About 31% of the BPO record has the police as a primary or secondary actor. That share has been stable across the dataset, with some elevation in election years and during large enforcement campaigns. The 31% figure is the most-cited statistic the BPO has produced. It is also the most-misread.
Read it carelessly and it implies a third of Bangladesh's violence is committed by the police. That's not what it says. The record codes police as an actor when police are present in the incident — making an arrest, investigating an attack, responding to a call, dispersing a crowd, taking a complaint. Most of those entries are the police doing the job they exist to do.
Three categories that share the line
1. Police as responder. A reported assault, a traffic incident, a missing-person report. Police arrive, write it up, file a case. The incident appears in the BPO record with police as a secondary actor. This is the largest fraction.
2. Police as enforcer. Drug raids, anti- narcotics operations, organised-crime busts, illegal-activity crackdowns. The police are the primary actor; the violence coded in the record is the act of the raid itself, including any casualties incurred. This is the BPO category that overlaps most with the contested term "extrajudicial".
3. Police as party to a confrontation. Political clashes where police take sides, custodial deaths, complaints of disproportionate force during demonstrations, and post-arrest violence. Smaller in count, larger in public controversy. The BPO desk codes these with both V/NV markers and case context where the source supports it.
What the structural comparison shows
Compare the police-involved tail of the record against civilian-only incidents — political clashes, communal violence, intra-family violence, and so on — and a stable pattern emerges. Civilian incidents have lower casualty fill rates (about 25% on injuries, 34% on killed) because the press less often reports a number. Police incidents have higher fill rates (closer to 55%) because police reporting includes arrest counts almost by definition.
That asymmetry inflates the apparent severity of police- involved incidents in any naïve count-by-casualty aggregation. Filter the BPO record to police-involved incidents and the casualty totals look larger; filter to civilian-only and they look smaller, but the underlying per-incident severity is similar. The reporting practice, not the underlying event, is producing most of the visible difference.
What you should read carefully
Two slices of the police record are worth separate attention:
Custodial deaths. A small but recurring category. The BPO desk codes them when reported; the source bias is severe (deaths in custody go unreported routinely). Treat the count as a lower bound.
Post-poll arrest waves. The Election Story on this site discusses these; the police-incident filter narrows further. Spikes here are not "police violence" in the day-to-day sense — they are the operational footprint of a political transition.
For everything else: read police-presence as a coding fact about who was at the incident, not a moral claim about who committed it.