The Levels Register Falls Too: The 4.6pp Baseline Gap Is Committee Composition (FE Estimate +0.7pp, Driven by One Single-Ministry Cell), the Dose Test Lands Between the Two Theories (+1.3pp/SD, Below the +2.5pp Diluted-Continuity Bar), and the Cohort-3 Stratum Is Null Everywhere

round: R26 (Arc 4, depth round, Season 2)
tasks: Critic 075 Commitments 9b (dose test), 9c (baseline-gap hardening), 9d (cohort 3 stratum), codebook entry
standing_result: pooled opposed-vs-supportive DiD -0.9pp [-3.0, +1.3], prior overturned R25 - UNCHANGED this round
new_this_round: baseline-gap finding overturned (retreat logged); dose test inconclusive-positive below threshold; cohort 3 null
artifacts: knowledge/hand_coding/round_26.jsonl (325 rows, cohort-3 dictionary, written before analysis per C5);
  workspace/r26/{build_cohort3,dose_baseline}.py; workspace/r26/cohort3_panel.csv;
  kr-hearings-data/docs/CODEBOOK.md term-snapshot caveat added; knowledge/retreats.jsonl entry appended
series_note: no ideal points used this round

Rejected Paths

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

• Standardize the dose across the full sample rather than within committee. Rejected because hearing length varies mechanically by committee (multi-day 법사위-style hearings vs single-session ones), so a global SD would load committee size onto the slope; Scout 076 Section 4 fixed within-committee standardization and I follow it, adding raw-count, log, and top-tercile variants as robustness only.
• Rescue the baseline gap by reporting the 여가위-inclusive pooled FE estimate (+1.53pp, p=.057) as "marginal support." Rejected: the placebo levels test shows opposed members also ask +2.23pp more of named non-confirmed agencies, so even the residual gap is generic rather than ministry-specific, and Commitment 9c pre-committed us to dropping the levels register if the FE check failed.
• Code cohort-3 opposition from leg_ruling_status or from bare party names. Rejected live: my first build coded ruling as {국민의힘, 국민의당} and produced 11 supportive units, because PPP members carry their term-start labels 미래통합당/미래한국당 in leg_party. This is the exact R25 hazard, and I hit it myself before catching it (Section 4).
• Headline the log-dose specification (+1.34pp/SD, p=.039), the only dose variant whose interval excludes zero. Rejected as one significant cell among six specifications of the same slope; the R17 multiple-testing debt forbids promoting it, and it enters the survival table as a fragility note.

1. What this round did

Depth only. The standing result (pooled DiD -0.9pp [-3.0, +1.3], N=278, prior overturned) is not re-litigated. I ran the three tests Critic 075 assigned: the within-opposition dose test (9b), the committee-FE hardening of the 4.6pp baseline gap (9c), and the May-2022 cohort-3 stratum under a nominating-president coding (9d), plus the codebook entry. Baselines for 9b and 9c were fixed in Scout 076 Section 4 and Critic 075 Commitment 9c before I computed anything.

2. Predictions written before computing

• 9b (dose): Continuity in diluted form requires a slope of at least +2.5pp per SD of hearing engagement among opposition questioners; its failure condition is a 95% interval excluding +2.5pp and including zero on 150+ opposition units (Scout 076 Section 4). Party-theater predicts a zero slope.
• 9c (baseline gap): If the cohort-1 raw gap of +4.61pp survives committee fixed effects, it becomes the arc's positive finding; if it collapses, it was composition and the paper reports the null alone (Critic 075 Commitment 9c).
• 9d (cohort 3): No pre-set threshold; labeled an Eldes-style ruling-status stratum, descriptive-plus, per Commitment 9d.

3. Survival Table

All estimates committee FE, SEs clustered by legislator, units = legislator-ministry pairs present in both audits. Code: workspace/r26/dose_baseline.py, workspace/r26/build_cohort3.py.

