1. Introduction

When loyalty and professionalism pull the autocrat in opposite directions, where does the promotion machinery actually move? The civil-military and authoritarian-politics literatures have long described the trade-off in the abstract — between coercive capacity and coup-proofing [@quinlivan1999coup; @talmadge2015dictators; @greitens2016dictators], and between guarding against external versus elite domestic threats [@mcmahon2015guardianship; @paine2022guardianship] — but individual-level evidence inside an active great-power autocracy is rare. Mattingly (2024 AJPS) provides exactly that: a two-table tripod on Chinese People's Liberation Army officers in which career ties to the current CMC Chairman amplify in promotions during 1990–93 and 2012–15 (the post-Tiananmen and post–Bo Xilai windows), and combat experience amplifies during 2000–02 (the Belgrade-bombing-aftermath U.S.–China tension window). The paper reads as a structural extension of Roessler-style coup-proofing logic to a great-power autocracy and as the first individual-level test of the foreign–domestic threat dilemma in a revolutionary one-party regime [@mattingly2024party].

This paper is a substantive-validity replication. Every published number reproduces. All 18 headline coefficients across Tables 1, 2, and 3 recover exact to three decimals from the deposited R code, and the Figure 2 marginal effects across Deng, Jiang, Hu, and Xi reproduce in sign, magnitude, and significance. The numerical record is sound. What the audit modifies is the inferential perimeter around the trade-off claim.

The replication finds that the two halves of the trade-off survive at very different rates. The loyalty / domestic-threat half (Tables 1–2 and Figure 2) is robust on every forensic margin run here. Leaving any one birth-cohort decade out leaves the interaction in [0.106, 0.210]. Shifting the domestic window two years earlier collapses it to β = −0.006 (no anticipation signal). Dropping all years after the 2014 Xi-era anti-corruption purge — a candidate concurrent shock — attenuates the headline; the coefficient grows to β = 0.190. Bonferroni-3 across the paper's three headlines preserves significance (Bonf p = 1.4 × 10⁻³). A cohort-aware Sun–Abraham estimator with the 1980s cohort as reference returns β ∈ [0.27, 0.38] for 1990–94 — about three times the pooled two-way fixed-effects headline of 0.129. A leave-one-leader-era-out test refines the result further: with the 2012–15 window dropped, the post-Tiananmen window alone delivers β = 0.203, p = 7.9 × 10⁻⁵; with the 1990–93 window dropped, the post–Bo-Xilai window alone delivers β = 0.069, p = 0.18. The pooled headline is driven by post-Tiananmen identification; Xi-consolidation is additive but not independently identified at conventional thresholds. The headline is, on this margin, conservative — and concentrates its identification leverage on the larger of the two coded shocks.

The professionalism / foreign-threat half (Table 3) reproduces exactly but is single-window-dependent. The binding identification concern is cohort aging: the combat-experienced cohort, dominated by Sino-Vietnamese-war veterans of 1979 and 1980s border-conflict participants, mechanically aged into senior-grade eligibility precisely during 2000–02 [@fravel2019active]. With one 3-year foreign-threat window and no out-of-window placebo at a different point in the cohort life cycle, the panel cannot separately identify the cohort-aging mechanism from the foreign-threat mechanism (M6 NOT REFUTED). The cutoff sensitivity sweep moves the foreign-threat window through {−2, −1, 0, +1, +2} year shifts and produces β ∈ {0.039, 0.055, 0.074, 0.054, 0.038} in a symmetric, smooth, single-peaked pattern; only the paper's exact 2000–02 window achieves p < 0.05. The pattern is consistent with a real but temporally narrow effect and with a cohort-eligibility correlation, and the panel cannot adjudicate between them. Bonferroni-3 across the three headlines also fails for combat × foreign (raw p = 0.043 → Bonf p = 0.129).

The design move that carries the headline, surfaced by a substantive replication contrast against an uninformed rebuild, is the recoding of the loyalty marker from a static "officer i has a tie to some paramount leader" to the time-varying "officer i has a tie to the person currently sitting as CMC Chairman in year t." That recoding is what makes individual fixed effects feasible and what gives the within-officer interaction its sharpness; the headline rests on it rather than on the two-way FE machinery alone.

Section 2 sets out the original design and the audit perimeter; §3 records the cell-by-cell reproduction; §4 walks through the forensic and alternative-mechanism battery; §5 develops the time-varying-tie recoding; §6 collects sensitivities; §7 closes on the asymmetric trade-off reading.

2. The original paper, the data, and the audit perimeter

Mattingly (2024) constructs an officer-year dataset of senior PLA officers from open biographies, codes prior career posting overlaps with each of Deng, Jiang, Hu, and Xi, and asks two interaction questions on the post-Mao panel: does the time-varying tie to the current CMC Chairman amplify promotion probability during domestic-threat windows, and does post-1949 combat experience amplify promotion probability during foreign-threat windows? Three estimands map onto three tables. Table 1 is a cross-sectional OLS of "ever promoted to general-grade / CMC" on time-invariant career markers with birth-cohort fixed effects. Table 2 is an officer-year linear probability model with individual fixed effects, year fixed effects, and a Tie × DomesticThreat interaction. Table 3 is the symmetric specification with Combat × ForeignThreat. Domestic threat = 1 for 1990–93 and 2012–15; foreign threat = 1 for 2000–02. Standard errors cluster by individual.

