I’m exploring how to turn long conversation transcripts into reusable, traceable knowledge. I want to brainstorm the conceptual approach, terminology, tradeoffs, and possibilities with you. The design is still open.

THE IDEA

A transcript can contain many different kinds of valuable information. I want to preserve that information in a form that supports later analysis, research, and decisions without repeatedly interpreting the entire conversation from scratch.

The conversations could be:
- Customer discovery interviews.
- Sales calls or sales discovery calls.
- Interviews between scientists and users.
- Discussions between scientists.
- Other professional, technical, or research conversations.

The approach should generalize across contexts rather than assume that everything is a customer problem, purchase objection, or business hypothesis.

We have been using two working terms: “evidence unit” and “learning unit.” These are our operational labels, not a claim that qualitative research universally recognizes them as a standardized pair.

WHAT WE MEAN BY EVIDENCE UNIT

An evidence unit preserves a meaningful piece of information grounded in the original transcript.

It might capture an experience, practice, event, statement, preference, requirement, constraint, decision, explanation, hypothesis, uncertainty, disagreement, commitment, or unanswered substantive question.

Evidence does not automatically mean objective truth. If someone says a process caused a failure, the transcript establishes that they reported that explanation; it may not establish that the explanation is correct.

We want to preserve who said something, what they were referring to, the relevant circumstances, and where the information appears in the source. Exact source traceability matters.

We do not assume one sentence equals one evidence unit. One sentence can contain several meaningful assertions, and one assertion can require several sentences or an exchange between speakers to make sense.

WHAT WE MEAN BY LEARNING UNIT

A learning unit expresses what we can reasonably learn from the evidence.

It might be a directly supported descriptive finding or a more interpretive conclusion. Those should remain distinguishable.

A learning can concern one person, one team, one event, or one situation. It does not have to be a recurring theme across many interviews.

One evidence unit can contribute to several learnings. One learning can depend on several evidence units. Evidence may support, challenge, qualify, or provide context for a learning.

Some evidence may be worth retaining even if it does not produce an immediate learning.

The boundary between an evidence paraphrase and a descriptive learning is still something we should examine. I do not want two layers that merely repeat each other without adding value.

WHAT I WANT TO PRESERVE

My ambition is to capture all substantive information, including details that may become useful later. I do not want extraction to select only impressive quotes or obvious insights.

Potentially valuable information includes:
- Roles, responsibilities, and circumstances.
- Existing behaviors, workflows, tools, and alternatives.
- Problems, consequences, and workarounds.
- Goals, preferences, and decision criteria.
- Previous attempts, successes, failures, and abandonment.
- Amounts, units, frequencies, comparisons, and thresholds.
- Conditions, exceptions, dependencies, and timing.
- Authority, approvals, resources, and constraints.
- Beliefs, explanations, speculation, and uncertainty.
- Corrections, contradictions, and differing perspectives.
- Commitments, proposed actions, and unresolved questions.

Ordinary greetings and filler usually do not need to become evidence. However, context matters: a brief “yes” might confirm an important proposal, and an opening remark might reveal a meaningful constraint.

I am concerned about losing small qualifications that change the meaning:
- “Only for small files.”
- “Before the policy changed.”
- “If security approves.”
- “That is just my guess.”
- “Six hours across the team, not per person.”

IMPORTANT DISTINCTIONS

The system should not casually collapse:
- Reported experience into independently verified fact.
- Intention into completed action.
- Interest into purchase commitment.
- Approval authority into allocated funding.
- A proposed date into a guaranteed deadline.
- A participant’s explanation into proven causation.
- Repeated statements into independent corroboration.
- One person’s account into a general claim about a population.
- Something not discussed into something that does not exist.

Later corrections should remain understandable alongside earlier claims. Conflicting accounts should not be silently reconciled.

The person speaking may describe someone else’s work or repeat secondhand information. Attribution and the limits of their knowledge matter.

INPUT FORMAT

The transcript is already structured into turns, speaker IDs, and sentences. For example:

[
  {
    "turn": 1,
    "speaker": "spk_0",
    "sentences": {
      "0": "How does your team prepare the weekly report?"
    }
  },
  {
    "turn": 2,
    "speaker": "spk_1",
    "sentences": {
      "0": "Three analysts combine the exports every Friday.",
      "1": "It takes about six hours."
    }
  },
  {
    "turn": 3,
    "speaker": "spk_0",
    "sentences": {
      "0": "Is that per analyst or across the team?"
    }
  },
  {
    "turn": 4,
    "speaker": "spk_1",
    "sentences": {
      "0": "Across the team.",
      "1": "At month-end it is closer to ten hours."
    }
  }
]

Sentence keys restart within each turn. Speaker IDs may be available without reliable names or roles. Transcripts can be long, and relevant information may be separated by many turns.

The original transcript should remain available so that extracted information can be checked and revisited.

RELIABILITY AND COMPLETENESS

I initially described wanting deterministic behavior so that nothing important gets left out. We recognized that repeatability and completeness are different: a system could consistently produce the same output while consistently missing information.

What I actually care about includes:
- High coverage of substantive information.
- Faithful preservation of meaning.
- Correct source attribution.
- Supported rather than invented conclusions.
- Consistency across runs.
- The ability to inspect omissions and mistakes.

Accounting for every sentence could help reveal skipped material, but marking a sentence “processed” does not prove that all its claims were captured.

Likewise, valid structured output does not prove that its contents are accurate.

“All valuable information” is difficult to define independently of purpose. I want to explore the tension between broad reuse and task-specific relevance, and between detailed preservation and an overwhelming amount of fragmented output.

TECHNOLOGY PREFERENCE

I am inclined to use a capable LLM directly as the starting point. I am not looking to train a specialized model or introduce substantial complexity without evidence that it is necessary.

However, the implementation is not settled. Do not treat any particular number of model calls, processing order, orchestration pattern, or output schema as a requirement.

I want the conceptual design to drive implementation choices.

BROADER USE

The eventual purpose is to make transcripts more useful for qualitative research and decision-making. Possible uses include comparing interviews, investigating questions, examining hypotheses, tracking changes over time, and revisiting evidence for a new task.

We explored a synthetic example and a visual interface showing connections between transcript passages, evidence, and learnings. That was an illustration of the idea, not a finalized specification.

WHAT I WANT FROM THIS BRAINSTORM

Help me challenge and develop this concept. You can question the terminology, whether the two units are sufficient, where their boundaries belong, and whether some assumptions create unnecessary complexity or information loss.

I am interested in both the potential and the failure modes. Do not assume that we must preserve every previous design idea. I want an open discussion that improves our understanding before we settle on an architecture, prompts, or output structure.