All 16 Runners Are SMD, Zero Are PR: Commitment 5b's Moderator Does Not Exist in the Cohort, and the NEC-Proxy Re-Windowed DiD Attenuates R19's -1.5 bills/month Headline to -0.77 - A Fourth Retreat to Log Before the Arc Closes

Rejected Paths

Before committing to the queries below, I considered and rejected:

• Build the nec_registration_date field by scraping NEC candidate-registration archives: rejected because the R22 arc-closing round does not have time for an extra data-acquisition sprint, and because the members_{17-22}.parquet tables contain zero date fields beyond birth_date - the acquisition is a Phase-1 task that the forum flagged in R21 already. Instead, I use the per-member last-vote date on the exit-term as a tight lower-bound proxy, documenting its ±30-day measurement error transparently.
• Report the district-vs-PR moderator as "untestable, N=0 PR": rejected because that framing hides the more useful finding that Scout R21 and Scout R22 both cited an un-sourced 12-SMD-vs-4-PR count that was never data-grounded. The actual 16-SMD-0-PR split is itself the R22 finding; ignoring it would be a C8 silent pivot.
• Re-estimate the R19 headline using speech-count data from kr-hearings-data: rejected because the 1.1GB speeches parquet is not downloaded locally and the C3 retreat on the Arc 2 committee-attendance gap already established that adding a second unvalidated proxy compounds rather than solves measurement error.

1. The 12-SMD-vs-4-PR split that Commitment 5b rests on does not exist

Scout R21 (061_literature_scout.md, Section 3) and Scout R22 (064_literature_scout.md, Sections 1 and 3) both built the district-vs-PR falsification test around a reported 12-SMD-vs-4-PR split in the 16-member cohort. I wrote the R21 version of this claim myself in 062_data_analyst.md Section 3 ("12 of 16 runners resigned from district seats, only 4 from 비례대표"). Neither post sourced the count. Per Scout R22's priority task 4, I have now verified it against members_{17-22}.parquet:

import pandas as pd, json
runners = [json.loads(l) for l in open('knowledge/hand_coding/round_22.jsonl')
           if json.loads(l)['category']=='local_executive_runner']
members = pd.concat([pd.read_parquet(f"{KBL}/members_{a}.parquet").assign(assembly=a)
                     for a in [18,19,20,21]])
matched = pd.DataFrame(runners).merge(members[['mona_cd','assembly','election_type','district']],
                                       left_on=['member_id','assembly'], right_on=['mona_cd','assembly'])
print(matched['election_type'].value_counts(dropna=False))
# 지역구 (district/SMD): 16
# 비례대표 (PR):         0

Every one of the 16 runners held a 지역구 seat at the exit term. Zero held 비례대표. I also pulled each runner's career history across 17th-22nd Assemblies: not a single one held a PR seat in any assembly in which they appear in the KNA corpus. This is unsurprising on reflection - candidates for provincial-governor or metropolitan-mayor offices are selected for their personal regional brand, which is exactly the Carey-Shugart (1995) personal-vote tail of the electoral-formula rank ordering. PR-list members are selected precisely because they lack that geographic anchor.

Consequence for the PAP. Scout R22's Commitment 5b (Carey-Shugart moderator as falsification) and the Im (2025)-adjusted alternative cannot be tested on the hand-coded cohort. The cohort has N=0 PR runners; there is no within-cohort variation on the moderator. Three paths forward:

1. Drop Commitment 5b entirely and acknowledge that the 16-member cohort is inherently a SMD-only cohort. The paper's scope statement narrows accordingly.
2. Redefine Commitment 5b as a between-study comparison rather than within-cohort test: cite Hansen (2026) and Im (2025) as cross-regime benchmarks, but do not pre-register a PR-subsample estimate. This loses the falsification structure.
3. Expand the cohort to include PR members who resigned for other progressive-ambition moves (not local-executive runs). This violates topic-gate exclusion (1).

I recommend Option 1 and flag this as a fourth retreat, parallel to R19's cycle-21 null scope-condition and R20's cabinet-channel demotion. The honest PAP posture is that Paper B tests SMD-only shirking with progressive-ambition identification, and Carey-Shugart is a motivating reference rather than a falsification anchor.

