The Learning Prior Fails at Its Premise: There Is No First-Term Passage Gap to Close - the Year-1 Gap Is +0.3pp, the Interaction Is Zero at an MDE That Binds, and the One Significant Trajectory Runs the Wrong Way, on the Absorption Channel

round: R28 (Arc 5, opening round, Season 2 - prediction test)
prediction_tested: Scout 082 Section 4 (year-1 first-term gap -2 to -5pp; gap halves by year 3;
  first_term x proposal_year interaction positive)
population: member-sponsored law bills (의원발의 법률안), 17th-22nd NA, N=93,572 bills, 1,128 sponsor clusters
merge_key: rst_mona_cd -> members mona_cd (uid, homonym-safe per R27 standing policy; validate='m:1' asserted)
artifacts: workspace/r28/{build.py, analyze.py, robust.py, bills_panel.csv}
series_note: no ideal points used (arc exclusion 2)

Rejected Paths

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

• Defining first-term status from cumulative bill history instead of the reelection field. Rejected: reelection (선수) is the register the Korean literature itself uses, is complete for all 1,933 member-terms, and inferring tenure from bill records would conflate sponsorship activity with the treatment.
• A survival/hazard model of time-to-passage instead of the LPM on passage within proposal year. Rejected: the signed falsifier names the first_term × proposal-year interaction on passage rates; a hazard model changes the estimand and would let the prior escape its own test. The 12-month-horizon outcome covers the censoring concern the hazard model would address.
• Restricting to bills where the first-termer is sole or lead drafter (small cosponsor count). Rejected for the headline: it conditions on a post-treatment choice (coalition size). It belongs in R29 depth if the null needs stress-testing.
• Excluding the 22nd Assembly upfront because its term is in progress. Rejected: the within-proposal-year design gives both groups identical time-to-expiry, so the 22nd adds clean year-1/year-2 cells; I instead report the drop-22nd specification as robustness (it changes nothing).

1. Baseline, written before computing

From Scout 082 and the signed arc prior, before any estimation I committed to: (a) pooled member-bill passage rate (원안가결+수정가결) in the low-to-mid teens; (b) a year-1 first-term gap of -2 to -5pp; (c) the learning prior: a positive first_term × proposal_year interaction large enough to halve the year-1 gap by year 3 - given (b), an implied slope of +0.5 to +1.25pp per year. The falsifier: interaction indistinguishable from zero or negative, with the within-year censoring control in place.

2. Observed

export KBL_DATA=$KBL_DATA
python3 workspace/r28/build.py    # 93,572 member law bills, 17th-22nd, uid-merged to members
python3 workspace/r28/analyze.py  # descriptives + headline LPM, SEs clustered by sponsor
python3 workspace/r28/robust.py   # decomposition, member-level weights, absorption channel, MDE

Proposal year = years since term start (1-4) from ppsl_dt; outcome primary = 원안+수정가결 (strict); robustness = absorption-inclusive (+대안반영폐기, 수정안반영폐기) and passage-within-12-months. Headline model: passed ~ first_term × prop_year + assembly FE + party bloc + election_type + committee FE, SEs clustered on sponsor uid (1,128 clusters). 494 bills (0.5%) drop for missing committee_nm.

The premise fails first. The pooled strict passage rate is 6.2%, not low-to-mid teens (the teens figure matches the absorption-inclusive rate of 30.9%, or the 17th Assembly's strict 12.2% before the proposal explosion diluted rates - a definitional point Paper E must fix against Ka 2025a). And the year-1 gap barely exists: raw -0.70pp (first-term 6.83% vs re-elected 7.53%; N=10,323 vs 24,733), and +0.27pp [-0.58, +1.12] with covariates. The entire predicted range of -2 to -5pp lies outside the confidence interval.

The interaction is zero. Strict outcome, full specification: -0.06pp/yr [-0.61, +0.48], p=.82, N=93,078. The MDE (80% power) is 0.78pp/yr, which sits inside the prior's implied +0.5 to +1.25pp/yr band: a learning slope in the upper half of the predicted range would have been detected. The censor-clean 12-month outcome gives the same null (-0.04pp/yr); dropping year-4 proposals (heaviest censoring) gives -0.05pp/yr; dropping the in-progress 22nd gives +0.08pp/yr. Per assembly, the interaction is positive nowhere except trivially (A19 +0.50, A20 +1.22, neither excluding much) and negative in A17 (-2.74pp/yr, driven by a year-4 first-term collapse to 5.9% vs 14.3%, ft N=322). The member-level equal-weight version (each member-year cell counted once, 6,264 cells) leans negative: -0.64pp/yr, p=.064.

3. Baseline vs Observed

Verdict on the prediction: the learning prior is overturned, at its premise. First-term members do not pass member bills at a lower rate at any point in the term; there is no gap for learning to close. The Korean position-and-timing account (Kim and Lee 2023; Ka 2025b) predicted a zero interaction and is confirmed - and more than confirmed, because even the constant level offset it would have tolerated is absent. Padró i Miquel and Snyder's (2006) learning-by-doing, extended within-term, finds no purchase here: whatever tenure buys in North Carolina, the first Korean term buys nothing measurable in direct passage.

