Execution lineage
AgentExecution, employee version, principal, and trace/correlation must close.
← All capabilities
Experience Distillation / Governed MoE Readiness
Distill real work into training assets.Make MoE training, connection, and fallback evidence-backed.
NexusClaw extracts experience from formal Agent executions, human corrections, tool trajectories, and business outcomes, producing training assets with lineage, redaction, split, and evaluation contracts. A trained local MoE then enters the digital workforce through the shared model router.
The current product governs training data, deterministic export, evaluation assets, Provider connection, probing, metering, and fallback. Weight training or LoRA jobs run in a customer or mutually approved training environment; NexusClaw does not schedule them automatically.
Product UI · learning inboxNexusClaw · Product evidence
01 / Formal Evidence
Training assets cannot be assembled from anonymous logs or polished conversations
Only terminal formal executions qualify: workspace, employee, employee version, principal, steps, tool calls, trigger, and outcome must be complete. Preview conversations, incomplete executions, and anonymous fallback are rejected.
Behavior and tools
Human actions, ReactSteps, ToolCallRecords, and writeback stay in real order.
Business outcome
Outcome events link back to execution so value is judged by results, not frequency alone.
Least disclosure
Credentials, secrets, unapproved PII, provider bodies, and internal stacks never enter export.
02 / Experience Distillation
Turn “one person does this well” into reviewable shared experience—not one more prompt
Behavior capture connects enterprise application actions, corrections, failures, and outcomes. Pattern recognition finds cross-user success paths with conservative small-sample calibration. Outputs may become SOP, knowledge, eval cases, or training trajectories.
01 · Capture
Formal actions and outcomes
02 · Recognize
Step sequences and success patterns
03 · Review
Source, confidence, and risk
04 · Materialize
SOP · knowledge · eval · trajectory
Product UI · experience signal distillationNexusClaw · Product evidence
03 / Governed Dataset
Training data becomes an unambiguous version before it leaves the system
A version records source query, split policy, example references, and manifest digest. PII, unauthorized data, missing lineage, train/holdout leakage, or open review blocks approval. The export explicitly records that no trainer job was scheduled by the platform.
Deterministic
The same source and policy produce the same digest; any change creates a new version.
Leakage-safe
Customer/case leakage groups govern train and holdout; overlap is rejected.
Replayable
Formal executions project to employee-eval-cases/v1 and reuse the candidate test lifecycle.
Retention-bound
Export bytes and receipts follow retention policy; no anonymous public download URL is created.
Product UI · evaluation and release gateNexusClaw · Product evidence
dataset-manifest.json
schemaVersion: cognitive-dataset-manifest/v1
splitPolicy: train · holdout · leakage-group
quality: redaction · lineage · review
digest: sha256 / deterministic
trainerOrJobScheduled: false
04 / External Training + Local MoE
NexusClaw governs training assets and runtime connection—it does not pretend to be the trainer
Approved manifests are consumed by a customer or mutually approved training environment for MoE, LoRA, or other offline training. The resulting service endpoint connects as a local_moe Provider; weights, trainer jobs, and GPU scheduling remain owned by the training environment.
NexusClaw data plane
- Formal execution evidence
- Redaction and lineage
- Dataset version and eval suite
- Deterministic export receipt
Customer / approved trainer
- Choose base model
- Training and hyperparameters
- GPU and weight storage
- Produce OpenAI-compatible endpoint
NexusClaw runtime plane
- Register local_moe
- Capability and health probe
- Usage, latency, and cost trace
- Cloud or other-local fallback
A task entering MoE stays inside the same governance chain
Agent task
identity · permission · context
Local MoE router
Expert A
sales tasks
Expert B
service tasks
Expert C
operations
Expert D
general
Governed output
tools · approval · audit · attribution
05 / Evaluation, Canary & Rollback
A model that knows the business better still cannot bypass permission, approval, or fallback
Local MoE is one model Provider. Employee identity, knowledge permission, tool permission, approval, guardrails, execution audit, and outcome attribution stay intact. Routing learning may emit recommendation, draft, simulation, and approval—but never auto-change production provider policy.
Quality
Compare task quality and human edit rate with frozen holdout and formal execution cases.
Health
Probe endpoint availability and latency; failure does not block later Providers.
Metering
Record providerKind, modelId, token, cost, routing source, and fallback position.
Fallback
Timeout, connection, rate-limit, or service errors try the next Provider in explicit order.
Product UI · learning and usage auditNexusClaw · Product evidence
Current capability boundary
NexusClaw currently does not include weight training, LoRA job scheduling, adapter-weight registry, or GPU cluster management; learning does not auto-write production routing. The platform provides governed training assets, deterministic export, evaluation and candidate release, plus controlled local MoE Provider connection, metering, probing, and fallback after training. Vision is denied for local_moe by default unless a future explicit capability contract adds it.
Bring formal executions, one target task, and your training environment
Together we will define exportable evidence, data-quality gates, offline-training ownership, evaluation baseline, local endpoint, and fallback policy.