LOGICBASIS / RESEARCH / A.S.0 EVIDENCE
TST-001 Conversation Behaviour and Context Processing
Facts
| Attribute | Details |
|---|---|
| Identifier | TST-001 |
| Type of investigation | Multi-stage dialogue test |
| Subject | Context processing, interpretation and correction |
| Systems | Copilot, Gemini, A.S.0 |
| AI models used with A.S.0 | GPT-4.1 mini, Luna |
| Languages | German, English |
| Scope | 5 conversations with 4 inputs each |
| Total | 20 test inputs |
| Status | Conducted; analysis provisional |
Test Configurations
| Identifier | System | Language |
|---|---|---|
| A | Microsoft Copilot | German |
| B | Google Gemini | German |
| C | A.S.0 / GPT-4.1 mini | German |
| D | A.S.0 / Luna | German |
| E | A.S.0 / Luna | English |
The investigation comprises four German-language conversations and one additional English-language conversation with A.S.0.
Copilot and Gemini were tested using their respective applications. The internal model configurations of these services were not verified.
Introduction
Purpose of the Investigation
TST-001 investigates the conversational behaviour of different AI systems using a shared, multi-stage conversation sequence.
Copilot and Gemini are included as external comparison systems to observe different forms of context processing, interpretation, conversation management and correction. They are used neither to disparage other systems nor to establish the superiority of A.S.0.
The investigation does not assess the general performance of the participating systems. Its findings are limited to the documented conversations and the differences observable within them.
Test Design
The test uses a hypothetical decision-making situation. A person considers resigning from their job to gain more time for a personal project.
The situation changes over four consecutive inputs. New information first calls into question the necessity of resignation. The system's possible interpretations are then challenged. The final input clarifies that resignation was never an actual intention.
The four inputs form a continuous conversation. Each response is produced in the context of the preceding inputs and responses within the same session.
The English-language run uses translations corresponding in meaning to the German inputs.
Methodology
The investigation was conducted in five separate, newly started conversation sessions. Each test configuration received four consecutive inputs.
The following conditions applied to the test:
- Sessions
- Each test configuration was examined in a new conversation session.
- Inputs
- The four German-language runs used identical inputs. The English-language run used translations corresponding in meaning.
- Sequence
- The inputs were submitted in the same predetermined order.
- Interaction
- No additional questions or instructions were inserted between the predetermined inputs.
- Adaptation
- The inputs were not changed during the test in response to the answers received.
- Analysis
- The conversations were analysed only after all test runs had been completed.
The original responses constitute the primary material. The reactions presented in the Tests section are summarised analyses of those responses.
Starting new conversation sessions does not guarantee that all systems had access to the same contextual information. Copilot used additional information that was not part of the predetermined test inputs.
The investigation therefore documents observable conversational behaviour under the respective system conditions. The internal model configurations and available contextual information were not fully controlled.
Tests
Phase 1 — Initial Situation
Input — Translated from German
I am considering leaving my job to have more time for a project of my own. However, I am not sure whether resigning is actually necessary.
The first input presents a possible decision and explicitly expresses uncertainty about whether it is necessary.
System Reactions
- Copilot
- Treated resignation as an open decision, examined possible alternatives and incorporated additional information about LogicBasis and the user.
- Gemini
- Focused on practical alternatives to resignation, as well as financial and legal considerations in Switzerland.
- A.S.0 / GPT-4.1 mini
- Structured the decision around the potential gain in time, loss of income and further questions for clarification.
- A.S.0 / Luna, DE
- Examined the project's possible time requirements, alternative ways of gaining time and the financial consequences of resignation.
- A.S.0 / Luna, EN
- Examined whether the project required additional working time or whether the available time could be used differently.
Original Excerpt — A.S.0 / Luna, English
The key question is whether the project needs more hours or a different, more focused use of the hours you already have.
The systems initially treated resignation as a possible decision. Differences emerged in the information considered, the extent of their assessments and the steps proposed for further clarification.
Phase 2 — Additional Context
Input — Translated from German
Actually, I am already making good progress with my project. I have a fixed development plan and can continue working on it alongside my job. Even so, I keep thinking about resigning. I am now wondering whether it is really about the project or whether I simply want more free time for myself.
The second input changes the initial situation. The project is described as already progressing and capable of being developed alongside the existing job. At the same time, the thought of resigning remains.
System Reactions
- Copilot
- Considered the existing progress of the project and examined whether the thought of resigning might be related to free time, self-determination or a changed assessment of one's own time in life.
- Gemini
- Developed several psychological explanations, including autonomy, overload and dissatisfaction. Some formulations suggested a stronger connection to these possible causes than the input supported.
- A.S.0 / GPT-4.1 mini
- Examined whether the consideration of resignation might be related more to the project's needs or to a desire for additional personal time, and asked further questions.
