Branching & flip-tests

Branch a conversation

POST /api/v1/insights/branch forks a fresh persona chat at a fault line (e.g. a miner branchPrompt), runs a short deepening exploration, and reports back — without touching your main thread.

POST /api/v1/insights/branch

{ "audienceId": "aud_f1_1",
  "branchPrompt": "You mentioned trust — what exactly would make you trust the numbers?",
  "subject": "Trust", "depth": 2, "seed": 7 }
# → data: { subject, opened, transcript:[{role,content}], summary, ranLive }

Field	Meaning
`branchPrompt`	The question that opens the branch (required).
`subject`	Label for the fault line (for the report header).
`depth`	Follow-up turns to explore (1–4, default 2).

The flip-test loop

POST /api/v1/insights/flip-test runs find → branch → verify in one call:

Find

The smallest model hypothesizes what condition would flip an opinion in your turns.

Branch

Forks a fresh chat: captures the persona's baseline stance, introduces the hypothesized condition, re-asks.

Verify

Code-computed stance delta (before vs after) → a verdict. The movement is measured from real text; the hypothesis + prose are model-generated.

POST /api/v1/insights/flip-test

{ "audienceId": "aud_f1_1",
  "topic": "$29 finance app; price-sensitivity & scaling",
  "turns": [{ "speaker": "Riley", "text": "Cheap means it won't scale for my team" }] }

# → data: {
#   hypothesis: { message, branchPrompt, confidence },
#   baseline:   { question, answer, stance },
#   afterCondition: { question, answer, stance },
#   verdict: "flipped" | "softened" | "unchanged" | "hardened" | "no-test",
#   delta: -1..1, report: "…measured in code…" }

Honest by construction. A real example returned unchanged (Δ0) because the persona revealed the true objection was data lock-in, not scale — so "prove it scales" wouldn't move them. The verdict is a measurement, not a model claim.