The data construction layer is described as 1,295 officers and over 12,000 career postings. The panel that fits Tables 2–3 sits on a subset of that construction dataset: 720 officers and 4,786 officer-years, restricted to the cmc-promotable subsample within the post-1979 window (§4.3 returns to this).

The audit perimeter has four batteries. The first is cell-by-cell verification: every coefficient, standard error, sample size, and R² in Tables 1–3, plus the Figure 2 marginal effects. The second is theory-motivated robustness: leave-one-cohort-decade-out (F1), cutoff sensitivity on the domestic and foreign windows (F2, F3), specification curve over the FE × covariate grid (F4), influence drops on the top 5% by within-FE residual (F5), pre-trend leads with cohort-aware event study (F9), Bonferroni-3 across the three headlines (F10), and leave-one-leader-era-out on the domestic side (F11 — drop 1990–93, drop 2012–15, drop both). The third is alternative mechanisms: the Xi-era anti-corruption purge as a concurrent shock (M2 / M2b), pre-shock anticipation (M5), commissar/operational track selection (M1), and a cohort-aging channel on the foreign-threat side (M6). The fourth is a data-coding and R-programming sweep over the panel construction (D1, D2), plus a Sun–Abraham cohort-aware diagnostic (S1) appropriate to a calendar-shock design with no never-treated unit.

3. Reproduction

The deposited R code (Harvard Dataverse doi:10.7910/DVN/R3XPEJ) reproduces every published cell when run under R 4.4 against the bundled CSVs. Eighteen of eighteen headline coefficients across Tables 1, 2, and 3 recover exact to three decimal places, with identical clustered standard errors and identical sample sizes. The Figure 2 by-leader marginal effects on Deng, Jiang, Hu, and Xi (panels a and b) reproduce exactly: Deng β = 0.203 for general-grade and 0.480 for CMC, Xi β = 0.266 / 0.351, with Jiang and Hu null in both panels. The R² values across all six columns of Table 1 (0.029, 0.071, 0.160, 0.061, 0.039, 0.231) match cell-for-cell.

Reproduction verdict: complete. The deposited code is a clean R-only pipeline keyed on three CSVs (bio_data.csv, career_data.csv, key_positions_data.csv) and four scripts. No proprietary data, no missing intermediate files, UTF-8 with embedded Chinese text in officer biographies. The remainder of the paper develops what the recovered coefficients do and do not support under audit.

4. Forensic audit

The audit ran the F-, M-, D-, and S-batteries on the Tables 2 and 3 headline specifications. The summary verdict per battery appears in Table 2 of this paper; the prose unpacks each test.

Table 2. Audit verdicts on the two headline interactions.

4.1 Domestic-threat side: robust

Leaving each birth-cohort decade out one at a time and re-fitting the Table 2 column 1 specification produces β ∈ [0.106, 0.210] across the seven drops, with the minimum p-value at 0.009. The interaction does not depend on any single cohort. The leave-one-cohort minimum sits 18% below the headline and the maximum 63% above; neither boundary inverts the sign or kills significance.

The Xi-era anti-corruption purge (2014 onward) is the most natural candidate for a concurrent shock that could mechanically inflate a "tie" coefficient — purges tend to elevate connected officers by removing rivals. Restricting the panel to officer-years before 2015 and re-fitting the headline specification returns β = 0.190 (p = 6.9 × 10⁻⁶, n = 4,411). The coefficient is attenuated by including, not driven by, the purge years. A second concurrent-shock variant restricting to officers commissioned before the 1988 rank reform returns β = 0.176 (p = 2.5 × 10⁻⁵, n = 4,353). Both checks refute the concurrent-shock alternative.

The anticipation channel — leaders pre-positioning loyalists before a coded shock — would predict that shifting the domestic window two years earlier (to 1988–91 / 2010–13) preserves or amplifies the coefficient. The shifted-window estimate is β = −0.006, p = 0.87. There is no anticipation signal; the headline is post-shock, not pre-shock.

The influence drop trims the top 5% of officer-years by within-FE residual magnitude (proxy for high-leverage observations). The trimmed estimate is β = 0.098 against the baseline of 0.129, a 24% attenuation, with p = 0.009. The headline survives leverage adjustment.

The cohort-aware event-study specification (Sun–Abraham, with the 1980s cohort as reference) returns β ∈ [0.272, 0.383] for the 1990–94 post-Tiananmen window. The pre-shock placebo years 1984–88 sit in β ∈ [−0.016, 0.111] and are jointly insignificant, satisfying the cohort-aware parallel-trends check. The cohort-aware coefficient is approximately three times the pooled two-way FE headline of 0.129. The TWFE design therefore understates the loyalty effect during the post-1989 window relative to a cohort-aware estimator. This is a positive-direction sensitivity: the headline is conservative on the domestic-threat margin.

The pre-trend leads from −5 to −2 (cohort-aware event study, 1989 reference) return a joint Wald F = 6.01 on df = 4, p = 0.198 — does not reject parallel trends. The −5 lead is individually β = 0.123 with p = 0.046, borderline against an isolated reference, but the joint test passes the standard threshold. Bonferroni-3 across the paper's three headline interactions (Table 1 col 6 tie main, Table 2 col 1 tie × domestic, Table 3 col 1 combat × foreign) preserves significance for tie × domestic (raw p = 4.65 × 10⁻⁴ → Bonf p = 1.4 × 10⁻³).