4. The anomaly: first-termers fall behind on the absorption channel

The single significant trajectory in the round runs opposite to learning. Isolating absorption (대안반영/수정안반영 폐기, excluding direct passage): interaction -1.51pp/yr (SE 0.41, p<.001) on a base rate of 24-29%. First-termers match re-elected members on absorption in year 1 (+0.13pp raw) and then lose ground (year 2 -3.2pp). The sign is negative in five of six assemblies and significant in the 19th (-2.37pp/yr) and 20th (-1.94pp/yr) individually - not one-assembly-driven. Given R11's finding that absorption is the chair's constructive-bundling channel, the natural reading is that access to 대안 packages is positional and accrues to incumbents as the term progresses - but per exclusion criterion 3, no mechanism is promoted before Critic rules on whether this trajectory becomes R29's depth target. It is reported as preliminary.

5. Data gaps and limitations

1. By-election entrants have a miscoded clock. months_in counts from term start, but a first-termer seated via 보궐선거 in year 2 starts their personal clock later; this attenuates any true learning slope. The members files carry no seat-entry date - quantifying this needs an election-registry merge (candidate: NEC data from the Arc 2 pipeline).
2. Attribution is representative-sponsor only. rst_mona_cd credits the 대표발의자; party-drafted bills fronted by first-termers add noise to treatment.
3. "Passage rate" is definition-dependent by a factor of five (6.2% strict vs 30.9% absorption-inclusive). Any comparison to Ka's (2025a) member-level rates must state the definition; the literatures may be talking past each other on this number.
4. The 22nd Assembly contributes only years 1-2 (term in progress, data through 2026-03); all headline results are robust to dropping it.

6. What Critic should evaluate

1. Is "overturned at the premise" the right closure? The falsifier condition (interaction ≈ 0 with censoring control) is met, but the deeper failure is that Scout's -2 to -5pp premise gap does not exist. Does the arc now report the level-gap null as its spine (per the signed falsifier's fallback: "the arc then reports whether the level gap itself exists"), making this a second null-paper candidate?
2. Whether the absorption-channel divergence (Section 4) is admissible as R29's depth target or whether it trips exclusion 3 (mechanism before gap) / drifts toward the R11 bundling arc.
3. The MDE asymmetry: 0.78pp/yr detects the upper prior band but not the lower (+0.5). Should the draft state the null Rainey-style against the +1.25 upper bound only?
4. Whether the strict-vs-absorption definitional gap (limitation 3) deserves a named paragraph - the Korean seniority literature's "mixed" level effects may partly be a definition artifact.

7. Completion checklist

• [x] 2+ analyses run with results (build, analyze, robust: 3 scripts, 10+ specifications)
• [x] Key statistics with N (93,572 bills; 1,128 clusters; per-cell Ns in Section 2-4)
• [x] Baseline stated before computing (Section 1; arc prior + Scout's numbers)
• [x] Baseline vs Observed table (Section 3, opening-round format)
• [x] Data limitations (Section 5, four items)
• [x] Reproducible code for every finding (workspace/r28/, commands in Section 2)
• [x] N≥10 guardrail: smallest inferential cell is A17 year-4 first-term, N=322 bills
• [x] uid merge with m:1 validation asserted (R27 homonym policy)
• [x] Critic evaluation points (Section 6)

References

Ka, Sangoon. 2025a. "Analyzing Legislative Activities and Behavior of National Assembly Members: Focusing on the Number of Bill Proposals, Bills Passed, and the Passage Rate." Journal of Research Methodology 10 (3). doi:10.21487/jrm.2025.11.10.3.1

Ka, Sangjoon. 2025b. "Analysis of Lapsed Bills Within the Institutional Time Structure of the National Assembly." Korean Party Studies Review 24 (2). doi:10.30992/kpsr.2025.6.24.2.5

Kim, Yanghun, and Dongseong Lee. 2023. "An Analysis of the Impact of Bill Initiators' Position in Subcommittees on the Passage of Bills: From the Perspective of the Distributive Benefits Theory." Korean Political Science Review 57 (1). doi:10.18854/kpsr.2023.57.1.005

Padró i Miquel, Gerard, and James M. Snyder. 2006. "Legislative Effectiveness and Legislative Careers." Legislative Studies Quarterly 31 (3): 347-381. doi:10.3162/036298006x201841

Rainey, Carlisle. 2014. "Arguing for a Negligible Effect." American Journal of Political Science 58 (4): 1083-1091. doi:10.1111/ajps.12102

Volden, Craig, and Alan E. Wiseman. 2014. Legislative Effectiveness in the United States Congress: The Lawmakers. New York: Cambridge University Press. doi:10.1017/cbo9781139032360