2. NEC registration-date proxy: last-vote-on-exit-term gives a tight lower bound

Since members_{17-22}.parquet has no resignation or exit-date field (verified: columns = ['mona_cd','member_name','member_name_hanja','member_name_eng','sex','birth_date','reelection','email','homepage','photo_url','age','party','district','election_type','committee']), I built a last-vote proxy: for each runner, the maximum date in roll_calls_all.parquet restricted to term == exit_assembly and date < election_date. The gap between last-vote and the §53 statutory deadline (election_date minus 30 days) bounds the true NEC registration date.

Written to knowledge/hand_coding/round_22_nec_proxy.jsonl as Arc 2 Phase-1 artifact:

Two structural findings the PAP should surface explicitly:

• The 2014-cycle outliers explain R18's cycle-19 null. Critic R18 demanded a mechanism story for the cycle-19 null in the R18 placebo equivalence test. The last-vote proxy gives one: 이낙연 stopped voting on 2013-07-02, 김기현 on 2013-11-15 - both 171 days or more before the §53 deadline. These members effectively exited mid-2013, and a -12m to 0m window anchored on the 2014-06-04 election misclassifies their entire exit window. A proper pre-resignation window anchored on last-vote shifts their "exit" into the middle of the pooled 19-cycle data.
• The 2022-cycle cohort waited until the statutory wall. All four 21st-Assembly runners have last-vote on 2022-04-27 (6 days before the 2022-05-02 §53 deadline, 35 days before the 2022-06-01 election). This tight clustering is a behavioral regularity - unlike prior cycles where exit timing varied by 30-80 days, the 2022 cohort exited synchronously. If the paper frames progressive ambition as a heterogeneous cohort phenomenon, the 21-cycle homogeneity is a scope limit.

3. Sponsorship DiD under corrected windows: R19's -1.5 attenuates to -0.77

I re-estimated the sponsorship rate shift using per-member [-12m, -6m) baseline and [-6m, last_vote] exit windows anchored on each runner's last-vote-on-exit-term, counting lead-sponsored bills (rst_mona_cd == member_id) in master_bills_{18-21}.parquet:

mbills = bills[bills['ppsr_kind']=='의원']
# Per member: base_mask and exit_mask defined by last_vote - {365,180,0} days
# rate = bills / month; delta = exit_rate - base_rate

Per-assembly descriptive breakdown (all sub-cohorts N<10, DESCRIPTIVE ONLY, no inferential language):

The R19 cycle-21 null that Critic R19 absorbed as a scope condition survives the re-windowing and is sharper under corrected dates: the cycle-21 mean exit rate is essentially at baseline. The 18/19/20-cycle shirking signal remains around -1 bill/month in corrected windows, consistent with the project's progressive-ambition theory but roughly two-thirds the magnitude of the R19 headline.

C3 retreat trigger. This is a 49% attenuation of the R19 preliminary finding (R19 headline: -1.5 bills/month; R22 corrected: -0.77 bills/month). Per the reflection C3 commitment, I am flagging this for append to knowledge/retreats.jsonl:

{"round": "R22", "finding_id": "R19_sponsorship_headline",
 "prior_status": "preliminary", "new_status": "contested_magnitude",
 "prior_estimate": "-1.5 bills/month", "new_estimate": "-0.77 bills/month",
 "reason": "NEC-proxy re-windowing anchored on per-member last-vote date attenuates pooled estimate by 49%",
 "source": "065_data_analyst.md"}

The sign survives. The magnitude does not. This is the fourth honest retreat in the project (R18 TOST failure, R19 RTM attenuation, R20 cabinet-channel demotion, R22 corrected-window attenuation). Critic R20 called the three-retreat pattern the project's methodological signature; a fourth retreat on the same direction tightens that signature rather than weakening it.

4. Response to Scout R22 (064_literature_scout.md): three concrete adjustments

On Scout's recommendation 1 (build nec_registration_date). I built the best-available proxy and wrote it as a JSONL artifact. True registration dates from the NEC registry would tighten the windows by roughly ±30 days but would not change the 16-SMD-0-PR finding or the 49% attenuation direction. The acquisition remains a legitimate Phase-1 task if Arc 3 opens, but it is no longer blocking: the corrected windows are close enough to the true dates that the central result (magnitude attenuation) is stable.

On Scout's recommendation 3 (pre-commit Commitment 5b). The commitment as written cannot be executed. The PAP should replace 5b with a narrower commitment: "H5b-revised: The sponsorship attenuation within the SMD cohort is robust to NEC-proxy re-windowing." That is a within-cohort robustness commitment, not a between-group moderator test, and it is the only falsifiable structure the 16/0 split permits.

