnel learned-provider list
python -m mnel demo --workspace build/demo
python -m mnel ledger verify build/demo/evidence.jsonl
cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace

Safety and claim boundary

The current command adapters use argument arrays with shell=False, but MNEL is not an operating-system sandbox. Untrusted experiment execution belongs in a hardened runner with network restrictions, resource controls, immutable verifiers, and disposable workspaces.

The provider runtime includes a small native-trusted dynamic loader with explicit ABI, artifact, pointer/length, output, and diagnostic-authority checks. It is not an operating-system sandbox: malformed native code can still crash the host process, so untrusted experiment execution belongs in a hardened runner with network restrictions, resource controls, immutable verifiers, and disposable workspaces.

A local MNEL result or learned-provider observation can describe bounded development context. It cannot by itself establish independent evaluation, protected custody, real-world safety, general recursive self-improvement, formal MNCS/MNCDS status, certification, or promotion.

See Architecture, Learning model, Learned micro-providers, learned-provider runtime, Threat model, and Roadmap.