Capture the Narrative 2026

Back to projects
Capture the Narrative 2026
Archived
Python/LLMs/sentence-transformers/LightGBM/NetworkX/PostgreSQL/Docker Compose/GitHub Actions

Top 3 of 168 teams in a UNSW-backed CTF-style contest on AI disinformation, where teams run LLM agents on a simulated social platform to sway thousands of simulated voters.

LLM & data science

39k
fact-grounded LLM replies
~55%
of our official engagement score, from my analysis
90%
precision of the team's bot reporter, fed by my groups
  • 39k fact-grounded replies at 2.1 LLM calls each: built a two-stage plan-then-write pipeline. Every claim must cite a fact id from the supplied sources, and code validates the ids before writing starts, so the model cannot invent facts or links. Each reply keeps a full trace, and the code checks rejected 2.2% of drafts.
  • Analysis behind ~55% of our official engagement score: analysed the platform's users and posts with MiniLM sentence embeddings, KMeans topic clustering, author-fixed-effects Poisson regression and LightGBM with SHAP.
  • Randomised experiments on 4k replies: tested each change with randomised experiments and baseline-controlled metrics, and dropped ideas that failed to replicate out of sample.
  • Bot-network detection behind a 90%-precision reporter: built an action graph of the platform and found other teams' bot networks with Louvain community detection. These groups fed the team's automated reporter, which reached 90% precision in the organisers' rulings.

Backend & infrastructure

1.1 min
median reply latency, down from 16
0.1%
permanent failures over 49k posts
170+/h
sends against a 180/h platform cap
  • Reply latency cut from 16 to 1.1 min: designed a 10-second feed collector that picks up new posts in 6 s instead of 5 min.
  • 49k posts in 14 days with 0.1% permanent failures: owned the publishing pipeline, a PostgreSQL task queue with shared 429 holds and backoff, where timed-out sends are reconciled before any resend.
  • 170+ sends an hour against a 180/hour platform cap: ran the team's 20-service Docker Compose stack on a self-hosted NAS, with no hour missed over 151 h. Every push to main auto-redeploys through GitHub Actions and recreates only the affected services. Cloudflare tunnels share the dashboard and database with teammates, and heartbeat-based alerts notify Discord when a service goes down.

The strategy and code are not public. I’m happy to walk through them in person though :).

Profile Image of the Author
nonameTAT
olleH :)
Site Statistics
Posts
2
Total Words
3,039
Running Days
0 days
Last Activity
0 days ago
Site Info
Build Platform
GitHub Actions
Blog Version
Firefly v6.16.9
License
All Rights Reserved
Table of Contents