How Tathya Shastra works

This section exists because the rest of the page is worthless if this part is vague.

There is no continuous background crawler. Six scheduled jobs are declared in the deployment config, but claim rating is not one of them — the scheduled agent cycle runs the deterministic matcher only. Language-model rating happens when a claim page is requested, capped at a small number of new claims per request, and the result is cached against the claim.

Without a model API key configured, the language-model path is inert and the entire live engine is a deterministic keyword matcher against a fixed list of official series. That matcher is why most claims read unverified: if a statement contains no term the matcher recognises, there is no series to compare it against, and the honest answer is that we have not checked it.

Most stored records are development fixtures. Any figure that is a fixture rather than a retrieved observation is labelled as one wherever it appears. Retrieved series carry their publisher, source URL and retrieval date.

Evidence is ranked, and the rank is part of the record.

Where a Tier 2 or 3 document cites an official series, we prefer the original series as the numerical evidence. A think-tank paper is not automatically correct because it is a think-tank paper.

AI assists research. Sources remain inspectable. A model’s opinion is not evidence and is never presented as one.

The model is constrained to classification against a supplied list of official series. It may not:

• compute statistics — percentage change and CAGR are computed by a deterministic library and stored as calculation records;
• invent a number, a publisher or a URL — those are looked up from the retrieved series after the fact;
• return a verdict outside a fixed set of labels;
• issue a decisive verdict without citing a series that actually exists and has an observation;
• cite a series unrelated to the claim — a relevance check rejects that and downgrades the result to unverified.

Two limits we have not yet closed, stated plainly: the model’s free-text explanation is not machine-verified against the number it cites, and the publication gate that enforces “no uncited verdict” currently guards stored database rows rather than everything the pages render.

Most political statements are not cleanly true or false — they are true about the wrong period, true in nominal terms but not real ones, or true of a state and asserted of the country. Forcing those into a binary is itself a distortion.

A verdict is always accompanied by the reason, the series it was compared against, and how the check was made — hand-checked editorially, matched to a series, or AI-assisted. That label appears on every claim.

Context failures we look for specifically: selective endpoints, base effects, nominal versus real values, percentages versus percentage points, national figures asserted of a state, revised datasets, methodology changes, partial quotation, and per-capita versus aggregate.

The same pipeline must point at everyone or it means nothing. The monitoring roster currently registers 21 entities across government, opposition, state leadership and commentary, of which 10 have ingestion feeds wired. That gap is a real limitation of the current build, and it is shown on the roster rather than hidden.

A person with no monitored feed is not a person with nothing to answer for — it means we are not yet looking. Those two things are never allowed to look the same in this interface.

Party affiliation never determines a colour, an icon or an ordering anywhere in this product. No page ranks people by how often they were wrong.

A verdict is versioned. When one changes, the previous version, the new version, what changed and the evidence that caused the change are all retained — evidence is never silently overwritten.

A public corrections log is not yet built; corrections submitted through a fact check are currently received but not published. Until that log exists, this sentence is the honest statement of where the policy stands.

2014 is a marker on a longer series, not year zero. Series are shown from the earliest available observation. Observed change is not causal attribution to any government.

Any chart with a value axis is zero-based. Where an official series does not exist, the page reads data not available rather than substituting an estimate.

These stay at the level of a specific event, quotation or record. The system does not generalise from an incident to a community, and it does not apply “anti-national” labels to people, religions, parties or publications. Where facts remain genuinely disputed, the verdict says so.