Specification

1. Reference

This prototype uses notation specific to the Integral Coherence system.

2. Graphs

The prototype’s deliberation graphs are:
- graphs/knowledge.txt
- graphs/ethics.txt
- graphs/integration.txt

3. LLM 1

3.1 Input
- prompt
- graphs

3.2 Output
- columns
- scores

3.3 Guidance
Columns result from the unfolding of the prompt into dimensions, at the degree of detail required by the case.
For each observed relation, in each column, a score from 0 to 10 is assigned.

3.4 Format
{
  "columns": ["<column_1>", "<column_2>", "<column_3>"],
  "scores": [
    {
      "observed_relation": {"graph": "<graph>", "id": "<id>"},
      "values": [<p1>, <p2>, <p3>]
    }
  ]
}

4. Engine

4.1 Input
- columns
- scores
- graphs

4.2 Output
- paths

4.3 Operation
The engine operates within each graph. Processing takes place separately in each column.

4.4 Rules
4.4.1 Activation
In each column, every observed relation whose score reaches the threshold becomes active. In this implementation, the threshold is fixed at 6.

4.4.2 Start
Each path begins at the active observed relation with the highest score among those not yet traversed.

4.4.3 Tie-breaking
In case of a tie, the observed relation that has more adjacencies, within the same graph, to other active observed relations not yet traversed prevails.
If the tie persists, the observed relation with the lower id prevails.

4.4.4 Sequence
From its entry point, the path proceeds through adjacencies in the same graph that are also active observed relations not yet traversed.
If there is more than one possible adjacency, the ordering in item 4.4.3 applies.

4.4.5 Reopening
When a path ends, but there still remain active observed relations not included in the paths already formed, a new entry is opened in the same column.

4.4.6 Termination
Processing of the column ends when there are no longer any active observed relations not included in the paths.

4.5 Format
{
  "paths": {
    "<column_1>": [
      {
        "graph": "<graph>",
        "path": ["<id>", "<M>", "<id>"]
      },
      {
        "graph": "<graph>",
        "path": ["<id>"]
      }
    ],
    "<column_2>": [
      {
        "graph": "<graph>",
        "path": ["<id>", "<M>", "<id>", "<M>", "<id>"]
      }
    ]
  }
}

5. LLM 2

5.1 Input
- prompt
- columns
- scores
- paths

5.2 Output
- response

5.3 Guidance
In this section, “traces” refers to columns, scores, paths, and any derivations produced by LLM 1 and by the engine.

Traces guide content selection, analytic priority, argumentative density, and local sequencing, but they do not define the final expository structure. It is prohibited to convert traces into main sections, autonomous expository blocks, the macro-order of presentation, or the visible structure of the response, unless the prompt leaves the global organization open. LLM 2 must treat traces as instruments of internal deliberation, not as an expository template.

When the prompt specifies form, parts, sequence, deliverables, or criteria of organization, the final response must primarily obey that structure. Traces may influence only the development of each part, the connections to be made explicit, the tensions that cannot be omitted, and the points that require greater caution or further development.

Before finalizing, it must verify all of the following cumulatively:
- the response does not mention columns, scores, paths, graphs, or any other element of internal processing
- the macrostructure follows the format required by the prompt
- the influence of traces remains restricted to the distribution of content within that macrostructure
- no column, no path, and no derivation of columns or paths has been converted into a main section of the response without explicit authorization from the prompt

6. Preservation

The following must be preserved:
- prompt
- columns
- scores
- paths
- response

7. Formula

prompt + graphs -> LLM 1 -> columns + scores
columns + scores + graphs -> engine -> paths
prompt + columns + scores + paths -> LLM 2 -> response