4. The three results in words

9c is the round's headline and it is a retreat. The cohort-1 gap that R25 flagged as the arc's candidate positive finding - opposed legislators devoting 4.6pp more of their audit questions to the confirmed ministry before the hearing - is committee composition. One cell does almost all the work: 여가위, where the confirmed ministry is effectively the whole witness list (mean before-share 0.963) and 7 of 11 units are opposition members. Drop that one committee and the raw gap is +0.61pp; hold committee constant and it is +0.69pp with a CI straddling zero. The pooled FE residual (+1.53pp, p=.057) fails the specificity test: opposed members show a same-signed, larger gap on placebo agencies (+2.23pp), so what little remains is "opposition members ask more questions of named government witnesses generally," not ministry-directed attention. Per Commitment 9c's pre-commitment, the levels register of Scout 076's adjudication paragraph drops, and the retreat is logged in knowledge/retreats.jsonl. Note what this does to the theory adjudication: R25 read the levels gap as discriminating evidence for the party-theater reading. That support is now gone. What survives of the adjudication is the changes register alone - continuity's positive DiD prediction failed, party-theater's zero-DiD prediction held - and the simplest description of levels is that allocation follows committee jurisdiction, full stop.

9b lands between the theories and cannot formally resolve. The dose slope is positive in every specification (six of six), placebo-clean, and roughly +1.3 to +1.8pp per SD - about half of Scout's +2.5pp diluted-continuity bar. Neither pre-set decision rule fires: the survival bar is not reached (point estimates well below +2.5), and the failure condition cannot fire because the pooled interval still includes +2.5pp and the opposition sample is 138 units, short of the 150 Scout specified. The honest verdict is weakened, not killed: if a within-opposition intensity effect exists, it is too small to satisfy the diluted-continuity threshold, but the design at this N cannot push the upper bound below it. The one zero-excluding cell (log dose, p=.039) is one of six specifications and stays a footnote.

9d is null everywhere, and instructive twice over. With the corrected nominating-president coding (opposed = not in {국민의힘, 국민의당, 미래통합당, 미래한국당}), 125 units survive the 원구성 reshuffle in both audits, splitting 50 supportive / 75 opposed - both pooled cells clear the N>=10 guardrail, though all 32 nominee-by-side cells sit below it and stay descriptive. DiD +0.09pp, placebo +0.14pp, baseline gap +0.32pp: nothing moves, even across a full government change in which the "opposed" Democrats went from ruling party (Oct 2021 audit) to opposition (Oct 2022 audit). As a labeled Eldes-style stratum this says the ruling-status flip did not move allocation - consistent with Eldes, Fong, and Lowande (2023), whose party effects concern confrontation style, not targeting. The second lesson is procedural: my first build coded ruling by current party names and produced 11 supportive units instead of 114, because the term-snapshot leg_party field still labels PPP members 미래통합당/미래한국당. The R25 hazard caught its own documenter.

Codebook entry shipped. kr-hearings-data/docs/CODEBOOK.md previously described leg_party as "party at time of speech"; it is a term-start snapshot. The caveat block now documents the 88/97 cohort-2 inversion, the satellite-label problem, and the required recoding rule (president's party at speech date, predecessor labels included).

5. Reproducibility

cd kna-research-agents
python3 workspace/r26/build_cohort3.py    # writes knowledge/hand_coding/round_26.jsonl FIRST, then cohort3_panel.csv
python3 workspace/r26/dose_baseline.py    # 9b dose test + 9c baseline-gap hardening on workspace/r25/analysis_sample_corrected.csv

Dose = total hearing question dyads a legislator directed at the ministry's nominee(s), summed across the 변창흠 and 노형욱 hearings for 국토위 (median 31 questions, IQR 17-51, cohort-1 opposition).

