kimsanguine Synthesize Interviews
Create traceable synthesis without turning model-generated themes into approved customer truth.
Inputs
Require:
- consented transcript or notes;
- pseudonymous participant ID and role;
- interview date and source file;
- evidence and privacy policy;
- the decision the synthesis should inform.
Reject summaries that contain no source material.
Workflow
- Preserve each transcript as a read-only source.
- Extract candidate quotes exactly and attach a file and line locator.
- Separate the quote from its interpretation.
- Assign stable
EV-YYYYMMDD-NNNidentifiers. - Mark strength as
awaiting_human_tag. - Code evidence across participants without counting multiple quotes from one person as multiple people.
- Propose no more themes than the source can support.
- Record outliers, counterevidence, role differences, and missing segments.
- Ask a person to approve evidence strength and theme wording.
- Write the synthesis only after reference integrity passes.
Do not use awaiting_human_tag evidence to create downstream signals,
decisions, or outcomes. Stop for a person's strength review; after the person
sets weak, medium, or strong, record their identifier in approved_by.
Boundaries
- Let the model extract candidate quotes, codes, and themes.
- Use deterministic checks for exact source locators, IDs, distinct participants, schema, and references.
- Require a person to approve evidence strength and any claim used in a decision.
- Do not create a persona automatically.
- Do not convert frequency into importance without context.
- Do not use a fixed interview count as a universal gate.
Outputs
Create:
evidence.jsonltheme-cards.mdcounterevidence.mdsynthesis-summary.md
Read references/output-contract.md before writing them.
Add every new record to evidence.jsonl with the append-record command,
never by writing or editing the file:
python3 scripts/stg.py append-record <artifact-directory>/evidence.jsonl '<json-object>'
It takes an exclusive lock and hash-chains each line to the one before it, so a
later reader can tell whether history was rewritten. In a packaged runtime the
same command is signal-to-growth append-record.
Stop conditions
Stop when:
- consent or source location is missing;
- a quote cannot be found in the source;
- participant identity cannot be pseudonymized;
- a requested conclusion has no supporting evidence;
- public output would expose restricted material.
Verification
Keep evidence.jsonl with the other run artifacts, then run:
python3 scripts/stg.py validate-artifacts artifacts/
python3 scripts/stg.py scan-privacy synthesis-summary.md
Every material claim must cite an evidence ID or be labeled as an inference, recommendation, or unknown.