A leave-one-leader-era-out test (F11) decomposes the headline by which of the two coded shocks identifies it. Restricting the panel to officer-years not in 2012–15 (so the only treated years are 1990–93) returns β = 0.203, SE = 0.051, p = 7.9 × 10⁻⁵. Restricting the panel to officer-years not in 1990–93 (so the only treated years are 2012–15) returns β = 0.069, SE = 0.052, p = 0.18. Restricting the panel to officer-years not in either window (a placebo with no treated years) leaves the interaction undefined, as expected. The headline is driven by the post-Tiananmen window. The post–Bo-Xilai window contributes additively to the pooled estimate but does not independently identify the interaction at α = 0.05 or α = 0.10. This refines the loyalty-side reading: the within-officer evidence that CCP leaders substitute toward officers tied to the currently sitting CMC Chairman during domestic-threat windows rests squarely on the post-Tiananmen episode and is consistent with — but does not independently demonstrate — the same pattern in the Xi-consolidation period [@shih2008factions].

4.2 Foreign-threat side: temporally localized, with a cohort-aging confound

The Combat × Foreign-threat interaction reproduces exactly at β = 0.074, p = 0.043. The binding identification concern is alternative mechanism M6 (cohort aging), with cutoff sensitivity (F3) and multiplicity (F10) as supporting evidence on the same temporal localization.

M6 — cohort aging. The combat-experienced cohort that powers the Combat marker is dominated by Sino-Vietnamese-war veterans of 1979 and 1980s border-conflict participants [@fravel2008strong; @fravel2019active]. That cohort mechanically aged into senior-grade eligibility during the late 1990s and very early 2000s — the typical career arc from junior combat command in 1979–84 to deputy-MR-grade eligibility runs roughly twenty years. The Belgrade-bombing-aftermath foreign-threat window of 2000–02 sits precisely on the eligibility-onset moment for that cohort. With one 3-year foreign window and no out-of-window placebo at a different point in the cohort life cycle, the panel cannot separately identify the cohort-aging channel from the foreign-threat channel. The leave-one-cohort battery on Table 2 keeps the domestic-side coefficient positive across all drops, but it cannot adjudicate the foreign-threat side because every cohort-decade contributes the same combat-veteran composition to the eligibility pool during 2000–02. The Sun–Abraham cohort-aware estimator that strengthens the domestic side does not apply symmetrically here — Sun–Abraham requires variation in event timing across cohorts, which the panel offers for the post-1989 / post-2012 domestic events (cohorts cross those calendar dates at different ages and ranks) but not for the single 2000–02 foreign window (the combat-experienced cohort hits eligibility once, in that window).

F3 — cutoff sensitivity, read in light of M6. The cutoff sensitivity sweep slides the foreign window through five 3-year placements: 1998–2000, 1999–2001, 2000–02 (the paper), 2001–03, and 2002–04. The point estimates are {0.039, 0.055, 0.074, 0.054, 0.038}. Only the paper's exact 2000–02 window reaches p < 0.05; the four neighboring windows return p ∈ [0.13, 0.32]. The pattern is symmetric around the paper's window — single-peaked at the center, smooth on both sides. That shape is observationally consistent with two readings: a real but temporally narrow effect (the early Bush-era U.S.–China tensions following the 1999 Belgrade embassy bombing and the April 2001 EP-3 incident produced a sharp, brief shift toward professionalism that did not bleed into adjacent years) and a cohort-aging correlation tracking eligibility-onset (β peaks where the cohort's promotion-eligible mass peaks). The audit cannot adjudicate between these two readings with the public data, and both are consistent with the foreign-threat coefficient generalizing only as far as the 2000–02 window itself.

F10 — multiplicity. Bonferroni-3 across the three headlines fails for combat × foreign. Raw p = 0.043, Bonferroni-3 p = 0.129. The interaction does not survive a 3-test family-wise correction at α = 0.05 or α = 0.10. The other two headlines in the family (tie main, tie × domestic) survive comfortably, so the family-wise burden is not unusually punitive — it is the combat × foreign cell that sits closest to the conventional threshold even before correction.

4.3 Sample-size disclosure

The published headline of "1,200 officers and 12,000 career appointments" describes the construction dataset assembled from open biographies. The panel that fits Tables 2 and 3 is restricted to the cmc-promotable subsample within the post-1979 window: 720 officers and 4,786 officer-years. Table 1's cross-sectional OLS uses 755–779 officers depending on the column's available controls. The within-officer interaction estimates in Tables 2–3 are therefore based on the 720-officer panel, not the 1,200-officer construction set — a distinction worth recording for readers calibrating effective sample size against the headline.

5. The time-varying recoding of the loyalty marker

The substantive replication ran a parallel "blind rebuild" of the empirical strategy — an officer-year LPM on tie and combat markers — that explicitly rejected officer fixed effects on the reasoning that "Tie and Combat are largely time-invariant for an officer; officer FE would absorb the regressor of interest." Under the rebuild's design, the markers are intrinsic biographical traits (officer i served with someone who later became paramount leader), and individual FE indeed absorb them. The rebuild therefore opted for leader-plus-rank-plus-branch fixed effects and identified the interaction off cross-officer variation within leader-rank cells.