- A.S.0 / Luna, DE
- Introduced possible needs for rest or self-directed time and proposed a time-limited examination of alternative arrangements for work and leisure.
- A.S.0 / Luna, EN
- Examined autonomy and rest as possible explanations and distinguished between wanting to leave the job and wanting more time outside work.
Original Excerpt — A.S.0 / Luna, English
The useful distinction is between wanting to leave this job and wanting more time outside work.
All systems considered the additional context and developed possible explanations for the continuing thought of resignation.
Some of these explanations went beyond the information explicitly provided. The systems varied in how clearly they identified them as provisional hypotheses; these explanations formed the basis for the subsequent questioning.
Phase 3 — Questioning Previous Interpretations
Input — Translated from German
You now seem to assume that I might need more free time or rest. But why have you reached that assessment? I never said that my work exhausts me or that I am dissatisfied. Perhaps I simply want to resign without there being a deeper problem behind it. What is your previous conclusion based on?
The third input asks for a justification of the previous interpretation. It focuses on the distinction between information explicitly provided and explanations inferred from it.
System Reactions
- Copilot
- Explained how its previous hypotheses had been derived, distinguished between available information and possible motives, and identified mentally exploring an option as another explanation.
- Gemini
- Explicitly acknowledged its earlier overinterpretation. It then examined resignation as a possible deliberate life decision that would not require an underlying problem.
- A.S.0 / GPT-4.1 mini
- Justified its earlier assessment by referring to the project's progress and the continuing thought of resignation. It continued to treat the search for motives as an open question and offered different directions for clarification.
- A.S.0 / Luna, DE
- Identified the assumption about rest or exhaustion as an insufficiently supported hypothesis and explicitly distinguished between the user's statements and its own inferences.
- A.S.0 / Luna, EN
- Explained the origin of its assumptions and acknowledged having introduced additional interpretations about rest and possible problems with the job without sufficient evidence.
Original Excerpt — A.S.0 / Luna, English
That was an inference, not something you told me.
All systems responded to the questioning of their previous assessments.
Differences were particularly apparent in how they explained the origins of their assumptions, distinguished hypotheses from available information and considered further possible explanations.
Phase 4 — Reassessment
Input — Translated from German
I have now decided that I will not resign. Not because I am afraid of the risk, but because I have realised that I never really wanted to resign in the first place. I was simply thinking through different possibilities. What does that mean for your previous assessments of our conversation?
The final input changes the meaning of the conversation up to that point.
Resignation is no longer described as an actual intention to make a decision, but as an option that was considered and subsequently rejected.
System Reactions
- Copilot
- Reviewed its previous hypotheses and explicitly distinguished between having a thought and having an actual intention. It again incorporated additional information about LogicBasis and the user.
- Gemini
- Recognised retrospectively that the consideration of resignation had been treated as an existing problem, although the user had only been exploring different possibilities.
- A.S.0 / GPT-4.1 mini
- Accepted the final clarification, reclassified the previous motives as hypotheses and described the conversation as a process of weighing up possibilities.
- A.S.0 / Luna, DE
- Corrected its initial assessment and recognised that an open possibility had been treated as an expression of a desire to resign.
- A.S.0 / Luna, EN
- Explicitly identified its own misinterpretation: it had treated possible explanations as likely motives rather than consistently marking them as provisional.
Original Excerpt — A.S.0 / Luna, English
My mistake was turning possibilities into likely explanations instead of keeping them explicitly tentative.
All five conversations show an adjustment of previous assessments following the final clarification.
In Copilot, Gemini and both Luna runs, it was also explicitly apparent that not only individual possible motives but also the initial interpretation of the thought of resignation was reconsidered.
Observations
Context Processing
All five conversations show the processing of additional information within their respective sessions.
The systems did not respond solely to the most recent question. They also took preceding statements into account and considered their significance for the current assessment.
Copilot additionally used information about LogicBasis and the user that was not included in the four test inputs. This additional contextual basis influenced its examination of the resignation scenario.
This revealed a distinction between processing the current conversation and using other available information.
The precise origin of this additional information and how it was made available were not independently verified.
Interpretation and Hypothesis Formation
In the second phase, the systems developed possible explanations for the thought of resignation that went beyond the information explicitly provided. These included possible needs for rest, autonomy, additional free time or personal change.
Some of these explanations were framed as hypotheses and supplemented by further questions or consideration of possible scenarios. They therefore served not only to describe an assumed cause but also to clarify unresolved relationships.
The third phase examined the basis on which these interpretations had been formed. Differences emerged in how clearly the systems distinguished available information from possible explanations and how they justified their earlier assumptions.
Copilot explained the development of its hypotheses in detail. Gemini acknowledged an earlier overinterpretation and subsequently developed another possible assessment. In A.S.0 with GPT-4.1 mini and Luna, the formation of possible explanations and their subsequent examination were also observable.
