Dataset studio
Trace original scenarios to their evidence. Separate calibration, training and held-out evaluation by task family.
Local pipelineDOMAIN MODELS · ENGINEERING ARCHITECTURE
Engineering intelligence.
Built on evidence.
Turn architecture scenarios and expert judgment into better small language models. A transparent path from Qwen baseline to supervised tuning, preference optimization and verifiable reinforcement learning.
Development preview · local review tools available · managed training and deployment planned
NEXT RUN / MODEL DEVELOPMENT
01 / DATAArchitecture scenariosConstraints · evidence · failure modes
↓
02 / JUDGMENTExpert feedbackRubric scores → adjudicated preferences
↓
03 / LEARNQwen + LoRA
04 / VERIFYRelease gates
SFT → DPO → optional online RL ↺ Evaluate again
Requirements & trade-offsReliability & recoveryData & tenant boundariesSafe agent execution
ONE DEVELOPMENT LIFECYCLE
Make the decisions visible.
Make the improvement measurable.
Architecture quality is more than fluent prose. Evaluate constraints, failure handling and evidence before turning a response into a training target.
Human evaluation
Compare blinded responses, score anchored rubrics and resolve disagreements before exporting preferences.
Open browser demo →Post-training
Choose the objective that fits the evidence: approved answers for SFT, comparisons for DPO, tested rewards for RL.
GPU integration plannedEvaluation & release
Compare against a frozen baseline. Inspect regressions, critical failures and serving costs before promotion.
Release designEXPLORE THE PIPELINE
From raw context to a tested policy.
Select a stage to see its inputs, outputs and readiness gate.
PURPOSE-BUILT TASKS
Train for the hard architecture decisions.
Reason through failure.
Retry-safe payments, regional failover, queue backlogs and recovery tests. Reward correct constraints and executable validation.
Protect the boundary.
Tenant isolation, permission-aware retrieval, schema migration and event ordering. Track critical violations separately from average quality.
Connect design to evidence.
Target applications include Avcel architecture assistance and ProofFacet evidence assessment. Product integrations remain under development.
REFERENCE ARCHITECTURE
The harness runs the agent.
The trainer updates the model.
Start with direct Qwen generation and human feedback. Add an agent harness when the task actually requires tools, state and multi-step execution.
Versioned task registry
Approved sources → original scenarios → blinded candidates → rubric review → adjudication → immutable splits
Source rights, evidence IDs, reviewer provenanceQwen + PEFT + TRL
SFT teaches approved behavior. DPO learns preferences. GRPO is an optional on-policy stage for tasks with validated rewards.
Model revision, adapter, seed, dataset hashStrands harness
Tool execution, sessions and traces for an architecture agent. It is an integration candidate, not an RL optimizer or an installed dependency here.
Explicit Qwen provider · sandbox · tool allowlistPromotion by evidence
Frozen tests, human preference, executable checks, latency and cost. Promote a versioned artifact; keep the previous model available for rollback.
Separate inference endpoint and release registryWhen does Strands belong in the RL loop?
Use it for tool-using rollouts after the environment has reset semantics, bounded execution and reproducible observations. A training adapter must preserve policy tokens, action masks, rewards and policy-version identity. Ordinary agent traces are useful for evaluation, but are not automatically valid on-policy training trajectories. Offline SFT and DPO do not require a harness.
Disable default tools and remote model defaults until explicitly configured. Verify Qwen tool calling and trace capture against a pinned harness version.
EXPERIMENT DESIGNER
Choose the next justified step.
Build a downloadable run specification. This planner does not launch compute or upload data.
Download Python training starter ↓
Dataset validator, SFT/DPO/GRPO scaffold, narrow reward tests and setup guide. GPU execution has not been validated.
TRANSPARENT BY DESIGN
What is available today.
Available Browser-local review, rubric scoring, adjudication and demo exports. 24 actual Qwen3-4B baseline outputs, local inference, and a downloadable Python training scaffold.
In development Expanded architecture datasets, GPU validation, shared reviewer service and experiment tracking.
Not claimed A production-trained architecture model, measured architecture quality gains, a deployed reward model, or a managed training service.
The earlier Qwen2.5-0.5B action-selection experiment is a mechanics check. Its accuracy is not an architecture benchmark. The public workbench has no identity verification, shared review backend or live model endpoint.
RESEARCH & IMPLEMENTATION
Built from clear technical choices.
Primary documentation behind the reference workflow. These links describe third-party capabilities, not integrations already running in NextRun.