Ballot Brief

Teaching 7,515 candidate case files to tell the truth about severity — without inventing a scale the law never wrote.

A raw number — "12 pending cases" — tells a voter nothing.

Since 2003, Indian election law has required every candidate to swear an affidavit disclosing criminal cases, assets, and liabilities — a real transparency win, made public through ADR and myneta.info. But the data is published as raw legal text and aggregate counts, never classified. An affidavit lists IPC section numbers — "323, 506, 149" — not what they mean; a voter without legal training has no way to know that one of those is a minor offense and another is a serious one, so the actual severity of an alleged crime gets lost in translation before it ever reaches the person it's meant to inform.

Existing tools, including the Election Commission's own "Know Your Candidate," surface the disclosure; none of them interpret it. The gap voters actually face isn't access — it's translation. And translation done carelessly, by inventing an arbitrary "danger score," would just replace one kind of opacity with another. Ballot Brief's working thesis: grade every case against the law's own structure — the CrPC's bail and cognizance rules, the IPC's chapter taxonomy — not a number made up to sound authoritative.

Before building anything: does this already exist, what data is actually available, and what's the legal terrain to build inside?

Comparator research

Five domestic and international projects were reviewed against the same question: does any of them combine structured legal case data with a plain-language severity explainer, grounded in the law's own taxonomy rather than an invented one?

Data-source diligence

Before committing to any feature, every prospective data source was checked for an actual API, license terms, and rate-limit risk at scale.

Stress-testing the idea before scaling it

Three candidates were selected — by a fixed, neutral method (next page of myneta's own case-sorted list, first matching row), never by searching for a specific allegation — to deliberately exercise edge cases before any classification logic touched real scale.

The first design mistake, caught in early critique: collapsing severity into one score. A case needed two independent answers, not one number.

The edge case that defined the display ethic

A pending criminal case is an allegation, not a finding of fact — and that distinction has to hold regardless of how serious the alleged conduct is. So the system was built to apply one rule everywhere, without exception: classify strictly from what the source document says, and let severity and offense-type do all of the signaling — never let the interface itself add a second, unstated layer of judgment on top. IPC §366 — kidnapping or abducting a woman to compel marriage — is the case that put this rule to a real test. The section's text alone doesn't specify whether sexual intent is involved, so a tagging rule was needed: a case is only marked Sexual if the source affidavit explicitly states "illicit intercourse" — tagged from the document's own language, never inferred. And once tagged, that case is still rendered through the same severity and offense-type system as every other Serious-tier case — accurately labeled, fully visible in its own right, with no separate visual emphasis bolted on beyond what the facts in the record already establish. The system's job is to classify precisely and disclose fully; deciding how alarmed a reader should feel about an unproven allegation isn't a judgment the interface gets to make.

House rules (apply to every feature built after this point)

1. Naming avoids editorializing. "Criminal politician" became "Daily Case Spotlight." Names describe the mechanism, never render a verdict.
2. Charges ≠ convictions. Every surface touching criminal cases carries an explicit, visible disclaimer — pending cases are allegations, not adjudicated guilt.
3. Verbatim over paraphrase, for legal safety. Exact quotes with context, not summaries — accuracy is the strongest defense against defamation exposure.
4. Bias-blindness is designed in, not assumed. Sourcing and selection logic apply uniformly regardless of party — flagged repeatedly as the single biggest credibility risk in the whole project.
5. Unverified ≠ guessed. Any legal fact not yet checked against a primary source is marked verified:false and surfaced in the UI — never silently defaulted to a "safe" assumption.

Three real builds, not one — a hardcoded prototype, a scale-up to a real backend, and a full visual redesign, each correcting something the last one got wrong.

The severity ramp couldn't survive contact with real data

The redesign launched with four tiers — Minor, Moderate, Serious, Grave. Reviewing real case records immediately surfaced a problem: armed rioting and weapon-involved hurt were landing in "Moderate." That reads as far too lenient for anyone actually holding a weapon during an offense.

The fix wasn't a fourth tier — it was collapsing to two (Serious / Grave, plus an honest Not Determinable) and defining an explicit rule: a case is Grave if it's life/death-eligible, or involves a weapon, or involves grievous injury — regardless of what a formal sentencing-years field happens to say. The years field measures an offence's ceiling, not how grave the act actually was; an attempted-murder charge shouldn't read as lesser than a completed one just because the attempt's maximum sentence is shorter on paper.

Closing that gap surfaced a sharper one: IPC §303 — punishment for murder by a person already under a life sentence — was struck down as unconstitutional in 1983 (Mithu v. State of Punjab). Its replacement, BNS §104, uses an entirely different numbering scheme that happens to share section numbers with unrelated, still-active IPC provisions. §303 → §104 is safe to map today because old IPC §104 is a private-defence clause nobody would cite as a charge — but that's a narrow escape, not a structural fix. Logged as an explicit open risk rather than quietly patched and forgotten, in keeping with the project's own 'unverified ≠ guessed' rule — applied here to itself.

Visual system, rebuilt around legibility

Two earlier lessons that changed how the project is run

What's actually live, and — in the same spirit as the project's own disclosure ethic — what isn't finished yet.

Disclosed, not hidden

A reflexive read on the project as it stands — not a sales pitch.

A full feature-brainstorm session, mapped on Miro, evaluated against the same house rules established in Exhibit B. View the live board →

Every feature above is evaluated against the same rules set in Exhibit B: names describe the mechanism, not a verdict; charges are never displayed without a conviction disclaimer; sourcing pipelines apply uniformly regardless of party; and dispute-correction flagging isn't a one-off feature, it's a standing safeguard attached to anything that could misrepresent a real person.