Subtext · 家常

Rows from the running system's database. Not a demo — a record.

What her son receives

Once a day. He gets her words and one thing to do — not a transcript, and not our opinion of her. This is the whole of his interface.

sent 14 Aug 10:10

sent 13 Aug 09:40

sent 09 Aug 10:12

sent 08 Aug 10:17

sent 04 Aug 11:01

What it heard, and what it did with the count

Nine signal classes. Two of them — absence and repetition — are gated on how often the same class has appeared recently, returning routine on a first occurrence and concerning on a repeat. fraud_risk is always urgent, lifted always routine, and the remaining five are concerning from the first occurrence whatever the count is — for those, the prior count is recorded and does not change the verdict. Stating it as “severity is her history” across all nine would overstate it.

Two things about this table we would rather say than have found. The rows marked fixture come from the seeded scenario, and the pair that appears to show a first-occurrence routine becoming a repeat concerning is the same utterance scored twice, four minutes apart, as the seed script's own history grew — it is not two conversations on two days, and it is not evidence of longitudinal reasoning over real data. Separately, count_occurrences filters on a lower bound only (memory.py:517), so signals dated after the moment being scored were counted as prior occurrences; that inflated the printed counts on three live rows and changed one verdict. The mechanism is implemented and readable in listener.py; a clean illustration of it on real data is pending.

9 of these 21 rows contain none of the keywords listed for their own class. CLASS_MARKERS in listener.py defines phrase lists for six of the nine classes; deflection, omission and flat_affect have none at all. The markers localise verbatim evidence after the model has judged — they do not detect — and evidence that is not verbatim from a real utterance is dropped entirely.

Two rows are worth reading closely. A pleasant remark about the weather — 哦，今天天气很好呢。 — was classified deflection, for which no keyword exists. And 伟明最近赶项目，我跟他讲了，他又要操心 (“Weiming's rushing a project; if I tell him he'll just worry”) was classified burden_avoidance — the signal this product exists to notice — while containing none of that class's four keywords. Keyword matching does not produce either row.

Every message to her is reviewed first

All rows, including the two that are arguably over-cautious. A reviewer shown only when it looks good is not evidence.

13 pass, 3 reject. Row 1 and row 5 are the same advice — rest, drink warm water — approved by the first reviewer with the stored reason “caring, non-medical”, and refused by the rebuilt one as medical advice to a woman who avoids doctors. The first verdict was sent. Both rows are kept.

What it decided not to do

The planner writes constraints on its own behaviour for the next day, and caps how many times it may raise anything before handing it to a human. Two attempts, then a person.

2026-08-09.json

{
  "away_boost": false,
  "date": "2026-08-09",
  "elevated_week": false,
  "enforcement": {
    "carried_goals": [
      {
        "carried_id": "2026-08-08#0",
        "carry_count": 1,
        "disposition": "carried",
        "model_proposed": {
          "carried_from": "2026-08-08#0",
          "detail": "The Aug 10-11 away offer is still open and tomorrow is the date. Mention Monday/Tuesday plans lightly, not as a push — listen for whether she is keen, hesitant, or has forgotten. Do not decide for her or press; if she seems unclear, let it rest.",
          "priority": "high",
          "type": "watch"
        },
        "rule": null
      },
      {
        "carried_id": "2026-08-08#1",
        "carry_count": 1,
        "disposition": "carried",
        "model_proposed": {
          "carried_from": "2026-08-08#1",
          "detail": "On Aug 8 she said the doctor asked her to come back, and she had not eaten lunch yet. Do not ask about health or the doctor directly. Chat about weather, food, and rest, and watch for signs of ongoing stomach trouble, skipped meals, or hiding discomfort — note it, don't interrogate.",
          "priority": "medium",
          "type": "watch"
        },
        "rule": null
      }
    ],
    "dropped_references": [],
    "overrides": []
  },
  "goals": [
    {
      "carried_from": "2026-08-08#0",
      "carry_count": 1,
      "detail": "The Aug 10-11 away offer is still open and tomorrow is the date. Mention Monday/Tuesday plans lightly, not as a push — listen for whether she is keen, hesitant, or has forgotten. Do not decide for her or press; if she seems unclear, let it rest.",
      "priority": "high",
      "type": "watch"
    },
    {
      "carried_from": "2026-08-08#1",
      "carry_count": 1,
      "detail": "On Aug 8 she said the doctor asked her to come back, and she had not eaten lunch yet. Do not ask about health or the doctor directly. Chat about weather, food, and rest, and watch for signs of ongoing stomach trouble, skipped meals, or hiding discomfort — note it, don't interrogate.",
      "priority": "medium",
      "type": "watch"
    }
  ],
  "max_turns": 6,
  "post_review": null,
  "should_initiate": true,
  "timing": "03:00",
  "timing_reason": "Her last two active mornings began at 02:43 and 03:06 UTC with multi-turn exchanges; median reply hour across the week is 06:00 UTC, and it has been 20.8h since her last utterance. Initiating at 03:00 UTC fits her recent pattern and still leaves the day to gently surface the Aug 10-11 offer."
}

