Live numbers
Impact
What Sach has done, counted by the server itself. Generated 2026-11-12 20:27:07+00:00 UTC.
Read this first
Counting since 2026-11-12T17:46:07+00:00. This server's counters are persistent.
Numbers are computed from the ledger and counters on every page load; nothing here is typed in. Counters live in a shared remote store and survive server restarts. If that store cannot be reached, counters read as zero.
Checks
checks counted
ran the full pipeline
answered from the cache (50.0% of all checks)
had nothing checkable
Verdicts
- TRUE: 1, 12.5%
- FALSE: 4, 50.0%
- MISLEADING: 0, 0.0%
- UNVERIFIED: 3, 37.5%
Shares are out of 8 (the sum of the four verdict counters).
Languages of the forwards
- English: 7, 87.5%
- Hindi: 0, 0.0%
- Bengali: 1, 12.5%
- Marathi: 0, 0.0%
- Telugu: 0, 0.0%
- Tamil: 0, 0.0%
- Gujarati: 0, 0.0%
- Urdu: 0, 0.0%
- Kannada: 0, 0.0%
- Odia: 0, 0.0%
- Malayalam: 0, 0.0%
- Punjabi: 0, 0.0%
- Assamese: 0, 0.0%
- Konkani: 0, 0.0%
- Nepali: 0, 0.0%
- Manipuri (Meitei): 0, 0.0%
- Tinglish: 0, 0.0%
Shares are out of 8 (the sum of the language counters).
Evaluation
- Claims in the evaluation set: 31
- Judged exactly right: 26
- Percent right: 83.9%
- Overclaimed FALSE: 0
- Overclaimed TRUE: not available
- Said unverified instead of guessing: 5
- Run on commit: not available
- Run date: not available
Not in this evaluation file (older format): overclaimed_true, commit, date, by_lang. Shown as not available.
Read from data/eval_summary.json, which scripts/run_eval.py writes after a full run of the evaluation set. Percent correct = correct / n.
Public ledger
14 claim rows in the public ledger: 2 TRUE, 8 FALSE, 0 MISLEADING, 4 UNVERIFIED.
Latest FALSE verdicts
- Government is giving Rs 1000 to everyone under Corona Sahayata Yojana.2026-11-12 · PIB (Press Information Bureau)
- A Health Ministry office memorandum on state holidays has been issued.2026-11-12 · PIB (Press Information Bureau)
- Union Minister Sarbananda Sonowal's office has issued a letter on this matter.2026-11-12 · www.pib.gov.in
- Government is giving Rs 1000 to everyone under Corona Sahayata Yojana, click the link to register.2026-11-12 · PIB (Press Information Bureau)
- WHO doctors are advising against wearing masks and social distancing.2026-11-12 · www.who.int
Source rules in sach/fetch.py: 3 official suffixes, 3 official hosts, 1 official zones, 16 more official domains; 6 fact-checker sites and 6 more with path rules (secondary, labelled). Open the full ledger.
How each number is computed
- checks total: Counter checks_total: incremented once for every completed check, including cache hits.
- checks served fresh: checks_total minus cache_hits: checks that ran the full pipeline.
- cache hits: Counter cache_hits: incremented when an identical forward was answered from the verdict cache.
- no claims: Counter no_claims: forwards with nothing checkable (greetings, emoji, bare links). They are not in the verdict breakdown.
- verdicts: Counters verdict:TRUE, verdict:FALSE, verdict:MISLEADING, verdict:UNVERIFIED (overall verdict of each check). Each share = that count / the sum of the four, and the denominator is shown.
- languages: Counters lang:<code> for every supported language code, e.g. lang:en, lang:hi, lang:bn (language of the forward as detected). Each share = count / the sum of those counters.
- ledger: Rows returned by ledger.load(): the bundled starter ledger plus rows recorded by this server. One row per claim (a forward can hold several claims), so this is not the same number as checks.
- recent false: The latest rows in that ledger whose verdict is FALSE, newest first, with the first citation on each.
- counting since: Counter first_event_at: the time of the first check this store ever counted.
- sources policy: Counted from the lists in sach/fetch.py (OFFICIAL_SUFFIXES, OFFICIAL_HOSTS, OFFICIAL_ZONES, EXTRA_OFFICIAL, FACTCHECKER_HOSTS, FACTCHECKER_PATHS). Fact-checkers are secondary: they are labelled and can never produce TRUE or FALSE on their own.
- evaluation: Read from data/eval_summary.json, which scripts/run_eval.py writes after a full run of the evaluation set. Percent correct = correct / n.