BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.com//scipy-2026//talk//9UQN9C
BEGIN:VTIMEZONE
TZID:CST
BEGIN:STANDARD
DTSTART:20001029T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10;UNTIL=20061029T080000Z
TZNAME:CST
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
END:STANDARD
BEGIN:STANDARD
DTSTART:20071104T030000
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11
TZNAME:CST
TZOFFSETFROM:-0500
TZOFFSETTO:-0600
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000402T030000
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=4;UNTIL=20060402T090000Z
TZNAME:CDT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
END:DAYLIGHT
BEGIN:DAYLIGHT
DTSTART:20070311T030000
RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3
TZNAME:CDT
TZOFFSETFROM:-0600
TZOFFSETTO:-0500
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-scipy-2026-9UQN9C@pretalx.com
DTSTART;TZID=CST:20260715T112500
DTEND;TZID=CST:20260715T115500
DESCRIPTION:Automated data enrichment\, filling missing fields in structure
 d records from unstructured sources\, is the canonical case for pointing a
 n autonomous agent at a database and letting it fill every blank. In high-
 stakes data that instinct is dangerous. A confidently wrong value is worse
  than a blank\, and retrieval-grounded extraction reduces but does not rem
 ove the tendency to assert what the source never stated. An LLM can extrac
 t these fields\; this paper asks where agentic judgment earns its place an
 d where it becomes a liability.\n\nWe study this on the Texas Justice Init
 iative's police shooting databases\, where nearly two thousand records are
  missing the weapon\, the subject's race\, or the outcome\, whose fields v
 olunteers typically recover by hand\, fifteen to thirty minutes each. Our 
 LangGraph pipeline\, deterministic in its control flow\, searches\, valida
 tes\, extracts\, and escalates hard cases to a human. It completes 92% of 
 officer and 70% of civilian records and invents zero facts across twenty f
 abricated incidents. The recovery itself came from a deterministic prompt 
 fix without any agent. An autonomous agent pointed at the same fabricated 
 incidents\, with more freedom\, commits a wrong-article fabrication the pi
 peline escalates.\n\nIf the deterministic core does the recovering\, the a
 gentic layer earns its place by making those recovered values trustworthy.
  Agency lives only in this thin judgment layer above extraction\, and its 
 components act in one of two ways. One acts on the pipeline's control flow
 : a relevance judge reads the retrieved articles and\, when none actually 
 report this incident\, routes the record to a human instead of completing 
 it. The other two pass judgment on what extraction produced: one deletes a
  value the source never states\, and the other explains to the reviewer wh
 y the sources disagree on a value. Extraction calls an LLM too\, but becau
 se it only proposes values for these judges to rule on\, we do not count i
 t as agentic. Every judge had to clear a reward-hacking-resistant evaluati
 on gate before it shipped. The main contribution of this paper is a discip
 line\, an "earn-it" protocol\, for drawing the line between what a high-st
 akes pipeline should settle deterministically and where it is worth granti
 ng agentic judgment.
DTSTAMP:20260715T022753Z
LOCATION:Johnson Great Room
SUMMARY:Automated Data Enrichment for Police Accountability: Where Agentic 
 Judgment Earns Its Place - Hongsup Shin
URL:https://pretalx.com/scipy-2026/talk/9UQN9C/
END:VEVENT
END:VCALENDAR