1 of 15

DR. ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

> muck

A COMMAND LINE INTERFACE FOR AGENTIC JOURNALISM

COMPUTATION + JOURNALISM SYMPOSIUM · OCT 2026

2 of 15

WHAT MUCK IS

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

02 / 15

Muck is a set of tools

that enable AI agents

to find stories in data.

3 of 15

MISSION · CONSTRAINTS

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

03 / 15

Uncover stories

Don’t just extract facts
Find meaningful information

Run on a laptop

No APIs
No subscriptions
No GPUs

4 of 15

HOW IT WORKS · THREE LAYERS

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

04 / 15

BUILD

index once

peek

ingest

map

parse

index

INVESTIGATE

ask it questions

search

grep

entity

relations

aggregate

anomalies

VERIFY

check every claim

cite

verify

audit

finding

trace

5 of 15

BUILD · WHAT GETS BUILT

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

05 / 15

From raw records to searchable data

Step	What it produces	Count	Time
peek	A field map: which fields become text, names, dates, amounts	setup	—
ingest	Source files registered, deduplicated by content hash	61	1 s
map	Documents, one per record	715,191	77 s
parse	Pages from PDFs and Word files (none in this corpus)	0	16 s
index	Chunks, each embedded and keyword-indexed	2,041,925	~4.5 min
index	Resolved entities and their mentions	73,680 / 6.6M	~33 min
index	Typed relations: donated_to, lobbied_for	280,107	~33 min
		Total	39.6 min

6 of 15

WHAT THE AGENT DOES · EXAMPLES OF MUCK COMMANDS

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

06 / 15

muck status # what's in the index

muck fields # which fields can be searched or counted

muck relations --predicate donated_to -k 12 # who gives money to whom

muck grep "conviction|convicted|pleaded|sentenced|debarred" --all

muck aggregate --by covered_position --agg count -k 12 # revolving-door pool

muck entity "OpenAI" # one entity across all sources

muck verify --token "f6de82d891e.27778@0-1962#5db333c0" --quote "client.name: FRED DAIBES"

muck audit # re-check every recorded finding

7 of 15

TRIAL RUNS · FOUND FACTS, NOT LEADS

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

07 / 15

$ muck finding list # the first finding recorded, before the playbooks

registrant.name: CHECKMATE GOVERNMENT RELATIONS

client.name: INNOVAIRRS & CO.

income: 1000000.00

activity_desc: Issues regarding finances.

A big number. Not a story.

Statistically significant

≠ socially significant

8 of 15

PLAYBOOKS · TELLING THE AGENT WHAT A STORY LOOKS LIKE

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

08 / 15

Muck comes with 12 playbooks that point agents toward stories

Revolving door

Red-flag disclosures

Foreign influence

Peer-relative anomaly

One-shot whale

Say-vs-pay

Undisclosed relationship

Timing anomaly

Concentration / targeting

Follow-the-money

Boilerplate & filing mills

The slip-up

Each one is a hypothesis with the muck commands that test it.

9 of 15

44 FINDINGS · EXAMPLES OF FINDINGS FROM ONE PROMPT

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

09 / 15

9

Foreign influence

state-owned owners, from 2,276 foreign-owned filings

9

Red-flag disclosures

pardoned operatives, from 868 conviction disclosures

9

Revolving door

ex-committee counsel lobbying their old committees

4

Follow-the-money

400 reports fund both parties’ Senate committees

6

Peer anomalies

a $1.1M quarter, 22× the firm’s own median

7

Boilerplate & the slip-up

78% copies; one filer’s time sheet

The playbook helped the agent find newsworthy stories.

For the submission I gave the agent one prompt

to investigate the US senate lobbying disclosures and congressional releases

and produce a findings report of the most significant, defensible stories it contains.

It returned 44 findings,

I dont have time to go into full detail

It found 9 cases of foreign influence, including from foreign state-owned companies

9 cases of revolving doors

And a number of other newsworthy stories

Many of these have already been reported on but the data we were given goes back to 2022

-- so newsoutlets had a bit of a head start.

