Explore fragment-specific agent training and escalation of communication

Summary

Explore training local models on specific fragments of the emsemioverse so they become experts in that fragment. This is part of a wider exploration of how agent profiles vary across formalization levels, and how the existing “escalation of communication” concept maps onto skill maturity and progressive automation.

Motivation

The skill maturity concept already describes five stages of formalization: inference-heavy → structured inference → delegable → procedural → tool. Each stage shifts execution from a high-capability agent toward deterministic tooling.

A gap exists between stages 2 (structured inference) and 3 (delegable): the current delegate_task tool passes a single prompt with context to a generic local model. The model has no domain expertise — it transforms supplied content without understanding the domain’s vocabulary, conventions, or relationships.

Fragment-specific training would create models that are fluent in a specific fragment’s vocabulary and patterns. A model trained on semiotic-endeavor content, for example, would understand terms like “method component,” “closure condition,” and “lifecycle guard” without needing them explained in every prompt. This fills the gap between “needs a capable general model” and “can be scripted deterministically.”

Escalation of communication

Emsenn’s 2019 essay “An Escalation of Communication” describes how ideas should progress through increasingly public venues: private thought → family discussion → friends → public presentation. Each stage tests the idea against a wider, less sympathetic audience.

This maps onto the skill maturity axis:

Communication stage	Skill maturity stage	Agent profile
Private thought	Inference-heavy	High-capability agent (Claude Opus) exploring a new task
Family discussion	Structured inference	Same agent with typed I/O contracts
Friends / associates	Delegable	Fragment-trained local model with domain expertise
Public presentation	Procedural / Tool	Deterministic script or MCP tool

The fragment-trained model occupies the “friends” position: it can be trusted with domain-specific work because it speaks the language, but it doesn’t have the full judgment of the “family” (high-capability agent). Ideas that pass through the fragment-trained model without errors are ready for further formalization.

This is also perpendicular to the agent-skill-proficiency axis (which measures trust in a specific agent for a specific skill). A fragment-trained model has higher proficiency for skills within its fragment and lower proficiency outside it.

Research phase

What exists

1. Skill maturity concept at specifications/agential-semioverse-repository/concepts/skill-maturity.md — already defines the 5-stage lifecycle
2. Agent skill proficiency at mathematics/objects/universes/agential-semioverse/concepts/agent-skill-proficiency.md — already defines per-skill trust
3. Progressive formalization described in theory/agent-skill-interaction.md — conversation → specification → skill → script → formal proof
4. Escalation of communication essay at personal/writing/letters-to-the-web/2019-01-28-an-escalation-of-communication.md
5. Policy 001 (progressive automation) — “make actions accomplishable by less capable agents over time”

What needs research

1. Fine-tuning methods for small local models: what does it take to train a qwen2.5:7b variant on a corpus of ~100 markdown files? QLoRA? Full fine-tuning? RAG as an alternative?
2. Evaluation methodology: how do you measure whether a fragment-trained model is actually better for that fragment than a generic model? Compare enrichment quality? Frontmatter accuracy?
3. Fragment boundaries: what constitutes a “fragment” for training purposes? A single specification? A discipline? A layer of the conceptual stack?
4. Interaction with MCP: should a fragment-trained model be exposed as a separate MCP tool (e.g., delegate_to_endeavor_model) or as a model parameter to the existing delegate_task?

Steps

1. Survey: research fine-tuning approaches for small models (Ollama-compatible), especially QLoRA and corpus-based training. Produce a text documenting findings.

2. Design fragment boundaries: decide what “fragment” means. Options: single spec (semiotic-endeavor), single layer (agential semioverse), single discipline (technology), cross-cutting theme.

3. Pilot: fine-tune or RAG-augment a local model on the semiotic-endeavor specification (the most documented fragment). Test against the existing infer_triage_frontmatter task as a benchmark: does the fragment-trained model produce better frontmatter for endeavor-related triage files?

4. Formalize the escalation mapping: write a concept file that connects escalation of communication, skill maturity, and agent-skill-proficiency into a coherent theory of progressive formalization with agent trust at each stage.

5. Integrate into skill system: if the pilot succeeds, design how fragment-trained models fit into the skill manifest system (new runtime value? model parameter? proficiency metadata?).

Done when

• A text documenting fine-tuning/RAG approaches for fragment-specific training exists
• At least one pilot fragment-trained model has been tested against a benchmark task
• The escalation-of-communication mapping to skill maturity is encoded as a concept file
• The interaction between fragment training and the existing skill system is designed (even if not fully implemented)

Dependencies

Independent. Benefits from plan 0043 (script I/O standardization) for the benchmark task, and from plan 0045 (conceptual dependency mapping) for understanding fragment boundaries.

Log

2026-03-08 — Created from emsenn’s observation that training local models on specific fragments overlaps with the “escalation of communication” concept and the skill maturity lifecycle. The progression from general-purpose inference to fragment-trained delegation to deterministic scripting is a concrete instance of progressive formalization applied to agent profiles.