6. Data gaps and limitations

1. The dose test is powered against +2.5pp but not against +1.3pp. MDE at 80% power is ~2.0pp/SD on 138 opposition units. If the true intensity effect is the observed ~1.3pp, no within-arc sample can resolve it; the 22nd-NA hearings (only 4 minister hearings so far) will not add enough units this term.
2. Dose conflates engagement with speaking-time allocation. Question-dyad counts partly reflect committee-assigned question time and multi-day hearing formats, not voluntary intensity. A tone-weighted dose would need the gate amendment Critic already declined to run without a researcher signature.
3. 여가위 is unusable for any share outcome (ceiling at 0.96) and its removal changes raw descriptives materially; future arcs using ministry shares should exclude single-ministry committees by design, not as robustness.
4. Cohort-2 supportive side remains 26 units, so the pooled FE residual gap (+1.53pp) leans on a cell that cannot carry inference alone; I report it only alongside its placebo refutation.
5. The term-snapshot hazard is now documented but not fixed upstream. Until kr-hearings-data ships a time-varying party field, every cross-2022 analysis must hand-recode; both round dictionaries (round_25.jsonl, round_26.jsonl) contain worked rules.

7. What Critic should evaluate

1. Whether the adjudication paragraph must be rewritten null-only. Commitment 9c's pre-commitment says yes: the levels register drops, and the draft's contribution is the pre-registered null plus the composition diagnosis. Scout 076's drafted paragraph explicitly conditioned the levels clause on 9c, so this is an execution of the plan, not a pivot.
2. How to classify the dose result. I propose "weakened, unresolved": all-positive signs and a clean placebo argue against calling it a confirmed zero, but the diluted-continuity bar is not met and the formal failure condition cannot fire at N=138. Does the paper report it as a bounded residual (slope < +2.7pp/SD at 95%) or omit it?
3. Whether the R25 finding row "opposed baseline share exceeds supportive by 4.6pp" should move to overturned in the Findings Status, matching the retreat entry logged this round.
4. Whether cohort 3's null under a full government change deserves a sentence in the draft as an out-of-sample consistency check, given its descriptive-plus cap.

8. Completion checklist

• [x] Two-plus analyses run with results (dose test with 6 specs, FE baseline gap with placebo, cohort-3 build and stratum)
• [x] Key statistics with N throughout (278 / 138 / 194 / 125 units, cluster counts, MDEs)
• [x] Baselines stated before computing (Section 2: Scout 076 +2.5pp bar, Commitment 9c collapse rule)
• [x] Survival Table (Section 3, nine rows)
• [x] Data limitations (Section 6, five items)
• [x] Reproducible code at workspace/r26/*.py; dictionary at knowledge/hand_coding/round_26.jsonl written before analysis (C5)
• [x] N>=10 guardrail: all inferential cells N>=50 units / 26 minimum side; cohort-3 nominee cells (all <10) descriptive only
• [x] Retreat logged (C3) for the baseline-gap finding; codebook hazard shipped
• [x] Critic evaluation points (Section 7)

References

Birkland, Thomas A. 1998. "Focusing Events, Mobilization, and Agenda Setting." Journal of Public Policy 18 (1): 53-74. doi:10.1017/s0143814x98000038

Choi, Jun Young, Sangjoon Ka, Byoung Kwon Sohn, and Jin Man Cho. 2008. "The Executive-Legislative Relationship Reflected in the Prime Minister Confirmation Hearings: A Content Analysis." Korean Political Science Review 42 (2). doi:10.18854/kpsr.2008.42.2.008

Eldes, Ayse, Christian Fong, and Kenneth Lowande. 2023. "Information and Confrontation in Legislative Oversight." Legislative Studies Quarterly. doi:10.1111/lsq.12440

Jeon, Jin Young. 2018. "Analyzing the National Assembly-Government Relationship with Topic Modeling Methods: Focusing on Prime Minister's Confirmation Hearings." Journal of Parliamentary Research 13 (2). doi:10.18808/jopr.2018.2.1

Yoon, Young-Gwan, In-Kyun Kim, and Won-Taek Kang. 2020. "Politics of Confirmation Hearings: What Makes the National Assembly Approve or Reject Candidates for High Office in South Korea?" Korean Political Science Review 54 (2): 85-117. doi:10.18854/kpsr.2020.54.2.004