Working draft v5

Narrative analysis methodology

BIG EARS is designed to detect whether a respondent is curious — whether they can articulate, elaborate on, and extend a need, want, hope or expectation — rather than simply cataloguing which topics they raise. Topic content is still captured, but it is secondary to the engagement signal each answer carries. Absence of elaboration is treated as the finding, not a data gap.

The five-question sequence

Adaptive branching

A single threshold governs adaptivity: 250 words of narrative material, whether it arrives entirely from Q1 or accumulates across Q1+Q2. Q2 is generated adaptively only when Q1 alone reaches 250 words; otherwise Q2 falls back to a generic prompt ("Okay, that's interesting. Can you tell me a bit more about that?"). Q3 is generated adaptively whenever the combined Q1+Q2 narrative reaches 250 words, generic if not. Every respondent is asked a question of the same function at each stage — the interview move stays constant even when the literal wording differs.

Big Ears' R1–R16 pronoun-anchor scoring engine

When a follow-up runs adaptively, its underlying intent never changes — only its contextual framing does, chosen by scoring the respondent's own preceding text against a 16-cell matrix keyed to pronoun anchors (I, my, me, mine, it, them, their, our, you, we, your, us) and five gerund-tagged directions: Bridging (R1), Becoming (R5), Belonging (R7), Building (R10) and Integrative (R16). Exact anchor words score 3 points, related family words score 2, adjacent support phrases score 1, and specific pronoun co-occurrences (e.g. "my" + "them" in one clause) add a further 3+3. The dominant and runner-up cells, their confidence band, and the trigger words are logged for every adaptively-generated question — this log is what makes an AI-generated prompt auditable after the fact.

This 16-cell matrix is specific to Big Ears and unrelated to any other R1–R16 numbering used elsewhere in the firm's work — the naming is a coincidence, not a shared system.

Curiosity coding layer

Each response is scored on six dimensions: elaboration level (absent/vague/articulate), Q3 directionality (extension vs. flattening once handed their own words back — the primary curiosity signal), pronoun stance shift across Q1→Q4, specificity/concreteness, thread continuity between an adaptive prompt and the answer that follows it, and affect shift across the four answers.

Location & electorate resolution

State, suburb and postcode are captured last, after all five narrative questions, and are entirely optional with a visible skip — never asked before the narrative content, never a gate on taking part. When given, suburb is the primary resolution key; postcode is stored as a worst-case tiebreaker. Federal/state electorate resolution against current AEC and state boundary files is an open dependency flagged in the source methodology (licensing/build scope, not yet integrated) — this deployment stores location structurally and marks electorate as pending rather than guessing at one.

Rules of engagement

• This produces a reading, not a verdict — nothing here judges you as a person, and no single response is a final or permanent read.
• The six coding dimensions are reported separately and never collapsed into one composite "curiosity score."
• Follow-up questions only ever reflect your own words back to you — they're never rewritten or reinterpreted on your behalf.
• Your response carries no name or contact detail, and it's never used to target or follow up with you individually. The only optional exception is the "keep me posted" email signup at the very end — that goes to a completely separate list, never linked to what you told us.
• You can stop or skip at any point, including the location question, with no penalty.
• This is a listening instrument, not an automated decision system — a person reads what it surfaces and decides what to do with it.

Known limitations & open items (from the source working draft)

• Adaptive question governance: no presupposed sentiment, no leading framing, no assigned blame — this build follows that rule by keeping the fixed question intent constant and only varying a neutral framing lead-in.
• Auditability: every adaptively-generated question logs its dominant/runner-up R-cell, confidence and trigger words (visible in Admin).
• Electorate boundary data is a licensing/build dependency, not yet integrated — see above.
• Usefulness (U) and Opportunity (O) in AEIOUF, previously only inferred at coding, are now asked directly via Q5's LESS/MAINTAIN/MORE slider.
• Phone (CATI) collection is a legitimate second channel for the same instrument but is out of scope for this digital build.
• Voice: speaking an answer is transcribed by Deepgram before it reaches scoring; questions can be read aloud via the browser's own text-to-speech. Neither changes the scoring or coding logic above — voice is only ever an input/output layer around it.

Prepared as a working draft for review — not yet a final instrument.