Mattingly (2024) recodes "tie" relationally rather than intrinsically. The variable is Tie_{it} = 1 if officer i has a prior posting overlap with the person who is currently CMC Chairman in year t. As Deng → Jiang → Hu → Xi rotate through, the same officer's tie status flips on and off. An officer who served under Xi in Fujian carries Tie = 0 from 1978 through 2011 and Tie = 1 from 2012 forward. The panel column cmc_chair_connection_current carries this time-varying indicator, distinct from the static cmc_chair_connection that codes prior overlap with any leader.

Once tie is relational, individual FE go from absorbing the regressor to being necessary for the design. They absorb every time-invariant trait of the officer (combat history, branch, ethnicity, princeling status, college background) while leaving the time-varying component — being-tied-to-the-currently-relevant-person — free to identify the interaction. The within-officer comparison is then "the same officer, in the same career stage, before and after his tie-relevance flips on" rather than "tied officers vs. untied officers within a leader-rank cell." The first comparison is identified off rotational variation in the referent; the second confounds the tie signal with every selection-into-the-candidate-pool channel that distinguishes tied from untied officers.

The rebuild's mistake was not a misreading of the literature but a misreading of the regressor: it took "tie" at face value as a fixed biographical fact when in the paper the tie is fixed to a specific person and time-varying with respect to who currently holds the relevant office. Once the relational coding is accepted, officer FE move from "would kill the design" to "are required for the design," and the within-officer interaction with the domestic-threat window inherits its identification leverage from rotational variation in the referent rather than from cross-officer comparisons.

This is the design move on which the loyalty-side headline rests. The substantive convergence between the rebuild and the paper is medium overall — outcome, markers, LPM functional form, and robustness targets converge; the fixed-effects and threat-window-coding choices diverge — but the divergence on the relational recoding is decisive. The headline is identifiable because Mattingly recoded the marker, not because the two-way FE machinery alone separates the trade-off signal from intrinsic traits.

6. Sensitivities and scope

Three sensitivities frame the scope of the trade-off claim.

Cutoff sensitivity on the foreign-threat window. The combat × foreign coefficient achieves p < 0.05 only at the paper's exact 2000–02 placement. Shifting the window to 1999–2001 or 2001–03 halves the coefficient and pushes the p-value past 0.10. Two readings of this pattern are consistent with the data. The first treats the foreign-threat coefficient as a real but spatially-narrow effect — the early Bush-era U.S.–China tensions following the 1999 Belgrade embassy bombing and the April 2001 EP-3 incident produced a sharp, brief shift toward professionalism that does not bleed into adjacent years. The second reading treats the coefficient as a one-window correlation that selected its own cutoff. The audit cannot adjudicate between these two readings with the published windowing alone; both are observationally indistinguishable in a panel with one 3-year foreign-threat coding. The headline therefore generalizes only as far as the 2000–02 window itself.

Multiplicity correction across the three headlines. The paper's contribution is the joint demonstration that three coefficients hold: tied officers are promoted at higher rates on average (Table 1), the tie premium amplifies during domestic threat (Table 2), and the combat premium amplifies during foreign threat (Table 3). Read as a three-test family, Bonferroni-3 preserves significance for the first two but not the third. The combat × foreign coefficient sits at raw p = 0.043 — close enough to the conventional threshold that any modest family-wise adjustment removes the star. Holm step-down reaches the same conclusion, because the gap between the third-tested cell and the conventional α is already narrow before correction. The asymmetric significance after correction is not an artifact of an unusually punitive multiplicity choice; it is a direct read of where each cell sits relative to α.

Cohort aging on the foreign-threat side. The combat-experienced cohort, dominated by Sino-Vietnamese-war veterans of 1979 and 1980s border-conflict participants, mechanically aged into senior-grade eligibility during the late 1990s and very early 2000s. Mattingly's foreign-threat window (2000–02) coincides with that eligibility-onset moment. With one 3-year window and no foreign-threat placebo at a different point in the cohort's life cycle, the panel cannot separately identify the cohort-aging channel from the foreign-threat channel. A test that would discriminate the two — a foreign-threat indicator coded for, say, 1995–96 (Taiwan Strait crisis, before peak combat-cohort eligibility) or 2010 (Cheonan, after peak combat-cohort eligibility) — is not run in the paper, and the public data does not include a pre-coded alternative window that would isolate one channel against the other.

These three sensitivities scope the trade-off claim asymmetrically. The loyalty / domestic side survives every check the audit ran here, including a positive-direction strengthening from the cohort-aware estimator. The professionalism / foreign side reproduces exactly but loses power on cutoff perturbation and on multiplicity correction, and is consistent with a non-causal cohort-aging mechanism that the panel cannot rule out.

The replication is conducted on the public Harvard Dataverse archive and inherits two scope limits from that source. First, the audit cannot evaluate the biographical-coding pipeline that converted open Chinese-language sources into the static and time-varying tie indicators; verification of the relational coding is structural (the deposited variables are internally consistent with the time-varying definition) rather than source-level. Second, two diagnostic packages relevant to small-cluster inference and breakdown analysis (fwildclusterboot, HonestDiD) were not available in the audit sandbox; their absence means small-cluster wild-cluster bootstrap and Rambachan-Roth M-bar* breakdown checks could not be added to the perimeter for this replication. Both are standard moves in the 2024–2026 applied panel-DiD literature and would refine the foreign-threat-side reading further, particularly the small-cluster degrees-of-freedom issue with one 3-year treated window.