��The prompt, verbatim:��> You have a corpus of U.S. Senate lobbying disclosures (LDA filings + LD-203 contribution�> reports, 2022–2026) plus congressional press releases, indexed with the `muck` tool (installed�> as a Claude skill). Investigate it and produce a findings report of the most significant,�> defensible stories it contains.�> �> How to work:�> - Orient first: run `muck fields` and `muck status` to learn exactly what is in this index�> before assuming any field is present or absent.�> - Use the muck skill and work its playbooks. Prioritize RELATIONAL stories — follow-the-money,�> say-vs-pay (a member's public statements vs. the money and lobbying around them), the�> revolving door, foreign influence, and peer-relative anomalies — over raw dollar size.�> - Use the drill-downs: `muck entity "<name>"` to assemble one person/org's cross-source picture,�> and `muck relations` to walk the typed graph (who donated to whom, who lobbies for whom).�> - Every claim in the report must carry a citation token you have confirmed with `muck verify`,�> and be logged with `muck finding add` so `muck audit` shows it green. State nothing you have�> not verified against source.�> - Any possible legal or ethics issue is a LEAD requiring human/legal review, not a conclusion.�> �> Deliverable: write `findings-report.md` — findings ranked by significance, each with its verified�> citation(s) and the exact muck commands that produced it, plus the required notes (any outside�> data used, conflicts of interest, legal-flag caveats). Finish by running `muck audit` and�> reporting the verified/failed counts.

10 of 15

FINDING 1 · FOREIGN INFLUENCE

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

10 / 15

Foreign governments and state-owned entities are the disclosed principals behind U.S. lobbying on national-security-sensitive issues

$ muck grep "foreign_entities" --all # 3,108 matches

client.name: HIKVISION USA, INC.

foreign_entities: CHINA ELECTRONICS TECHNOLOGY GROUP CO., LTD

(A WHOLLY STATE-OWNED CHINESE COMPANY) 41.55% CN

description: Prohibitions on certain video surveillance equipment

in the National Defense Authorization Act

Hikvision USA, whose filing lists “a wholly state-owned Chinese company” as an owner, hired lobbyists to influence the NDAA ban on its own cameras.

11 of 15

FINDING 2 · RED-FLAG DISCLOSURES

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

11 / 15

Convicted and pardoned political operatives are active in the registered lobbying corps — often paired with notable clients

$ muck grep "convicted|pleaded guilty|sentenced|debarred|felony" --all # 868 filings

registrant.name: RRB STRATEGIES LLC

client.name: ANNE PRAMAGGIORE

conviction_disclosures: In February 2025, Mr. Blagojevich received a full

Presidential Pardon from President Trump

description: Client is seeking Presidential Clemency.

Rod Blagojevich, pardoned in February 2025, registered as a lobbyist to seek presidential clemency for Anne Pramaggiore, the former ComEd CEO.

12 of 15

PROMPT 2, “WHO LOBBIES ON AI?” · REVOLVING DOOR

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

12 / 15

The revolving door runs into the committees that write AI law — and into the Pentagon’s AI office

$ muck entity "Bobby Cornett" # 20 filings, all Q1 2026

covered_position: Head of Legislative Affairs, Chief Digital and

Artificial Intelligence Office, Department of Defense

registrant.name: ATLAS CROSSING LLC

client.name: AIR SPACE INTELLIGENCE Q1 2026

Bobby Cornett ran legislative affairs for the Pentagon’s AI office, then appeared on 20 Atlas Crossing lobbying filings in the first quarter of 2026.

13 of 15

VERIFY · HOW A CLAIM IS CHECKED

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

13 / 15

Claims are all checked against their sources

token

A result from a document carries a fingerprint that can be used to check the text is really in the document

muck finding add

A claim is recorded with its quote and token, and verified on the spot

muck verify

Re-reads the span, recomputes the hash, checks the quote is inside it

muck audit

Re-runs verify on every finding in the ledger: 58 of 58 passed

14 of 15

WHAT COMES NEXT · THIS IS JUST THE BEGINNING

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

14 / 15

For AI agents

More general adapters: court records, external databases, web search
Efficient models for agentic judgments
More accurate and searchable databases

For journalists

A dedicated user interface
Trace evidence visually
Monitor several agent threads at once
Give agents editorial direction

15 of 15

ULYSSES PASCAL · LARGE LANGUAGE LAB, UCLA

15 / 15

> muck

github.com/upascal/muck

two runs ≈ $30 · 52 min each · runs on a laptop