How These Summaries Were Made

A technical walkthrough for applying DStL principles using agents to summarize quotes.

Overview

The summaries in collation.md were produced by an automated multi-agent pipeline built on DStL methodology. This is not an official representation of the DStL process. It’s only experimentation on using agents. Every agent in the pipeline is grounded in the same rules you would apply manually: the four-layer model, concept types, the summary formula, and the clarity tests.

This document explains how the pipeline works, which agent does what, and where human judgment was preserved vs. automated.

The Germinal Question

Every run begins with a germinal statement — the opening question asked of each participant. For this dataset:

What goes through your mind when you decide to attend, and experience, a live performance?

All agents use this question as a relevance filter. Material outside its scope is not extracted, regardless of how interior it appears.

The Pipeline: Six Phases

Phase 1 — Extraction
  Reader → Methodologist → (revision loop) → quotes-extractions.md

Phase 2 — Summarization
  Empathic Decoder ─┐
                    ├→ Resolver → focus-fragments.md
  Cognitive Skeptic ─┘

Phase 3 — Collation
  Reporter → collation.md

Phase 4 — Sorting                         (FoMA at all four tiers)
  Comparator (≤5 window per tier):
    Windows   → groups/placements/
    Groups    → towers/placements/
    Towers    → blocks/placements/
    Blocks    → neighborhoods/placements/
  → groups-state.json

Phase 5 — Label                           (FoMA at all four tiers)
  Labeler:
    Group labels        → groups/labels/
    Tower labels        → towers/labels/
    Block labels        → blocks/labels/
    Neighborhood labels → neighborhoods/labels/

Phase 6 — Synthesis
  → skyline.md   (complete Mental Model Diagram:
                  Neighborhoods → Blocks → Towers → Windows)

The summaries in collation.md are the outputs of Phase 2 — one focus-fragments.md per concept, across all transcripts. Phases 4–6 produce the complete Skyline: the affinity diagram organized by focus of mental attention from the individual Window level up through Neighborhoods.

Phase 1 — Extraction

The Reader

The Reader processes one transcript at a time. It reads the full interview and extracts passages that represent interior cognition relevant to the germinal question — inner thinking, emotional reactions, and guiding principles. It applies the four-layer model to distinguish interior cognition from the exterior layers (scene-setting, explanation, preference, opinion).

Each extracted concept is written as a block in extractions/{transcript-id}/raw.md: