Monthly US payroll figures, as first published (dashed) and as revised (solid). The record moves
Suprabhat Rapolu
Hi, I’m Suprabhat. I work on information governance: when a person or an AI system makes a claim, can you see the evidence behind it, and does it still hold? I study that as an independent researcher in India, and I founded Dirmacs, where I design and build Eruka, a knowledge store that tracks what an AI agent knows and what it doesn’t.
My research measures whether knowledge is right, and still right, on official records, automated fact-checkers and AI agents. In one audit, 17.8% of news claims about US economic data that the record supported when they were published are contradicted by the revised record today.
At Dirmacs I also build Eekvul, an analysis engine whose findings carry their sources and their confidence. We make ARES, an open-source agent harness, and Haimdall, a model gateway. Checking what you’re told takes time and money that not everyone has. The mission is to change who carries that cost.
One claim, nine versions of the record
Official records are revised after people cite them. Pick a version of the record.
Real claim · audit dataA news report from a top-tier outlet, published 9 Jan 2026, says:
“The U.S. economy added just 50,000 jobs in December”
The claim matches the record as it stood at this version.
One of 4,669 claims in the audit. Explore them all
Research and building
The research asks whether claims can be checked; Dirmacs builds the tools that make checking possible.
Research
Can a claim be traced to what it rests on, and does it still hold?
- 17.8% of news claims about US economic data that were supported when published are contradicted by the revised record today; 43.3% for monthly payroll.
- Automated fact-checkers score 98.8% on claims copied from the evidence, and 33% or less when one number or word changes.
- An 8B open model with the right version of the record beat a 1.6T open model without it (preregistered): 64.5% of revised claims correct, against 51.4%; a post-review rerun on one host confirmed it (67.4% against 51.6%).
Building, at Dirmacs
Tools that make checking possible
- Eruka: a governed knowledge store for AI agents. Every fact carries one of four epistemic states and a timestamp.
- Eekvul: an analysis engine for public data. Every finding carries its sources, method and confidence.
- ARES: an open-source agent harness in Rust.
- Haimdall: one API across model providers, priced in rupees.
Working previews of the research and the systems
Each one comes from a paper or a system, with the data and code behind it.
Claims explorer
Checked against which record?
Published claims beside the record then and now, with a way to contest the verdict.
TestsCan a claim be traced to the version of the record it rested on?
Payroll revisions
The record moves
How first-print payroll figures were revised, month after month.
TestsWhy the version matters: the record a claim rests on keeps moving.
Financial claim verification
Fool the verifier
Change one number in a financial claim and see whether a verifier notices.
TestsDo automated verifiers really check the evidence?
System
Eruka: a governed knowledge store
A governed knowledge store: every fact carries an epistemic state and a timestamp.
TestsFor the AI systems that use information: every fact carries an epistemic state and a timestamp.
Now
OE-Agent workshop, ACML 2026
Camera-ready submitted
Camera-ready for "Today's Number, Yesterday's Claim" submitted; the workshop is 1 December 2026, Melbourne.
Eekvul
First brief November 2026
Building the analysis engine, towards its first brief.
Restated figures
Model runs pending
Whether language models judge a restated company figure against the record as it stood when the claim was made. Design complete.
Decision models as write-time gates
From November 2026
Building the benchmark, with input from Baalateja Kataru.
Bilingual hate speech detection
Paper in preparation
Consistency across English and Telugu, with Harshith Patnala and Prof. Atul Negi.
News
10 Nov 2026
OE-Agent paper accepted: “Today’s Number, Yesterday’s Claim”, OE-Agent workshop at ACML 2026. Research
28 Oct 2026
FinNLP paper presented at EMNLP 2026. ACL Anthology
Sep 2026
EMNLP 2026 Diversity and Inclusion Subsidy.