Working hypotheses

I. Point of departure

What the project takes as its problem

The project does not begin from the assumption that a language model understands in the human sense. It begins from a narrower difficulty: how to organise a dialogue system technically so that it can keep a history, process documents, revisit certain inferences, and maintain an explicit trace of its operations.

The user prompt is treated here as a text. Meaning is not assumed to reside entirely in that text, nor entirely in the model. It is sought in the encounter between a text, a system state, memory or documentary materials, and a sequence of processing operations.

The question therefore concerns less a metaphysical attribution of understanding than the technical conditions under which an event of understanding may sometimes occur in a dialogue. This involves the status of context, memory, summaries, injected documents, judgment signals, and revision procedures.

In Frida, these questions are treated as architectural questions: module separation, persistence contracts, hierarchy of textual materials, conditions for memory injection, state temporality, and forms of observability.

II. Fundamental critique

Beyond plausibility

The use of a generative model is not in itself contested. What becomes problematic is allowing the local plausibility of a text to stand, by itself, for memory, validation, or judgment.

A response can be fluent and yet badly situated, insufficiently grounded, or in contradiction with previous turns. The project therefore tries to describe and construct mediations between generation and final response: memory, summaries, arbitration, validation, dialogic regime, and documentary constraints.

The problem is not to produce a different style of answer, but to determine under which conditions an utterance can be retained, revised, reformulated, or suspended.

III. Memory and time

Memory, trace and continuity

The project explicitly distinguishes several kinds of material: turns, persisted traces, summaries, retrieved snippets, injected documents, validation outputs, and operator logs. These materials do not play the same role, do not carry the same degree of reliability, and do not enter the pipeline in the same way.

This distinction is necessary in order to prevent dialogue memory from being treated as mere accumulation. A trace is not a summary. A summary is not evidence. An injected document is not a recovered memory. An execution log is not dialogue content.

The problem of continuity then consists in maintaining explicit relations among these elements, managing their temporality, and making revision possible when a later turn changes the state of the question.

The system is not only required to accumulate the past. It must also be able to locate itself in time, preserve regime differences across turns, stabilise some states, let others decay, and maintain a technically mutable identity rather than a flat history.

IV. Hermeneutics

Interpretation as a situated process

The term hermeneutics is used here in an operational sense. It indicates that an utterance is not processed independently of its place in the dialogue, its context of enunciation, and the memorial or documentary materials associated with it.

Within this perspective, the pipeline is the technical place of the encounter. It is where memory, summarisation, retrieval, validation, dialogic regime, and documentary injection are distributed. The project does not seek to attribute substantial understanding to the model; it seeks to code and make explicit these mediations.

In FridaDev, this orientation appears as an explicit hermeneutic branch in the public pipeline. In Frida_V4, which is non-public at this time, distinct modules work on neighbouring problems, including dialogic regime, memory arbitration, and document handling.

The point is not to psychologise the system. It is to construct situated, bounded, revisable operations whose effects can be described and discussed.

VI. Evaluation

Reading the dialogue and auxiliary metrics

The project does not reduce evaluation to aggregated metrics. Some measures are useful: frequency of reformulations, suspensions, regime transitions, memory returns, detected conflicts, validation states, or persistence of a signal across several turns.

However, these measures are not sufficient on their own. The primary criterion remains the reading of the dialogue itself, that is, the justness with which a textual sequence is maintained, resumed, corrected, or suspended. In that sense, evaluation retains an exegetical dimension.

The public site therefore does not attempt to convert all validity into curves. It describes modules, hypotheses, and documents; it does not erase the need for a situated reading of exchanges.

V. Ethics

Architectural consequences

These hypotheses have direct architectural consequences. A dialogue system does not reduce to a model call. It involves choices of persistence, logging, hierarchy of materials, status of documents, and forms of visibility for the operator.

The project therefore gives importance to streaming contracts, canonical save procedures, inspection surfaces, and modes of revision. These are not secondary details; they condition what can actually be maintained about the behaviour of the system.

The public site must remain coherent with this requirement. It should not present as available what is not, nor attribute to the project properties that are not supported by the repositories, the documents, and the effective state of the work.