Venode / Research practice
We investigate compact language models, bounded agent execution and the evaluation methods required to move AI into accountable operational use.
Enter the observatoryResearch position
Capability is necessary. It is not sufficient. Useful AI must also be efficient enough to run, bounded enough to govern and legible when something goes wrong.
Model research, agent infrastructure and operational evaluation therefore sit in one practice. Each discipline changes what the others need to prove.
Research is not a product claim. Training is not availability. A controlled pilot is not proof at scale.Research observatory / Status July 2026
The lab is organised as connected research lines, not isolated product claims. Select a programme to inspect its present scope and maturity.
A compact language-model research programme focused on code comprehension, debugging and repository-scale software work. Dataset construction, training behaviour and evaluation methods for specialist code tasks.
A larger-scale model research programme for multi-step reasoning, planning, tool use and agentic workloads. Reasoning evaluation, tool-use behaviour and the controls required for longer task sequences.
A platform for coordinating agent runtimes, shared capabilities, state and verification across tools and operating environments. A framework for turning goals into constrained, testable task contracts for dependable agent execution and handoffs.
Applied evaluation connects lab work to operating conditions. We study where a system is useful, where control must remain human, which failures matter and what evidence is required before wider deployment.
Status is a July 2026 working record. It does not imply public availability or validated performance.
Evidence discipline / 03
A result becomes useful only when the task, conditions, baseline, failures and decision remain attached to it.
Define the work, constraint and decision the research must support.
Record the data, tools, compute and human oversight behind the result.
Measure current performance, cost, failure modes and intervention.
Test where performance breaks and which controls continue to hold.
State what the evidence supports, what it does not and what happens next.
This page does not present benchmark or reliability results. Future results should include the task definition, method and limitations needed to interpret them.
Open lines / 04
Current lines of inquiry—not papers, benchmarks or availability announcements.
Mara / Model systems
Mesh + Netrima / Agent reliability
Applied research / Operations
Publication register / 05
No paper is presented as published, peer reviewed or available as a preprint yet.
Venode will distinguish clearly between a published paper, preprint, technical report, methods note and work in progress. A record will not display authorship, venue, DOI or review status until each is confirmed.
Request publication updatesMara / Model systems
Mesh + Netrima / Agent reliability
Applied research / Operations
Venode AI policy / Working draft 0.1 / July 2026
Our mission is to research and build useful AI systems without obscuring who is accountable, what evidence supports their use or where their limits begin.
Policy governance
This record remains a working draft. It does not take effect until Venode completes formal review and approval.
Revision history
Initial public working draft and lifecycle trace.
Policy trace / Interactive working record
Select a principle to see the lifecycle practices and external reference inputs connected to it. The mapping describes design intent; it is not a certification or compliance assessment.
Why this relationship matters
Evidence discipline connects a claim to its task, conditions, baseline, failures and decision so another person can judge its limits.
Lifecycle practices
Define the use case, affected people and unacceptable outcomes.
Test performance, bias, security, abuse paths and material failure modes against explicit acceptance criteria.
Monitor in operation, record incidents and retain the ability to restrict, roll back or retire the system.
Reference inputs
Inclusion indicates a reference point only. It does not state that Venode is certified, independently audited or compliant with every requirement.
Evidence before claims selected. Evidence discipline connects a claim to its task, conditions, baseline, failures and decision so another person can judge its limits.
Ethical practice guidelines
Define the use case, affected people and unacceptable outcomes.
Assign an accountable owner and classify risk before implementation.
Document data provenance, quality, access, privacy and intellectual-property constraints.
Test performance, bias, security, abuse paths and material failure modes against explicit acceptance criteria.
Provide meaningful human oversight, intervention and override where consequences require it.
Disclose the role and limitations of AI and provide a path for questions, challenge or remediation.
Monitor in operation, record incidents and retain the ability to restrict, roll back or retire the system.
Reference frameworks
These are reference points, not claims that Venode is certified, independently audited or compliant with every requirement. References last reviewed July 2026.