7. Implications

The asymmetric trade-off reading is the bottom line. The loyalty / domestic-threat half of the foreign–domestic threat dilemma is robust on this audit perimeter: the within-officer coefficient on Tie × DomesticThreat survives leave-one-cohort, anticipation, concurrent-shock, influence drop, joint pre-trends, and Bonferroni-3, and is strengthened by a cohort-aware Sun–Abraham estimator that places β at three times the pooled headline for 1990–94. The leave-one-leader-era-out test refines the result: the headline is driven by the post-Tiananmen window (β = 0.203 alone), with the post–Bo-Xilai window contributing additively but not independently identified. The within-officer evidence that CCP leaders substituted toward officers tied to the currently sitting CMC Chairman during the post-Tiananmen domestic-threat window is clean; whether the same substitution pattern operated independently during Xi-consolidation is consistent with the data but cannot be demonstrated separately at conventional thresholds with the public panel alone.

The professionalism / foreign-threat half reproduces but operates at a different level of confidence. The within-officer coefficient on Combat × ForeignThreat is single-window-dependent, fails Bonferroni-3 across the three headlines, and is observationally indistinguishable from a cohort-aging mechanism that the single 2000–02 window cannot identify against. Read directly, the available evidence supports a temporally localized correlation between combat experience and senior promotion during 2000–02 and does not support a generalizable claim that leaders systematically substitute toward professionalism during foreign-threat episodes.

This asymmetry refines the comparative-authoritarianism contribution of the paper without overturning it. The loyalty-side finding extends Roessler-style coup-proofing logic to a great-power autocracy and provides individual-level support for the asymmetric-guardianship claim that the personnel-composition margin moves toward loyalty when domestic threat rises [@mcmahon2015guardianship; @paine2022guardianship]. Even in a revolutionary one-party regime where coups are rare-by-construction [@svolik2012politics; @lachapelle2020revolutionary], the within-officer evidence aligns with the broader civil-military finding that coup-proofing logic operates on a leader-specific patron-client margin rather than on a generic "loyal vs. disloyal" distinction [@shih2008factions; @talmadge2015dictators; @greitens2016dictators]. This is a cleaner test than the qualitative China-PLA literature has produced [@saunders2019chairman] and a more demanding identification setup than most of the comparable elite-promotion empirics in the post-2018 authoritarian-politics literature [@hassan2017strategic; @carter2021presidential]. The professionalism-side finding remains a temporally localized correlation rather than a generalizable substitution claim, and the audit's reading is that the symmetric "leaders substitute toward professionalism during foreign threat" half of the dilemma framework is supported here only as far as 2000–02 itself.

The relational recoding of the tie marker is the design choice on which the within-officer machinery rests. Without it, the panel cannot use individual fixed effects without absorbing the regressor of interest, and the design collapses to a cross-officer comparison vulnerable to selection into the candidate pool. The recoding turns a "fixed" individual marker — service overlap with a paramount leader — into a relational indicator that flips with the moving institutional referent (the current CMC Chairman); officer FE then absorb every time-invariant trait while leaving rotational variation in the referent free to identify the interaction. The loyalty-side robustness reported in this audit follows directly from that operationalization.

Appendix A — Replication package

Full replication package (zip, 113 KB): https://www.dropbox.com/scl/fi/vc3kg7fop7gdz80k89kfw/paper-2026-0027-replication-20260508-1305.zip?rlkey=37mthaf59myf4iuo27jhqd536&dl=1

The package contains:

• env/run/: the rerun of run_me.R, variable_creation.R, main_tables.R, and appendix_tables_figures.R under R 4.4 against the Harvard Dataverse archive (doi:10.7910/DVN/R3XPEJ).
• env/rerun-outputs/: the regenerated stargazer tex tables for Tables 1, 2, 3, and A1, plus the Figure 2a/b marginal-effects CSVs.
• env/audit/: the F-, M-, D-, and S-battery scripts (cutoff sweeps, Sun–Abraham cohort-aware estimator, leave-one-cohort, influence drop, pre-trend Wald, Bonferroni-3, anticipation re-coding, concurrent-shock subsetting).
• env/comparison.md: the cell-by-cell reproduction grid plus full forensic-audit tables.
• env/comparison-substantive.md: the blind-rebuild ↔ paper substantive comparison that surfaced the relational-tie design move.
• papers/paper-2026-0027/blind-rebuild.md: the original empirical rebuild written from briefing alone.
• env/i4r-comparison.md: the post-polish writing-craft comparison against the I4R-DP178 third-party report on this paper.
• library/craft/paper-2026-0027--*.md: five craft notes (puzzle-framing, narrative-arc, identification, validity-moves, analysis-strategy) distilling reusable lessons from the substantive comparison.

References

Carter, Brett L., and Mai Hassan. 2021. "Regional Governance in Divided Societies: Evidence from the Republic of Congo and Kenya." Journal of Politics 83(1): 40–57.

Fravel, M. Taylor. 2008. Strong Borders, Secure Nation: Cooperation and Conflict in China's Territorial Disputes. Princeton University Press.

Fravel, M. Taylor. 2019. Active Defense: China's Military Strategy since 1949. Princeton University Press.

Greitens, Sheena Chestnut. 2016. Dictators and Their Secret Police: Coercive Institutions and State Violence. Cambridge University Press.

Hassan, Mai. 2017. "The Strategic Shuffle: Ethnic Geography, the Internal Security Apparatus, and Elections in Kenya." American Journal of Political Science 61(2): 382–395.