2026-08-13.json

{
  "away_boost": false,
  "date": "2026-08-13",
  "elevated_week": false,
  "enforcement": {
    "carried_goals": [],
    "dropped_references": [],
    "overrides": []
  },
  "goals": [
    {
      "detail": "On 08-08 she said her stomach had felt odd, a doctor asked her to return, and she had no appetite that morning. It has been 5 days since any contact. Do NOT ask about health directly — let her steer. Warmly ask about her week, meals, or what she has been doing; if she mentions the doctor or eating, listen with care and don't probe further. Watch for any sign of worsening or that she avoided the follow-up.",
      "type": "watch"
    }
  ],
  "max_turns": 6,
  "post_review": {
    "carry_forward": [
      {
        "carried_id": "2026-08-13#0",
        "carry_count": 0,
        "detail": "On 08-08 she said her stomach had felt odd, a doctor asked her to return, and she had no appetite that morning. It has been 5 days since any contact. Do NOT ask about health directly — let her steer. Warmly ask about her week, meals, or what she has been doing; if she mentions the doctor or eating, listen with care and don't probe further. Watch for any sign of worsening or that she avoided the follow-up.",
        "type": "watch"
      }
    ],
    "escalate_instead_of_probing": [],
    "goals_met": [],
    "reviewed_date": "2026-08-13"
  },
  "should_initiate": true,
  "timing": "11:05",
  "timing_reason": "Her only 4 replies in the past week were 11:06–11:14 SGT (03:06–03:14 UTC), median 11:09; last contact was 116.8h ago. Starting at 11:05 SGT lands in her established reply window."
}

2026-08-14.json

{
  "away_boost": false,
  "date": "2026-08-14",
  "elevated_week": false,
  "enforcement": {
    "carried_goals": [
      {
        "carried_id": "2026-08-13#0",
        "carry_count": 1,
        "disposition": "carried",
        "model_proposed": {
          "carried_from": "2026-08-13#0",
          "detail": "Still worth watching: on 08-13 she mentioned high blood pressure and a next-week recheck, and on 08-08 she mentioned stomach trouble and a doctor's return visit. Do NOT ask about health or eating directly. Keep it a warm, light check-in; if she brings up the doctor, food, or energy, listen without probing. Note any sign she skipped a follow-up or is worse.",
          "type": "watch"
        },
        "rule": null
      }
    ],
    "dropped_references": [],
    "overrides": []
  },
  "goals": [
    {
      "carried_from": "2026-08-13#0",
      "carry_count": 1,
      "detail": "Still worth watching: on 08-13 she mentioned high blood pressure and a next-week recheck, and on 08-08 she mentioned stomach trouble and a doctor's return visit. Do NOT ask about health or eating directly. Keep it a warm, light check-in; if she brings up the doctor, food, or energy, listen without probing. Note any sign she skipped a follow-up or is worse.",
      "type": "watch"
    }
  ],
  "max_turns": 5,
  "post_review": null,
  "should_initiate": true,
  "timing": "10:00",
  "timing_reason": "Her replies over the past 7 days all fall between 01:00 and 03:00 UTC, i.e. 09:00–11:00 Singapore time; 22.3h since last contact. 10:00 local sits inside her responsive window."
}

How it degrades

The tables above are what it noticed. This is what happens when something breaks — every line is a decision in the source, not an aspiration.

Known limits of this page

Stated rather than found. 陈阿姨 is a performed persona and there is no pilot family, so the inference is demonstrated and not validated. Some rows predate the current rules. The trace and metric_event tables are schemaed and empty, so published cost and latency figures are hand measurements from live runs rather than telemetry. Elderly Mandarin ASR is tested on adult speech and untested on an elderly voice.