THE PROBLEM
How can an insight become a useful method without losing uncertainty, context and responsibility?
A method can spread faster than the explanation behind it. What changed, what evidence supports it and who decides whether it may be used? We want to understand how a challenge can affect the next step.
Fictional scenario
What if a challenge became the beginning of a better method?
Two fictional municipalities use different distance thresholds in guidelines referring to the same legal provision. An AI assistant proposes correcting one guideline. A reviewer objects: local circumstances have not been compared. The question becomes an intertween and seeds a method for inspecting context before judging differences. When the method is applied, an exception raises another question.
CONNECTIONS TO EXPLORE
What does the method build on, and what has actually been tested?
A fictional miniature example. Choose a connection and read its evidence. All texts and events below are illustrative.
EVIDENCE IN THE FICTIONAL SOURCE
Method M · version 1
Method M develops from question I.
Read the complete miniature record and connected sources
Question I · an intertween
How can we compare guidelines when local circumstances are missing?
Illustrative record · manniska-ai-larande:I · v1Method M · version 1
Method M develops from question I. Missing context must be stated before comparison.
Illustrative record · manniska-ai-larande:M · v1Evaluation plan E
Plan E intends to test method M, version 1. No results are available yet.
Illustrative record · manniska-ai-larande:E · v1LET’S EXPLORE
What we want to explore with the framework
We explore how an intertween can seed a skill and how that skill can be represented as its own Knowledge Object. The method links to its evidence; the original question and alternative explanations remain available.
Possible source material
Intentions, observations, intertweens, candidate skills, evaluations, decisions and outcomes.
An illustrative example
A candidate skill checks the shared source, local circumstances, versions and missing information before treating different guidelines as contradictory. A reviewer also evaluates cases where context is unknown. The evaluation and the decision to use the method are separate objects.
What we need to learn
Which local conditions, responsibilities and human judgments are missing from the example? A real case may change both the problem framing and the proposed response.
What should remain useful, and on whose terms?
Can a method and its contribution history remain useful after a colleague leaves, without becoming a permanent personal profile? How do we separate permission to use the method from permission to train a model, and give contributors a say in recognition?
Questions to test. The examples do not establish that these protections or decision rules are implemented. Read about responsibility and conditions ↗
An insight can become a method
An intertween can seed a skill. The method gets its own Knowledge Object with purpose, instructions, prerequisites, evidence and conditions of use. The original question is not replaced. Not every intertween becomes a skill; several methods may develop from one insight.
Observation → intertween → candidate skill → evaluation → adoption decision → outcome → new questions.
This is a branching network, not a mandatory sequence. A skill description does not prove capability, and an evaluation does not automatically authorise use.
Between human intention and machine activity
We explore how the protocol could make interpretation traceable in both directions. Human goals and conditions need to become usable instructions. Machine results need to become evidence people can understand and challenge.
The intention
Reduce handling time while preserving explanation quality.
An incomplete interpretation
Minimise response time.
The difference needs to be visible before the method is used. An interpretation record would link the intention to the instruction, assumptions and review. A summary likewise needs references to its evidence and material omissions.
Follow how a challenge changes the method
Fictional, scripted records. This example runs no AI and enforces no permissions.
The assistant proposes flagging different thresholds as errors. This is version 1 of a candidate, not a confirmed conclusion or an authorised method.
Two ways to improve an AI system
Learning around a model
Instructions, tools, context selection and workflows can change without changing model weights. The skill object describes the method and the version evaluated. This can work with both open-weight and closed-weight models.
Learning through model adaptation
Permitted material and experience may inform training or adaptation. The protocol could connect sources and transformations to the resulting model version and its evaluation, to the extent those records are available.
Open weights do not automatically provide full visibility. A closed service may offer some adaptation. Actual access, licensing and available evidence determine what can be changed and inspected. Skill descriptions and history can travel to a replacement model, but performance must be re-evaluated.
When learning changes the next learning cycle
A task skill improves an activity. An improvement skill changes how the next method is proposed, tested or selected. This is where recursive self-improvement, RSI, becomes relevant. We explore how changes to the evaluation process itself can have an explicit history and separate review.
The protocol is proposed as a way to carry evidence and responsibility through these steps. It does not itself establish that improvement is real, unlimited or safe.
Questions that remain
- What evidence is sufficient to call a method reusable?
- How can a short summary preserve disagreement and uncertainty?
- Which changes require renewed authorisation?
- What can be verified when parts of a system cannot be inspected?
- How does a challenge lead to actual reassessment?
Test the question together
Do you recognise the problem, or see it differently? Bring a real case, source material or a challenge. Together, we can examine the assumptions and decide whether a focused demo or pilot is the right next step. We would test what becomes easier, what still needs human judgment and what Intertween needs to change.
Start a conversation about this path ↗