Audit Ledger Sri Lanka: data dictionary
Version 0.1, demo, October 2026. Random samples of 20 findings and 20 published links, and the 23 recurring problems in the matching test, were checked by hand against the original reports (hand_check_results.json); other checks were done by AI. Released under CC0.

findings.csv (one row per audit finding)
institution        Name of the audited institution.
type               "corporation" (state corporation) or "department" (government department).
year               Audit year of the report the finding comes from.
ref                Paragraph number of the finding in that report.
section            Report section heading the paragraph sits under.
summary            One-line summary written by AI. Not the auditors' exact words; read them in the report at the paragraph in ref.
amount_rs          Main rupee amount stated in the finding, if any. Can overlap where a total and its parts are both given.
since_year         Year the problem dates from, if the report says.
category           Type of problem, a first draft for review by a chartered accountant.
recurring_problem_ids  IDs of the recurring problems this finding belongs to, separated by ";". Empty if not linked to another audit.
source_pdf         Link to the report PDF on the Audit Office website.
(The auditors' original wording is not stored in this version. Each summary points to its report and paragraph instead.)

dataset.json (everything shown on the site)
version            Version label of this release.
years              Audit years covered.
totals             Headline counts used on the site (institutions, reports, findings, recurring problems and the 2017 follow-up figures).
quality            Accuracy figures: the hand check (hand), and the 50-finding reading check and link checks done by AI.
institutions       One entry per institution: name, type, reports (with findings as in findings.csv), per-year counts and measures.
themes             One entry per recurring problem: id, institution, title, years raised, and the IDs of its findings.

accuracy_check.json  The 50-finding reading check, done by AI: each sampled finding and whether each field was right.
hand_check_results.json  The hand check: 20 findings, 20 published links and 23 test-set recurring problems checked against the original reports, with each result.
link_review.json     Every link the script proposed between two findings, with its AI verdict ("repeat" or not) and how it was decided.
match_eval.json      Results of testing the matching script, including the blind test on two departments.