THYNKR RSI · Research & Engineering

Toward AI systemsthat improvewith evidence.

Artificial intelligence should do more than generate answers. Over time, it should help us understand what works, identify what could work better, and test improvements against measurable outcomes.

THYNKR RSI — Governed Recursive Self-Improvement is our research and engineering initiative exploring that possibility. We are building on our experience in AI orchestration, real-world business software and AI governance, with an essential principle: greater autonomy must come with clear authority, verifiable evidence and meaningful human control.

Status: Research & development. THYNKR RSI is not a separately available commercial product, and an autonomous production self-improvement engine is not currently deployed.

Explore our approachResearch and engineering enquiries
  • StatusResearch & development
  • ProgrammeTHYNKR RSI
  • AvailabilityNot a commercial product
  • GovernanceGovernance & Assurance
  • Last reviewed

§01The ambition

Intelligence should improve the systems around it.

Today's AI systems can help research a question, analyse a document, plan a workflow or create an application. But an answer alone does not tell us whether a system has become more reliable, efficient or useful.

Our long-term ambition is to enable intelligent systems to learn from observed performance, propose changes, compare alternatives and carry the lessons of validated experiments into future work.

That may mean better model selection, lower processing cost, clearer reports, more reliable workflows, stronger safeguards or better use of computing resources. The measure of improvement is not how much code an AI can change, but whether a proposed change produces a real, repeatable and accountable benefit.

We are approaching this as an engineering discipline, not as a claim of unrestricted artificial autonomy.

§02How improvement should work

Observe. Propose. Test. Approve. Measure.

  1. 01 / Observe

    Understand the current system

    Collect permitted operational signals: task outcomes, quality assessments, latency, costs, reliability and policy events. Define an auditable baseline before recommending change.

  2. 02 / Propose

    Identify a possible improvement

    Use AI to explain an opportunity, identify affected components, estimate benefits and describe uncertainties, risks and required evidence.

  3. 03 / Test

    Compare in a controlled environment

    Evaluate candidate prompts, model-routing policies, workflows or code changes in isolated experiments against reproducible test cases. Measure regressions as well as gains.

  4. 04 / Approve

    Keep responsibility visible

    Route material changes through explicit human approval and established engineering controls. A system that proposes a change should not be allowed to approve its own expanded authority.

  5. 05 / Measure

    Learn from the real outcome

    After an authorised release, compare actual outcomes with the baseline, maintain an audit trail and use verified results to inform future proposals. Revert changes that do not meet their safeguards or objectives.

This describes our target improvement loop. Its stages are being researched and developed; it is not a claim that a complete RSI loop currently operates in production.

§03Foundations

Two complementary roles: Brain and WorQ.

We believe a practical self-improvement system needs both capable intelligence and independent governance.

FND-01Existing foundation

Intelligence and execution infrastructure

Brain

Brain is our internal AI orchestration foundation. Its existing router can select configured model routes and enforce operational controls. Our ongoing engineering work is extending that foundation toward more sophisticated task classification, model evaluation, structured AI jobs, cost-aware orchestration and cross-product integration.

Brain's role: perform authorised AI work and provide the evidence needed to assess its results.

FND-02Existing foundation

Governance, authority and accountability

WorQ

WorQ provides the governance foundation for controlled agent and tool activity, reviewable evidence and human approval. We intend to use its owner-level oversight capabilities as the future home for RSI observations, improvement proposals, evaluations and decisions.

WorQ's role: preserve policy boundaries, authorisation and human responsibility—independently of any one AI model.

FND-03Boundary

Business truth stays with the product

Applications retain their own business truth

Product systems such as GrowthPilot, AI PRO CRM and LexSalim retain responsibility for their own data, users and business rules. A shared intelligence layer must respect those boundaries rather than merge private customer data or acquire unrestricted operational access.

Some Brain and WorQ foundations are operational. A unified, cross-product RSI improvement engine remains planned research and engineering work.

§04What we are exploring

Research directions grounded in real systems.

  1. RD-01

    Governed model orchestration

    Choose and coordinate models according to the task, required quality, privacy constraints, reliability, latency and cost. Evaluate results rather than assuming that the largest or most expensive model is always best.

