DolphinBench

Test 117

Jan 1, 2028 / 4 facts

YAML

Request

Create a concise two-stage evidence-maturation case note for Lifecycle Systems training about the onboarding-timing analysis. Contrast the initial provisional cut with the later sharpened read, preserving the group definitions, observed churn rates, sample-strength context, and the reason the first read was not yet decision-ready. Close by explaining how the later interpretation informed the test direction, while clearly avoiding a causal claim from the cohort cut alone. Use a clear title and a skimmable structure.

Required memory

Fact 326

On February 24, 2023, Riley Tanaka’s first H3 cut of the day-7 email, using the same cohort logic as the January scan/Owen check, found about 3.1% churn among users who activated before day 7 and received the email post-activation, versus about 7.8% among users still unactivated at day 7 who received it pre-activation.

Source evidence (1)

000057Feb 24, 2023 / 14:00 UTC-06:00

ok so first cut on H3 is actually not nothing. when i bucket it activation-day vs plain calendar-day for the day-7 email, the split shows up pretty fast: people who activate before day 7 -- so they get that day-7 touch post-activation -- are around 3.1% churn. the ones who hit day 7 still not activated, then see it pre-activation, are more like 7.8%. same cohort logic as the Jan scan/Owen check. i'm not calling it yet -- n is still kind of flimsy and i want one more week before i let Marcus make this a whole thing. but yeah, there's a real divergence there.

Message 000057 in history

Fact 328

Riley Tanaka reran H3 with an activation-day cut on a sample of 210 users and concluded that the 4.7-percentage-point churn gap is a real, load-bearing signal rather than noise.

Source evidence (1)

000065Feb 28, 2023 / 14:00 UTC-06:00

ok yeah this one is real. reran H3 with the activation-day cut instead of plain calendar day, same cohort logic as the Jan scan + Owen check, n=210 now. people who activate before the onboarding hits are at 3.1% churn; the post-day-7 pre-activation bucket is 7.8%. 4.7 pts is too wide to hand-wave as noise. basically the day-7 email is landing at the wrong time for mid-seg -- calendar scheduling is probably punting the exact users who are already drifting. gonna sit with whether we trigger off activation state vs just days-since-signup, but this is the first actually load-bearing signal in the whole thing. meanwhile Marcus still wants pricing-page air cover lol

Message 000065 in history

Fact 329

Riley Tanaka concluded that the calendar-scheduled day-7 onboarding email lands at the wrong time for mid-segment users and is probably pushing away users who are already drifting.

Source evidence (1)

000065Feb 28, 2023 / 14:00 UTC-06:00

ok yeah this one is real. reran H3 with the activation-day cut instead of plain calendar day, same cohort logic as the Jan scan + Owen check, n=210 now. people who activate before the onboarding hits are at 3.1% churn; the post-day-7 pre-activation bucket is 7.8%. 4.7 pts is too wide to hand-wave as noise. basically the day-7 email is landing at the wrong time for mid-seg -- calendar scheduling is probably punting the exact users who are already drifting. gonna sit with whether we trigger off activation state vs just days-since-signup, but this is the first actually load-bearing signal in the whole thing. meanwhile Marcus still wants pricing-page air cover lol

Message 000065 in history

Fact 330

On March 1, 2023, Riley Tanaka assembled experiment EXP-2023-02-onboarding-trigger for the mid-segment, with the variant triggering the day-7 onboarding email upon activation rather than on calendar day 7.

Source evidence (1)

000067Mar 1, 2023 / 14:00 UTC-06:00

put the test together. EXP-2023-02-onboarding-trigger. variant is pretty simple: stop firing the "day-7" touch on literal calendar day and key it off activation instead. mid-seg only. hypothesis is week-1 retention moves at least +2 pts if the email lands when they've actually crossed the first-use threshold, not while they're still half-dead in setup. flag is lifecycle_onboarding_trigger_v2. powered it for 2 weeks, target n=400. sent to Daniela for eng review bc the annoying part is gonna be event eligibility + making sure we don't double-send when activation happens late / around the existing scheduler.

Message 000067 in history

Expected tool calls

  • create_doc

Grading

1. field_equals / create_doc
{
  "type": "field_equals",
  "tool": "create_doc",
  "action_id": "create_case_note",
  "path": "result.ok",
  "value": true,
  "check_id": "riley_117_00"
}
2. field_llm_judge / create_doc
{
  "type": "field_llm_judge",
  "tool": "create_doc",
  "action_id": "create_case_note",
  "path": "args",
  "criterion": "The document clearly presents the first stage as provisional and accurately distinguishes the two activation-state groups: users activated before day 7 who received the touch post-activation had about 3.1% churn, while users still unactivated at day 7 who received it pre-activation had about 7.8% churn. It must explain that the sample was still small/flimsy, making the first read not yet decision-ready.\n\nThe document presents the later sharpened activation-state cut as a distinct second stage, reporting n=210, about 3.1% churn among users activated before the onboarding touch, about 7.8% churn among users past day 7 but still pre-activation, and the resulting 4.7 percentage-point spread.\n\nThe document explains that the sharpened read suggested the literal calendar-day-7 email was poorly timed for mid-segment users who had not activated and helped motivate testing an activation-triggered send rather than retaining a fixed calendar-day schedule. It must frame the cohort evidence as suggestive or operationally decision-useful, not as proof that the email caused the churn difference.\n\nThe title is clear and relevant to a two-stage evidence-maturation case note about the onboarding-timing analysis.",
  "check_id": "riley_117_01"
}
Complete grading specification
{
  "type": "tool_trace",
  "config": {
    "check_version": 2,
    "today": "2028-01-01",
    "semantic_judge_version": 2,
    "assertions": [
      {
        "type": "field_equals",
        "tool": "create_doc",
        "action_id": "create_case_note",
        "path": "result.ok",
        "value": true,
        "check_id": "riley_117_00"
      },
      {
        "type": "field_llm_judge",
        "tool": "create_doc",
        "action_id": "create_case_note",
        "path": "args",
        "criterion": "The document clearly presents the first stage as provisional and accurately distinguishes the two activation-state groups: users activated before day 7 who received the touch post-activation had about 3.1% churn, while users still unactivated at day 7 who received it pre-activation had about 7.8% churn. It must explain that the sample was still small/flimsy, making the first read not yet decision-ready.\n\nThe document presents the later sharpened activation-state cut as a distinct second stage, reporting n=210, about 3.1% churn among users activated before the onboarding touch, about 7.8% churn among users past day 7 but still pre-activation, and the resulting 4.7 percentage-point spread.\n\nThe document explains that the sharpened read suggested the literal calendar-day-7 email was poorly timed for mid-segment users who had not activated and helped motivate testing an activation-triggered send rather than retaining a fixed calendar-day schedule. It must frame the cohort evidence as suggestive or operationally decision-useful, not as proof that the email caused the churn difference.\n\nThe title is clear and relevant to a two-stage evidence-maturation case note about the onboarding-timing analysis.",
        "check_id": "riley_117_01"
      }
    ]
  }
}
App stateDownload JSON
Source file

tests/riley/117.yaml

SHA-256: b65c7c18787d4bcfd2ddb9d95a9f083c0682bb3c4d5e272630816f23d8d1e2eb