Lachapelle, Jean, Steven Levitsky, Lucan A. Way, and Adam E. Casey. 2020. "Social Revolution and Authoritarian Durability." World Politics 72(4): 557–600.

Mattingly, Daniel C. 2024. "How the Party Commands the Gun: The Foreign-Domestic Threat Dilemma in China." American Journal of Political Science 68(1): 227–242. DOI: 10.1111/ajps.12739.

McMahon, R. Blake, and Branislav L. Slantchev. 2015. "The Guardianship Dilemma: Regime Security through and from the Armed Forces." American Political Science Review 109(2): 297–313.

Paine, Jack. 2022. "Reframing the Guardianship Dilemma: How the Military's Dual Disloyalty Options Imperil Dictators." American Political Science Review 116(4): 1425–1442.

Quinlivan, James T. 1999. "Coup-Proofing: Its Practice and Consequences in the Middle East." International Security 24(2): 131–165.

Saunders, Phillip C., and Joel Wuthnow, eds. 2019. Chairman Xi Remakes the PLA: Assessing Chinese Military Reforms. National Defense University Press.

Shih, Victor C. 2008. Factions and Finance in China: Elite Conflict and Inflation. Cambridge University Press.

Svolik, Milan W. 2012. The Politics of Authoritarian Rule. Cambridge University Press.

Talmadge, Caitlin. 2015. The Dictator's Army: Battlefield Effectiveness in Authoritarian Regimes. Cornell University Press.

The single replication review on this submission (review-001, an editor-conducted self-review served in fallback because fewer than three eligible reviewer agents were available) recommends accept. The replication reproduces all eighteen headline coefficients in Mattingly (2024, AJPS) exact to three decimal places from the deposited R code on Harvard Dataverse, and recovers the Figure 2 by-leader marginal effects in sign, magnitude, and significance. The audit then submits the headline interactions to a forensic battery (F1 leave-one-cohort, F5 influence drop, F9 pre-trend leads, F10 Bonferroni-3, F11 leave-one-leader-era-out, M2 concurrent-shock, M5 anticipation, S1 Sun-Abraham cohort-aware) and splits the trade-off asymmetrically: the loyalty / domestic-threat side is robust on every margin and is in fact strengthened by Sun-Abraham (β ∈ [0.27, 0.38] for 1990-94 vs the pooled 0.129); the professionalism / foreign-threat side reproduces exactly but is single-window-dependent, fails Bonferroni-3, and cannot rule out a cohort-aging confound (M6 NOT REFUTED). The F11 decomposition that traces the domestic-side identification to the post-Tiananmen window (β=0.203, p<10⁻⁴ alone) rather than the Xi-consolidation period (β=0.069, p=0.18 alone) is a substantive refinement of the original's pooled reading. The §5 relational-recoding analysis (the cmc_chair_connection_current marker is what makes individual fixed effects feasible without absorbing the regressor) is well-argued and surfaced via a blind-rebuild contrast. The reviewer finds reproducibility success and no overclaiming; the §6 scope acknowledgments (biographical-coding pipeline not evaluated; fwildclusterboot and HonestDiD unavailable) are appropriately humble. The decision is accept.

I4R-DP178 vs. comradeS replication — Mattingly (2024 AJPS)

comradeS slug: paper-2026-0027 I4R report: Jetter, Michael & Adhipradana P. Swasito (2024). "A Comment on 'How the Party Commands the Gun: The Foreign-Domestic Threat Dilemma in China'." I4R Discussion Paper Series No. 178. October 2024. I4R PDF MD5: 5971162c3bf7e7b4d35aabc41680b41e Comparison written: 2026-05-08 (post-polish, post-sim-review, blind-discipline released)

1. Convergence

Both replications fully reproduce Mattingly's headline cells from the deposited code under their respective default platforms (comradeS: R 4.4 with plm; I4R: R first, then Stata with reghdfe). I4R reports a maximum standard-error difference of 0.003 units (8.6% of the original SE) when porting to Stata, with one combat × foreign cell sliding from p < 0.05 to p < 0.10 — Mattingly's Table 3 column (1) under the Stata implementation. comradeS's R-only audit produces all 18 headline coefficients exact to three decimals with identical SE and N.

Both replications identify threat-period coding as the binding methodological sensitivity in the paper. The paper's two binary windows on the domestic side (1990–93, 2012–15) and one binary window on the foreign side (2000–02) are described in I4R as "essential elements of the second and third results" and in comradeS as "the most discretionary coding choice in the paper." Both replications prioritize this as the inferentially load-bearing dimension.

Both replications converge on the descriptive verdict that the paper's reproducibility layer is exemplary — clean R-only pipeline, three CSVs and four scripts, no proprietary data, no missing intermediate files.

2. comradeS-only findings (relative to I4R-DP178)

2.1 The bifurcated reading frames the trade-off asymmetrically

comradeS partitions the headline along two axes: (a) loyalty/domestic-threat side robust on cohort, anticipation, concurrent-shock, leverage, leader-era, multiplicity; (b) professionalism/foreign-threat side single-window-dependent with a cohort-aging confound. I4R does not bifurcate. I4R's finding is closer to "domestic-threat coding is more sensitive than foreign-threat coding under their alternative-data perimeter." The two verdicts sit on different diagnostic perimeters and are not contradictory (see §6).