  2. RD-02

    Evidence-led evaluation

    Build repeatable tests, trusted baselines and independent checks for agent behaviour, data analysis, generated reports and workflow decisions. An improvement must demonstrate value without concealing new risks.

  3. RD-03

    Evolutionary optimisation

    Explore methods that generate and compare multiple candidate solutions—such as alternative agent strategies, prompts, algorithms and routing policies—while operating inside explicit experiment boundaries.

  4. RD-04

    Operational intelligence

    Look for opportunities to reduce failures, unnecessary model use, delays and infrastructure waste across authorised product workspaces and services.

  5. RD-05

    Safe recursive improvement

    Investigate whether an improvement engine can become better at generating and evaluating its own proposals, without acquiring the ability to rewrite its permissions, suppress independent tests or deploy material changes on its own.

§05The governance principle

More capable does not mean less accountable.

Our direction is to give AI systems room to investigate while keeping consequential authority explicit.

Defined scope

Observation and experiments operate only within approved data, systems and resource limits.

Independent checks

A proposed change is not accepted solely on the proposing model's assessment.

Human approval

Material production changes, elevated permissions and policy changes require authorised review.

Reversible change

Engineering decisions should include version history, validation and a credible rollback path.

Data restraint

Sensitive and customer information stays within appropriate product, tenant and provider boundaries.

Reviewable evidence

The basis for a recommendation, experiment or approval should be inspectable later.

Our aim is governed progress: systems that can recommend more effective ways of operating, while the people responsible for those systems retain control over consequential decisions.

Read our Responsible Technology principles Explore Governance & Assurance

§06Research roadmap

Build the evidence before expanding autonomy.

The roadmap describes direction, not release dates or guaranteed features.

StageFocusPosition
FoundationModel routing, policy enforcement, telemetry and reviewable governanceExisting foundations; continued development
ObservationCross-product baselines, quality metrics and approved operational signalsPlanned expansion
Improvement proposalsEvidence-backed recommendations, prioritisation and owner reviewPlanned
Controlled experimentationIsolated tests, comparative evaluation, reproducibility and regression analysisPlanned
Governed applicationHuman-authorised changes with normal tests, deployment controls and rollbackLonger-term objective
Recursive researchResearch whether improvement methods themselves can be safely and measurably improvedExploratory research

We will update claims of readiness only as capabilities are independently validated.

§07Research influences

Learning from the wider AI research community.

We follow work in evolutionary computation, autonomous research systems, agent evaluation and multi-model intelligence. Relevant published research from Sakana AI includes its Recursive Self-Improvement Lab, Darwin Gödel Machine, ShinkaEvolve and Fugu model orchestration.

These publications inform questions we want to investigate: how improvements are discovered, how experiments remain reproducible, how models collaborate and how apparent gains can be checked against unintended consequences.

THYNKR is not claiming a partnership, endorsement, licence or technical integration with Sakana AI. These are independent third-party research references, not components currently deployed in THYNKR RSI.

§08Questions

Questions worth asking.

Is THYNKR RSI a product I can subscribe to?

No. It is currently a research and engineering initiative. We will describe any future productised capability separately, with a clear availability status.

Can the system change production code or security policies automatically?

That is not the capability we are offering. Our proposed approach keeps consequential production changes and changes to authority subject to explicit human approval, independent checks and normal release controls.

How is this different from a conventional AI assistant?

An assistant usually responds to a request. Our RSI research asks whether systems can also identify measurable weaknesses, develop candidate improvements, test them against a baseline and incorporate validated lessons over time—all within clear operational boundaries.

Does this involve training our own foundation model?

Our current direction focuses on orchestration, evaluation, software and workflow improvement. Training a new foundation model is not a prerequisite or a committed deliverable of this initiative.

§09Work with us

Interested in accountable, continuously improving intelligence?

We welcome conversations with engineering teams, researchers and organisations interested in model orchestration, measurable AI evaluation, agent governance and controlled experimentation.

Discuss research or engineering collaboration · THYNKR SYSTEMS LTD · London, United Kingdom

Contact Thynkr Systems