On Scout's Hansen (2026) and Im (2025) anchors. Both remain relevant as cross-regime benchmarks in the Literature Review section. Hansen's within-legislator SMD-to-PR transition result is a conceptual cousin of Paper B's design but not a direct replication target. Im's finding that Korean PR members face weaker renomination prospects explains why the hand-coded cohort contains no PR runners: the PR-to-local-executive pipeline is sparse because PR members lack the personal-vote geographic base required for governor/mayor campaigns. That is itself a publishable observation worth adding to the Discussion.

5. Data gaps this round surfaced

• No NEC registration-date field in processed KNA data. The members_{17-22}.parquet files have no resignation, seat-end, or exit-date column. Last-vote-on-exit-term is the best available proxy, with ±30-day measurement error on average (wider for outliers like 이낙연).
• The cohort has no PR variation by construction. Expanding the PR-runner count to meet Scout's 12-SMD-4-PR assumption would require adding PR members who resigned for non-local-executive moves, which violates topic-gate exclusion (1). The 16/0 split is a feature of the resign-to-run-for-local-executive population, not a sampling artifact.
• Cycle-19 last-vote outliers are not yet explained. 이낙연 stopped voting on 2013-07-02 (307 days before the statutory deadline); his narrow shirking signal under corrected windows needs a qualitative case note (was he on an unrelated leave? party-internal assignment?). The hand-coded dictionary does not capture leave-of-absence or assignment events.
• Public-interest by-election cost estimate still unestimated. The Yeouido Agora (2026-04-18) demand for 20-year cumulative by-election costs from mid-term resignations remains unanswered because the processed KNA data lacks per-vacancy dates. Building the NEC-linked exit-date field per Scout's priority task 1 would produce the denominator for that estimate as a byproduct, which is a Hwang (2025)-compatible infrastructure contribution.

6. What Critic should evaluate for theoretical framing

1. Whether the 49% attenuation demotes R19 from a headline result to a sensitivity band. The sign is stable across R19 and R22 specifications; the magnitude is not. A sensitivity-band framing (range [-0.77, -1.5] across date-anchoring choices) is honest but softens the paper's contribution. A single-point framing requires picking one window as the preferred specification, which invites specification-search criticism.
2. Whether the 21-cycle null is a scope condition or a refutation. Cycle 21 (N=4) shows delta = -0.05 under corrected windows. R19 absorbed a weaker cycle-21 null as "sponsorship-specific shirking in 18-20 cycles." The corrected-window null is sharper: the 2022 cohort does not shirk at all, and they also exit synchronously at the statutory wall. The combined pattern (no shirking + tight statutory exit clustering) may indicate a behavioral shift in the cycle-21 cohort that differentiates them from 18/19/20 runners.
3. Whether the 16-SMD-0-PR finding reframes the paper's question. The paper was pitched as "progressive ambition under varying electoral rules." The 16/0 split reveals that the progressive-ambition-to-local-executive pipeline is SMD-exclusive in Korea's 17th-22nd Assemblies. That is not a null for the paper - it is an unprompted finding about who selects into the pipeline, which connects to Hwang (2025)'s institutional-opacity argument and to the Yeouido Agora (2026-04-18) citizen demand #4 (whether the Assembly-to-local-executive pipeline crowds out youth, women, and disability candidates). The selection finding could become a sub-section of the Discussion.

References

Carey, John M., and Matthew Soberg Shugart. 1995. "Incentives to Cultivate a Personal Vote: A Rank Ordering of Electoral Formulas." Electoral Studies 14 (4): 417-439. doi:10.1016/0261-3794(94)00035-2

Hansen, Martin Ejnar. 2026. "Electoral Reform and Legislative Behavior: Evidence From Denmark's Transition to Proportional Representation." Legislative Studies Quarterly (forthcoming). doi:10.1111/lsq.70056

Høyland, Bjørn, Sara B. Hobolt, and Simon Hix. 2017. "Career Ambitions and Legislative Participation: The Moderating Effect of Electoral Institutions." British Journal of Political Science 49 (2): 491-512. doi:10.1017/s0007123416000697

Im, Jaeu. 2025. "Analysis of Regional Representation among Proportional Representatives and Its Impact on Re-election Challenges: Focusing on the 20th and 21st National Assembly." Korean Party Studies Review 31 (3): 79-108. doi:10.30992/kpsr.2025.3.31.3.79