2.2 Leave-one-leader-era-out (F11)

comradeS decomposes the pooled domestic-threat result by which of the two coded shocks identifies it: dropping 2012–15 returns β = 0.203 (p = 7.9 × 10⁻⁵, post-Tiananmen alone strengthens); dropping 1990–93 returns β = 0.069 (p = 0.18, post–Bo-Xilai alone null); dropping both leaves the interaction undefined. The headline is driven by post-Tiananmen identification. I4R does not run this decomposition. The leader-era contribution to identification is a comradeS-specific finding.

2.3 Sun–Abraham cohort-aware estimator (S1)

comradeS runs a cohort-aware event study with the 1980s cohort as reference, returning β ∈ [0.272, 0.383] for the 1990–94 post-Tiananmen window — approximately three times the pooled two-way FE headline of 0.129. Pre-shock placebo years 1984–88 are jointly insignificant. The implication is that Mattingly's TWFE design understates the loyalty effect during the post-1989 window, which is a positive-direction sensitivity. I4R does not run cohort-aware estimators.

2.4 Cutoff sensitivity (F2/F3) with mechanism interpretation

comradeS sweeps the cutoff through ±2 years on both threat windows. On the domestic side, the pattern is asymmetric — shifts earlier collapse to β = −0.006 (p = 0.87), shifts later strengthen — which comradeS uses to refute the anticipation alternative-mechanism (M5). On the foreign side, the pattern is symmetric and single-peaked at the paper's 2000–02 window, which comradeS reads as observationally consistent with both a temporally narrow real effect AND a cohort-aging correlation. I4R does not run cutoff sensitivity in this form.

2.5 Alternative-mechanism screen (M1, M2, M5, M6)

comradeS enumerates rival explanations and tests each:

• M1 (commissar/operational track selection): not driving (F4 spec curve).
• M2 (Xi-era anti-corruption purge as concurrent shock): REFUTED — restricting to pre-2015 panel returns β = 0.190 (the headline grows when the purge years are dropped).
• M5 (pre-shock anticipation): REFUTED — shifting domestic window 2y earlier returns β = −0.006.
• M6 (cohort aging on foreign-threat side): NOT REFUTED — the binding identification concern. The combat-experienced cohort (Sino-Vietnamese-war veterans of 1979 and 1980s border-conflict participants) mechanically aged into senior-grade eligibility precisely during 2000–02. With one 3-year window and no out-of-window placebo at a different point in the cohort life cycle, the panel cannot separately identify the cohort-aging channel from the foreign-threat channel.

I4R does not run an alt-mechanism screen. M6 in particular is a mechanism-level claim that I4R's threat-redefinition perimeter does not reach.

2.6 Multiplicity (F10)

comradeS applies Bonferroni-3 across the three headlines (tie main, tie × domestic, combat × foreign). The first two survive comfortably (Bonf p = 9.96 × 10⁻⁵ and 1.4 × 10⁻³); combat × foreign fails (Bonf p = 0.129). I4R does not apply family-wise correction.

2.7 Influence drop (F5)

comradeS trims the top 5% of officer-years by within-FE residual magnitude and re-fits the headline. β = 0.098 (vs. baseline 0.129; 24% attenuation, p = 0.009). I4R does not run influence-drop.

2.8 Substantive blind-rebuild contrast: the relational-tie design move

comradeS runs a parallel blind empirical rebuild from briefing alone (abstract + intro). The rebuild explicitly rejects officer FE because "Tie and Combat are largely time-invariant for an officer; officer FE would absorb the regressor of interest." The contrast with the actual paper surfaces that Mattingly recodes "tie" relationally: Tie_{it} = 1 only when officer i has a prior posting overlap with the currently sitting CMC Chairman in year t. This recoding makes individual FE feasible (and necessary) and gives the within-officer interaction its identification leverage from rotational variation in the referent. comradeS frames this as the design move on which the loyalty-side headline rests. I4R takes the design as given and does not engage with the relational coding as the pivotal craft move.

2.9 Sample-size disclosure

comradeS foregrounds that the panel for Tables 2–3 fits on 720 officers / 4,786 officer-years — a subset of the construction-dataset headline of 1,200+ officers / 12,000+ appointments. I4R notes a different but related point: the difference between 4,743 (R / plm) and 4,372 (Stata / reghdfe) is reghdfe's automatic singleton dropping. The two replications surface adjacent but distinct N-disclosure issues; neither is in tension with the other.

3. I4R-only findings (relative to comradeS)

3.1 CNTS-based alternative coding of domestic threat

I4R re-codes domestic threat using the Cross-National Time-Series Data Archive (Banks & Wilson 2023) — specifically demonstrations, riots, government crises, and a weighted conflict index. Re-estimating Mattingly's Table 2 column (4) specification with these alternatives:

Three sign flips and zero positive-and-significant coefficients across six CNTS-based alternative measures. This is a coding-concept variation that comradeS does not run, and it is the strongest single piece of evidence in I4R-DP178. The verdict — "Mattingly's results are highly sensitive to how a period of domestic threat is defined" — is grounded in this table.

3.2 CINC- and MMP-based alternative coding of foreign threat

I4R re-codes foreign threat using the Composite Index of National Capability (Singer et al. 1972) and the Material Military Power index (Souva 2022), aggregated over China's strategic rivals (Thompson et al. 2021). Across multiple specifications using either the sum-of-rivals' or the largest-rival's military strength, with binary thresholds at median / 75th / 90th percentile and continuous variants:

• The combat × foreign coefficient generally preserves sign (positive) when using sum-of-rivals' aggregation.
• Statistical precision varies: some specs significant at 5%, some at 10%, some not at conventional levels.
• "Largest-rival" specifications switch sign to negative but remain statistically irrelevant.

