{
  "_about": "How 'chat with my interviews' works on top of this structure. A question is turned into a retrieval over evidence units (filtered by codes, entities, segments, dates, modality), the answer is composed ONLY from retrieved units and must cite them, it must report how many INDEPENDENT sources it rests on, and it must say what was never asked. The answer is then persisted as a learning attached to the question (learnings.json), so the next person who asks gets the same answer with the same provenance - and can see that it was answered when the corpus had 3 transcripts, not 100.",
  "_answer_contract": {
    "citations": "Every factual sentence cites unit ids. Sentences without a citation are the model's synthesis and are marked as such.",
    "counts": "Number of independent sources (people at different organisations), not number of units or number of mentions.",
    "absences": "Which transcripts were never asked about the topic (from probes / cross_relations absent_in). Silence is never counted as a 'no'.",
    "modality": "Secondhand, guessed, or hedged evidence is labelled as such in the answer.",
    "persistence": "The answer becomes a learning with kind=chat_answer, scoped to the transcripts retrieved, generalises_beyond_transcript=true only if independent_sources >= 2."
  },
  "sessions": [
    {
      "id": "chat_01",
      "asked_on": "2026-12-12",
      "asked_by": "product lead",
      "question": "Across our interviews, would people pay for automatic column remapping?",
      "retrieval": {
        "filters": { "codes": ["C13", "C02", "C05"], "transcripts": "all", "status": "active" },
        "also_consulted": ["H2", "obs_05", "xr_11", "xr_13"],
        "units": ["ev_020", "ev_021", "ev_023", "ev_112", "ev_106", "ev_113", "ev_201", "ev_209", "ev_210", "ev_211", "ev_212"]
      },
      "answer": {
        "text": "Not clearly, and it splits by whether the team could solve it themselves. Two organisations have been asked (2 independent sources). At org_A the answer has moved from 'interested in looking at it' [ev_020] to a completed pilot with an 8k quote and money located but not signed off [ev_201, ev_209, ev_210, ev_211] - the closest thing to a purchase in the corpus, and still not one; the speaker's own January estimate is self-discounted [ev_212]. At org_B the answer is a flat no [ev_112]: they built a mapping table into a Power Query script and consider the problem solved [ev_106]; what they say they would pay for is mismatch flagging between sources [ev_113]. Synthesis (not cited): the dividing variable looks like in-house scripting capacity - org_A's script failed, org_B's works - which is the segment H2 should be split on. With n=2 this is a pattern to test, not a finding.",
        "independent_sources": 2,
        "absences": ["No one has been asked what they would pay - only whether they would. Price sensitivity is entirely unknown.", "org_B was not asked whether their script has ever broken (would change the 'solved' claim)."],
        "modality_notes": ["ev_210 (director found money) is secondhand via the speaker.", "ev_212 is a hedged forecast the speaker themselves flagged as unreliable."]
      },
      "persisted_as": "lr_c01"
    },
    {
      "id": "chat_02",
      "asked_on": "2026-12-12",
      "asked_by": "researcher",
      "question": "What changed for person A between the two interviews?",
      "retrieval": {
        "filters": { "entity": "person_A", "transcripts": ["tx_001", "tx_003"] },
        "also_consulted": ["obs_01", "obs_02", "obs_03", "obs_04", "obs_05", "obs_06", "obs_07", "obs_08", "xr_01", "xr_02", "xr_03", "xr_04", "xr_05", "xr_06", "xr_08", "xr_09", "xr_10"],
        "units": ["ev_005", "ev_202", "ev_006", "ev_203", "ev_018", "ev_205", "ev_025", "ev_206", "ev_207", "ev_010", "ev_208", "ev_023", "ev_210", "ev_211", "ev_008", "ev_213", "ev_028", "ev_214", "ev_026", "ev_201"]
      },
      "answer": {
        "text": "Eight things changed, one person, 87 days apart (1 source - this is a longitudinal read, not a cross-sectional one). Effort: weekly prep 6h -> 4h [ev_005, ev_202], month-end 10h -> 7h [ev_006, ev_203]. Gaps: still occur, now attributed to the fleet tracker export missing the 3 p.m. cutoff [ev_018, ev_205] - tx_001 never said what the gaps were, so this may be a definition rather than a change. Security review: the secondhand 'about six weeks' [ev_025] is replaced by a firsthand nine weeks that only started after two chases [ev_206, ev_207]. Column drift cause: the guess about internal template edits [ev_010] is now reported as found out, 'more or less' [ev_208] - how they found out is not stated. Purchase: from no known budget [ev_023] to funds located but unsigned [ev_210, ev_211]; pilot went from 'in principle' [ev_026] to done [ev_201]. Team: Priya left, finance spreadsheet now handled by the speaker directly [ev_008, ev_213]. Directors: third unprompted request for a daily version [ev_028, ev_214]. Synthesis (not cited): the remap tool removed roughly a third of the effort and none of the completeness problem; the completeness problem has an upstream cause the tool cannot touch.",
        "independent_sources": 1,
        "absences": ["Whether sample exports were delivered by October (the only concrete commitment in tx_001) was not asked (fu_006).", "The 'worse parts of the week' hint from tx_001 is still unprobed after two interviews (fu_007).", "Why the daily version is still declined, now that the process has changed, was not asked (fu_008)."],
        "modality_notes": ["ev_208's source is 'unclear' - the speaker says they found out, not how. Do not treat the cause as verified.", "ev_202's '4 from 6' was framed by the interviewer's recall; the speaker confirmed it."]
      },
      "persisted_as": "lr_c02"
    },
    {
      "id": "chat_03",
      "asked_on": "2026-12-13",
      "asked_by": "engineering",
      "question": "How long do security reviews take, and can we avoid them?",
      "retrieval": {
        "filters": { "codes": ["C11"], "transcripts": "all" },
        "also_consulted": ["H3", "obs_04", "xr_14"],
        "units": ["ev_024", "ev_025", "ev_206", "ev_207", "ev_114"]
      },
      "answer": {
        "text": "Duration: two data points, both from org_A (1 source for duration). One is secondhand and approximate - 'about six weeks, I heard' [ev_025]; the other is firsthand - nine weeks for the pilot's review, which did not start until the speaker had chased twice [ev_206, ev_207], so it includes queue time. Avoidability: both organisations (2 independent sources) gate the review on the tool connecting directly to a core system - the warehouse system at org_A [ev_024], the ERP at org_B [ev_114]; org_B states explicitly that tools which only read exported files need no review [ev_114]. Synthesis (not cited): a file-based architecture may bypass the review at both orgs. Whether the org_A pilot needed its review because of how it connected is unknown (fu_009).",
        "independent_sources": 2,
        "absences": ["org_B has never had a review of this kind described, so their duration is unknown.", "Whether the pilot tool's architecture triggered the review was not asked (fu_009)."],
        "modality_notes": ["ev_025 is secondhand and about a different, earlier review. It should not be averaged with ev_206."]
      },
      "persisted_as": "lr_c03"
    }
  ]
}