# Lesson 9 practice: Opening, accepting, and completing are different

[中文](README.md)

Start with [tasks.csv](tasks.csv), then compare [training](intervention-training.csv), [no penalty for justified refusal](intervention-rejection-policy.csv), and [broader information access](intervention-access.csv). The task table expands the public scenario’s 340/96/88/41/33/31 funnel into independent synthetic rows; all other details are new assumptions. Intervention tables are also constructed, not evidence that these changes worked in a real team. No customer interviews or live permission changes are required.

## Fields

| File / field | Meaning |
| --- | --- |
| All files: `source_type` / `task_id` | `independent_synthetic_exercise` / synthetic task ID |
| Tasks: `simulation_day` / `shift` / `simulation_operator` | Simulated day, `day` / `night`, and P-01 to P-12 simulated operator slots; not a mapping to Beichen’s 48 people |
| `eligible` / `opened` / `suggested` | Eligible task, assistant opened, recommendation generated |
| Tasks: `accepted` / `modified` | Accepted, including accepted with edits / edited subset; do not add them |
| `justified_rejection` / `private_note_bypass` / `human_handoff` | Assigned justified refusal, use of private notes, and human handoff; these flags need not be mutually exclusive |
| `completed` / `effective` | Assistant-related workflow completed / completed and effective under this simulation; nonuse does not mean the underlying work was unfinished |
| Interventions: `branch` / `group` / `period` | A / B / C; `target` / `comparison`; `before` / `after` |
| `accepted_or_modified` / `high_consequence_error` | Accepted directly or with edits / consequential error |
| `baseline_work_minutes` / `workflow_minutes` | Baseline comparison time / current workflow work, excluding the next three increments |
| `review_minutes` / `support_minutes` / `intervention_minutes` | Review, support, and allocated intervention time; include all three |
| `tool_cost_units` | Simulated tool-cost units; report separately, without inventing a currency or time conversion |

Boolean fields are numeric 1 / 0. Convert them before summing rather than treating the string `'0'` as true. `modified` is already included in `accepted`. Task and intervention tables represent different populations and must not be concatenated as one trial.

## Calculate and compare

1. The funnel is **340 eligible, 96 opened, 88 recommendations, 41 accepted, 33 completed, and 31 effective**. With eligible tasks as the denominator, the corresponding rates are **28.2% / 25.9% / 12.1% / 9.7% / 9.1%**. The 32.3% from 31 / 96 is conditional on opening, not overall coverage. Do not add the nine modified tasks to the 41 accepted tasks.
2. Compare shifts, refusals, and handoffs. Identify explanations to investigate without inferring motives from clicks. Rejecting an unsuitable recommendation can protect the user; it is not simply a failure to cooperate.
3. Each intervention file has four groups of 20: target/comparison × before/after, totaling 80 rows. Group by both fields. Before and after are not one combined trial.
4. Target groups start at **8/5/4/3** opened/accepted or edited/completed/effective. After intervention, A is **12/6/5/4**, B **14/12/10/9**, and C **16/11/9/7**. Comparison groups end at **9/6/5/4**. Subtracting comparison changes from target changes gives effective-task increments of **0 / 5 / 3**. This is a synthetic comparison, not real causal evidence.
5. Sum `workflow + review + support + intervention` minutes. Target groups start at **434**; after intervention, A totals **519**, B **432**, and C **478**. Comparison groups end at **432**. Relative to each group’s prior period and the comparison change, A adds **87**, B **0**, and C **46 minutes**. Report tool units separately and include C’s **two consequential errors**. More openings do not cancel risk.

## Check your reasoning

B warrants further investigation in this construction, but is not universally preferable. Random assignment, comparable task quality, actual intervention scope, and real costs have not been established. C is a hypothetical access-change branch, not advice to remove access controls. Review covers all generated recommendations; intervention time is allocated across every target-after task; unopened tasks retain baseline workflow work. Counting only successful completions would understate costs.

Submit a funnel, a comparison of outcomes, costs and risks, and a proposed next step with explicit limits. Mark real-world effectiveness as unverified. A peer can check denominators independently.

The next lesson addresses bad news: preserve confirmed and unknown scope, explain protection, and set an update plan.