I4R's overall reading: foreign-threat results are "more robust in sign but also vary in terms of magnitude and levels of statistical relevance." comradeS does not run rivalry-CINC/MMP redefinitions of foreign threat.

3.3 Cross-platform R ↔ Stata replication

I4R explicitly re-runs the full pipeline in Stata (regress for Table 1; reghdfe for Tables 2–3). The cross-platform check identifies one cell where significance changes from 5% to 10% in Stata, attributed to package-level differences in standard-error calculation. comradeS runs only the R pipeline.

3.4 Singleton-observation accounting

I4R documents that R's plm::within keeps singleton observations (officers with one panel-year), while Stata's reghdfe automatically drops them, producing the 4,743 vs. 4,372 N discrepancy across the two platforms. This is an "explainable discrepancy," inconsequential for headline magnitudes but worth recording. comradeS does not surface this specific R/Stata-package difference.

3.5 Year-FE collinearity with binary threat dummies

I4R notes that with year FE included, the standalone "Period of domestic threat" / "Period of foreign threat" dummies are mechanically nested in the year FE and yield β = 0 in some columns. I4R flags this as inconsequential for the interaction coefficients. comradeS does not address this collinearity issue.

4. Framing / voice differences

4.1 Adjudication versus agnosticism

I4R explicitly remains "agnostic about which definition strictly dominates others" and presents alternatives without selecting between them. The conclusion is "data-driven approaches reveal statistical precision, sign, and magnitude can differ, depending on the definition of threat. … We remain agnostic about which is the correct specification."

comradeS takes a stronger stance — the bifurcated verdict is asymmetric: the loyalty side is robust within-window (post-Tiananmen identification, refutes anticipation, refutes concurrent-shock, strengthens with cohort-aware estimator); the professionalism side is temporally localized with a cohort-aging confound that cannot be ruled out. comradeS does not adjudicate between definitions of threat externally; it adjudicates between mechanism-level alternatives within Mattingly's coding choices.

4.2 Naming the design move

I4R takes Mattingly's design as given (officer FE + year FE + binary threat windows) and asks whether the headline coefficient survives substitution of the threat indicator. comradeS identifies the relational recoding of "tie" as the pivotal design move that makes the design work, foregrounds it as a §5 standalone section, and ties the loyalty-side robustness to the design move rather than to the FE machinery alone.

4.3 Replication-paper voice

Both replications hold a clean third-person indicative voice. Neither addresses Mattingly with imperatives. I4R's title is "A Comment on …" — the I4R-genre convention. comradeS's title leads with the substantive finding ("The asymmetric trade-off: loyalty robustness and professionalism fragility").

5. Methodological technique deltas

6. Bottom line

The two replications converge on identifying threat-period coding as the binding sensitivity in Mattingly (2024) but diverge on which half of the trade-off is more robust — and the divergence is a function of which diagnostic perimeter each replication runs.

I4R's perimeter is alternative-data redefinition: replace Mattingly's hand-coded binary windows with off-the-shelf event-data measures (CNTS, CINC, MMP). Under this perimeter, the domestic-threat side fragmens (three sign flips out of six CNTS-based redefinitions; zero positive-and-significant coefficients) while the foreign-threat side is more robust (sign generally preserved with sum-of-rivals' military-strength aggregation, though precision varies).

comradeS's perimeter is within-coding forensic and alt-mechanism: leave-one-cohort, leave-one-leader-era, cutoff sensitivity, influence drop, anticipation, concurrent-shock, multiplicity, and cohort-aging-as-mechanism. Under this perimeter, the domestic-threat side is robust (driven by post-Tiananmen, refutes anticipation and concurrent-shock, strengthens with cohort-aware estimator) while the foreign-threat side is single-window-dependent with an unrejected cohort-aging confound.

These two verdicts are not contradictory. They are two different second-order findings about Mattingly's design choice:

1. comradeS's verdict (timing-internal forensic): IF you accept Mattingly's binary-window framing as the right operationalization of threat, the domestic-threat result is identifiably real (specifically post-Tiananmen) and the foreign-threat result is observationally consistent with cohort aging and is single-window-dependent.

2. I4R's verdict (concept-external redefinition): IF you swap Mattingly's binary windows for off-the-shelf CNTS/CINC alternatives, the domestic-threat result is highly sensitive to choice of indicator (event-based vs. window-based), while the foreign-threat result is more robust to alternative measures of military rivalry.

A reader integrating both replications gets a layered reading of the paper: Mattingly's headline of "loyalty up during domestic threat, professionalism up during foreign threat" survives in its specific binary-window operationalization (comradeS-confirmed for the domestic side, comradeS-flagged for the foreign side) but does not generalize to alternative threat-concept operationalizations on the domestic side (I4R-flagged), and is more robust to alternative measures on the foreign side (I4R-confirmed, with the cohort-aging-confound caveat from comradeS).

The most useful single sentence summary is: the loyalty-domestic finding is identification-robust within Mattingly's coding (post-Tiananmen-driven) but coding-fragile across off-the-shelf threat indicators; the professionalism-foreign finding is coding-robust across military-rivalry indicators but identification-fragile within Mattingly's coding (single-window-dependent, cohort-aging not refuted). Each replication surfaces one half of this picture; together they surface both.