The lab notebook: what was planted, what the agent read, how it was scored.

# Experience Study Workspace

## Who you are and what this is
You are the experience-study analyst at a life reinsurer. This workspace
contains a cedant's full-period data submission for a 100% quota-share YRT
treaty covering the cedant's Korean whole life and 5-year renewable term
portfolio, January 2019 – December 2025. The reinsurance premium in the
bordereaux is a monthly natural premium on the industry reference basis
(10th Korean experience life table reference net rates by attained age and
sex, applied to the ceded net amount at risk). Amounts are KRW thousands.

## Data
./data/policy_master.parquet
./data/reinsurance_premium_monthly/   (partitioned by year)
./data/reinsurance_claims.parquet
./data/policy_events.parquet
Treat the submission as read-only. Do all work in ./work/ (scratch) and
deliver to ./outputs/.

## Deliverables (all required)
1. ./outputs/cleaned/ — the four tables after your data-quality treatment,
   same schemas and formats as ./data/.
2. ./outputs/findings.json — every data-quality issue you identify and the
   action you took, conforming to ./schemas/findings.schema.json. Use the
   cedant_query flag for items that require confirmation from the cedant.
   Data issues only: methodology documentation (how a basis was derived,
   whole-book reconciliations) belongs in the report's methodology
   section, not here.
3. ./outputs/insights.json — your analytical findings with quantification,
   conforming to ./schemas/insights.schema.json.
4. ./outputs/report.docx — a professional experience-study report following
   ./templates/report_structure.md. Generate it programmatically
   (python-docx is acceptable).
5. ./outputs/retrospective.md — written last, following
   ./templates/retrospective_template.md.

## Working practices
- Follow the playbooks in ./playbooks/ (data quality, analysis,
  reporting) during the corresponding phase of the work.
- Verify your own work. Before finalizing, reconcile: cleaned-data row
  counts against your findings; every number in the report against
  insights.json; exposure and claim totals before vs after cleaning, with
  the movement explained by your findings.
- Every material judgment call gets a written rationale — in findings.json
  for data decisions, in the report for analytical ones. "Material" means
  it could change a reported result.
- State the credibility of what you report. Where claim counts are thin,
  say so and qualify or withhold conclusions accordingly.
- If the data conflicts with itself, decide on a consistent treatment,
  apply it uniformly, justify it, and flag it for the cedant when
  confirmation is needed.
- Before finalizing, run `python tools/validate_outputs.py` — it checks
  that all deliverables exist and that findings.json / insights.json
  conform to their schemas. Fix anything it reports.
- In the retrospective and in any rule you propose, generalize: describe
  patterns and checks, never specific record identifiers.

A cedant submission has arrived in ./data/ — full-period bordereaux for
our 100% quota-share YRT treaty on their Korean whole life and 5-year
renewable term book, January 2019 through December 2025.

Conduct the full experience study: assess and treat data quality, then
analyze mortality experience (actual vs expected on the treaty premium
basis) and lapse experience, across the risk dimensions present in the
data. Produce all deliverables specified in CLAUDE.md.

Work autonomously end to end; there is no one available to answer
questions mid-study, so record anything you would ask the cedant as
cedant-query items. You have up to 4 hours. Begin.

# Report Structure — Experience Study

Format contract for `./outputs/report.docx`. It fixes the section order and
presentation standards only; the methodology is yours to determine, apply,
and describe.

## Required sections, in this order

1. **Executive summary**
2. **Data received & quality assessment** — the issues you found, the
   treatment you applied to each, and the impact on exposure and claim
   totals (before vs after your cleaning).
3. **Methodology (as applied)** — describe what you actually did: exposure
   basis, expected basis, segmentation, credibility handling.
4. **Mortality experience results**
5. **Lapse experience results**
6. **Insights & recommendations**
7. **Cedant query list (appendix)** — all items requiring cedant
   confirmation, consolidated in one place.
8. **Limitations**

## Presentation standards

- Every quantitative claim carries its basis: the exposure and claim count
  behind each mortality figure, the decision count behind each lapse rate.
- The data-quality section precedes all results, so a reader knows what the
  results stand on.
- Cedant queries may be referenced from the body but are consolidated in
  the appendix.
- Every number in the report must be traceable to insights.json or
  findings.json.

# Study Retrospective

## What went well
## What was difficult or slow
(Where did you lose time? What would you check first next time?)

## Judgment calls I am least sure about
(Decision, alternative treatments considered, why you chose yours.)

## Proposed additions to this workspace's practices
(Checks, decision rules, or methodology notes that would have made this
study faster or more reliable. Write each as a GENERAL rule a future
analyst could apply to any submission — pattern-level descriptions only,
no specific record identifiers. For each: the rule, when it applies, and
what prompted it.)

## Open questions
(Anything unresolved, including items already flagged as cedant queries.)