Where Nepal's corruption accountability actually leaks

Corpus. Of ~12,600 Special Court records, most are procedural petitions. We isolate CIAA prosecutions as the Special Court's -CR- criminal register — 2,880 cases filed FY2069/70–2082/83 (the register is the definition; no plaintiff filter) — of which 2,726 are substantive corruption charges after removing money-laundering (a separate statute) and unclassified matters.

The funnel. Of 28,554 newly registered complaints in the year, only 947 (3.3%) went to a full investigation; most of the rest were screened out at intake — shelved or referred — much of it legitimate (outside the CIAA's jurisdiction, no supporting evidence, or duplicates). Of the complaints it fully investigated, the CIAA filed charges in 137 — about 1 in 7 (the CIAA reports this as 13% of its investigation decisions). So “0.5% of all complaints reach court” and “~1 in 7 of the complaints it investigates is prosecuted” are both true and measure different stages; the steep drop is concentrated at screening, not the courtroom.

Outcomes. Verdicts are coded per hearing as convicted / acquitted / partial; each case is taken at its terminal deciding hearing. The conviction rate is over the 2,728 register cases carrying an unambiguous ठहर / आंशिक / सफाई disposition — 18 more than the 2,710 whose case status reads फैसला (which the filed-vs-decided trend counts), because a few carry a deciding-hearing disposition without that status.

Dates. Verdict dates are parsed from the case status text; filings from the registration date. Bikram Sambat dates throughout; by-year charts bin by fiscal year (Shrawan–Ashadh).

Over time. Yearly rates are grouped by verdict fiscal year; the sharp rise in acquittals from FY2078/79 is a genuine surge in the record, not a coding artifact. Time-to-verdict is measured by filing cohort: cohorts through FY2079/80 are essentially fully decided, but recent cohorts are still open, so their apparent speed reflects only the cases already resolved (survivorship) and is drawn as provisional.

Per-justice. Attribution is bench-grain: every member of a panel is credited with the panel's outcome, so this describes the benches a justice sat on, not that justice's individual effect. It is descriptive, and small differences are noise.

Discrepancy with CIAA figures. The CIAA reports a ~53% “success” rate; that counts full + partial convictions together, whereas our full-conviction rate (45%) is cumulative and separates the two. Both bin by fiscal year, but CIAA's “cases filed” per year can still differ from ours because of case-versus-defendant counting and record timing. Never compare the two without aligning definition and period.

Identity. Only ~7% of distinct defendants (607 of 8,114) are resolved to a canonical, cross-referenced identity, so office-level and repeat-offender cuts are deferred as low-confidence.

Reading the source PDFs. The CIAA reports are Nepali-language PDFs set in legacy Devanagari fonts that ordinary tools garble. We convert them to clean, checkable Markdown with likhit — Jawafdehi's open-source universal Nepali document-to-markdown converter — then verify every figure by eye against the original page.

Limits. These are the records in the archive as of the snapshot date (2083 BS); figures update as new records are mirrored. Appellate outcomes are largely absent, and amount-recovered is untracked anywhere.