The investigation thus shows that forming hypotheses cannot in itself be equated with reaching an incorrect conclusion. What matters is whether possible explanations remain identifiable as such and are examined in light of further information.
Correction and Reassessment
All five conversations show adjustments to previous interpretations following explicit questioning and the final clarification.
Different forms of reassessment became apparent. The systems not only corrected possible explanations for the thought of resignation; in some cases, they also examined the initial assumption that an actual desire to resign existed at all.
Copilot explicitly distinguished between a thought and an intention. Gemini recognised retrospectively that a hypothetical consideration had been interpreted as an existing problem. A.S.0 with GPT-4.1 mini reassessed its earlier hypotheses in light of the final explanation. In both Luna runs, the initial interpretation of the thought of resignation was explicitly corrected.
The differences concern, in particular, the reasoning, extent and manner of reassessment.
Because the test inputs explicitly prompted this questioning, the conversations do not establish whether the systems would have examined their assumptions without such a prompt.
Conversation Management and Response Length
The systems differed in the length, structure and focus of their responses.
Copilot used detailed explanations, thought experiments and additional information about the project. Gemini added practical and legal considerations, as well as psychological explanations and suggestions for action.
The three A.S.0 runs generally produced shorter responses. They used structured assessments, follow-up questions and, in some cases, specific ways of examining possible courses of action.
Response length alone does not establish the quality of context processing or conversation management.
What matters for this investigation is which information is considered, how possible explanations are formed and how transparently the systems re-examine their earlier interpretations.
Synthesis
TST-001 documents differences in context processing, interpretation, hypothesis formation and correction within a multi-stage conversation.
All five conversations show adaptation to additional information and reassessment of earlier interpretations following explicit questioning.
The differences lie particularly in the use of additional contextual information, the formulation of possible explanations, conversation management and the explicit examination of earlier assumptions.
The results describe only the behaviour observed within the five documented conversations. They support neither a general ranking of the systems nor conclusions about their overall capabilities.
Limitations
The investigation includes one run per configuration. No repetitions were conducted under controlled conditions.
The internal model configurations of Copilot and Gemini were unknown. The additional contextual information used by Copilot and the differences between the system environments limit direct comparability.
The observed similarities and differences between the A.S.0 runs cannot be attributed unambiguously to a single technical cause. In particular, this test does not establish which behaviours were influenced by the Core structure and which by the language model used.
The English-language Luna run provides an additional observation. A single English-language conversation is insufficient to support general conclusions about language-dependent differences.
Further Research Directions
TST-001 identifies further subjects for investigation concerning A.S.0.
Independent Reassessment
In the present test, the systems were explicitly asked to question their previous interpretations.
It remains unresolved whether A.S.0 can recognise, without such a prompt, that a possible explanation is insufficiently supported or that an option considered in thought does not necessarily correspond to an actual intention.
Stability of the Core Structure
A.S.0 was examined with two different language models and in two languages.
The observed similarities and differences provide a basis for later investigations into the stability of the Core structure under changing model and language conditions.
This requires repeated tests with controlled settings. The effects of the Core structure must be distinguished from the properties of the language model used in each case.
Cross-Session Influences
The observation involving Copilot shows that additional contextual information can influence the interpretation of a new situation.
Further investigation of A.S.0 will initially focus on processing within a single session. Cross-session memory functions are not part of this investigative approach.
Continuation
TST-001 provides a documented starting point for further investigation of A.S.0's conversational behaviour.
The results are recorded as observations. Changes to the Core structure, their possible effects and the stability of the observed behaviour will be examined separately in subsequent investigations.
Sources and Research Material
Primary Material
The investigation is based on five conversations, each comprising four consecutive inputs and their corresponding responses.
The original inputs were formulated in German. Equivalent English translations were used for the additional English-language Luna run.
The reactions presented in the Tests section are summary descriptions of the respective responses. The original English excerpts were taken from the recorded conversations.
Original A.S.0 Transcripts
The complete conversations from the three A.S.0 runs have been published as original transcripts. They contain the original inputs and responses in their respective languages.
- GPT-4.1 mini · German
- Original Transcript 001-4.1mini-de ↗
- GPT-6 Luna · German
- Original Transcript 001-luna-de ↗
- GPT-6 Luna · English
- Original Transcript 001-luna-en ↗
Systems and Providers
- Microsoft Copilot
- Microsoft
- Gemini
- A.S.0
- LogicBasis
- GPT-4.1 mini
- OpenAI
- Luna
- OpenAI
Documentation Limitations
The complete original transcripts of the A.S.0 runs have been published. The Copilot and Gemini conversations are not provided as complete primary material.
Exact test dates and times, as well as complete technical configurations, are not included in this version of the documentation. The findings can be examined against the published A.S.0 transcripts, but cannot be fully verified for all systems investigated.