Replication package for: Trust in Whom? Only trust in the delivering entity predicts collective action

This package reproduces every numerical claim in the manuscript and its online supplement from the raw survey data through a single deterministic R script. It verifies 278 atomic quantities, every number the paper prints, under strict verification: each computed value is rounded to the precision at which the manuscript prints it, and the verdict requires exact equality. All 278 reproduce deterministically; the certified result fingerprint is recorded in the package and every later run self-checks against it. The package is self-contained and self-certifying. The four raw data files sit in the /data folder; the script 00_REPLICATE_ALL.R self-locates them, computes each claim, and writes an itemized verdict pairing every manuscript target with its computed value. On first run the script records a SHA-256 fingerprint of its results, rounded to reported precision; every later run recomputes and halts if any reported value has changed, naming the quantity that moved. The one stochastic step is seeded, so output is identical across runs. SHA-256 hashes of the four raw files are recorded in MANIFEST_SHA256.txt. Determinism is demonstrated rather than asserted. This build was run from the posted archive on two machines with different operating systems, R versions and locales, Windows 11 with R 4.6.1 ucrt and Linux with R 4.3.3, and produced verdict files identical byte for byte and the same result fingerprint, 278 LOCK and 0 CHECK and 0 FLAG on each. Every computed value agreed to the recorded precision, including the village-clustered standard errors added in this version. Contents: the four raw data files (the Dhading enrollment ledger is anonymized, household names removed, ward-level address reduced to the VDC, all analytic columns retained); the deterministic replication script; a one-time environment certifier and its renv.lock; the figure script R/02_FIGURES.R, which reproduces both manuscript figures from the raw data (PNGs included) and writes FIGURE_TEXT.txt, an inventory of every numeric string rendered in the figures; REVIEWER_START_HERE.md, a five-minute path stating what to run and what a successful run prints; a codebook documenting every analysis variable with its source file, storage type, the label recorded in the file, its value codes, and the definitions of the derived variables; a master replication-and-provenance record documenting the source of every claim with inline code; a data-provenance registry; a data provenance certificate recording owner, funding lineage, file checksums, the corrections made to the manuscript rather than to the data, a recorded exclusion scan, and the privacy rule applied to the withheld roster; and licence and citation files. Version 3.0.1 (this version) extends certification from the five tables and the main narrative to every number the paper prints, 252 to 278. The added claims are the p-values accompanying already-certified estimates, including those printed in scientific notation; the Table 2 composite channel index; the fixed-effects subsample estimated without fixed effects; the channel-by-generic interaction and the correlation between the two trust measures; the caste and gender interactions with their subgroup odds ratios and sample size; the site-level generic-trust estimate; the gatekeeper models and the count of households in villages with no self-help group members; and village-clustered standard errors computed with a CR1 sandwich implemented in base R inside the script. Three manuscript items were corrected to the data in the same pass, under the standing rule that where manuscript and data disagree the manuscript is corrected: the caste interaction p-value, which computes as 0.0395 and prints as 0.040 rather than 0.039; the village-penetration odds ratio controlling for channel trust, which computes as 0.135 and prints as 0.13 rather than 0.14, the bivariate estimate of 0.137 being the one that prints as 0.14 and the two having been conflated; and the direction of the caste interpretation, since the source file's value labels place Scheduled Caste and Scheduled Tribe households in the group with the odds ratio of 1.12, not 1.58. New documentation in this version: the reviewer quick-start file, the codebook, and the provenance certificate with its recorded exclusion scan, which confirms by filename and content scan that no material from a superseded and quarantined dataset family used in earlier, unrelated work is present in this package. After certification, five robustness subsections were moved from the manuscript to its online supplement to meet the journal's length guidance; no quantity changed and none was removed, the claims reporting them remain certified here, and the supplement names the claim identifiers. No data file and no analysis logic from v2.0.1 changed, and the 252 claims certified there are certified here with identical computed values. Version 2.0.1 corrected the figure annotations to the manuscript's before-rounding convention: the third-panel enrollment lift in Figure 1 reads +23.3 pp (computed before rounding, matching golden-verdict row V11 and the manuscript), where v2.0.0's figure showed +23.4 pp (a difference of rounded endpoints); it added an explicit x-axis title to Figure 2 (village SHG penetration) and aligned the figure vocabulary with the manuscript's. The figure script began emitting FIGURE_TEXT.txt so that every rendered numeric string is inspectable against the golden verdict. No data file, no analysis logic, and no certified value changed: the 252/252 certification and the golden fingerprint were identical to v2.0.0. Version 2.0.0 corrected and extended version 1: (1) coverage extended from headline claims to all 252 checkable numbers, including every confidence interval and the FDR-adjusted p-values of the falsification table; (2) the four raw data files included, so replication runs against the deposited data with no external dependency; (3) the multi-script pipeline replaced by one deterministic script with a golden-fingerprint self-check that halts on any drift; (4) two data-mandated manuscript corrections, both traced to the original analysis scripts, the out-group trust odds ratio reported under the normalized coding used throughout (1.08, not the raw-coded 1.03), and the Dhading peer-borrowing odds ratio 1.72, the value the analysis script computes; (5) the India sample geography stated correctly (seven sites in four states; 244 villages, 25 districts); (6) the Dhading enrollment ledger deposited anonymized. Correction to v1: the v1 file raw_data_files.md attributes the ECCP India dataset to NWO/WOTRO + BMZ; the correct funding is the EU-India Economic Cross Cultural Programme (EC grant ASIE/2004/095-995, with GTZ), as recorded in DATA_PROVENANCE_REGISTRY.md (rev. 2). Licence: MIT (code) + CC BY-NC 4.0 (data and documentation); see LICENSE.md.