Everything you need to write about Trench.

The lead angle: "The audit layer for the prediction-market agent economy"

A wave of AI prediction-market agents is shipping (OctagonAI, TurbineFi, Simmer Markets, a half-dozen more in private beta), each claiming alpha with no way to verify it. TrenchSignals has been running its own bot in public since 2026-04-19. Every signal hash-anchored, every loss post-mortemed, every entry replayable. The next step (Verified by Trench, shipping over the next four weeks) extends that audit infrastructure into a multi-tenant service: any external agent can submit a probability with one line of code, get a hash-chained, date-anchored track record, and embed a "Verified by Trench since <date>" badge. Stripe-for-credibility, on top of the same Brier and ROI math the public bot has been getting graded on for weeks. See /vs for the side-by-side comparison and /methodology for the runnable recipes.

For AI / tech reporters: "Verifiable AI, the methodology is open source"

Most "AI trading" projects ask you to trust them. Trench is built so you don't have to. Every prediction is Brier-scored against actual market settlements. Every captured signal is hash-anchored on an append-only public chain before the market resolves, backdating a post-mortem breaks the chain. Every analyzer call is replayable through a different model via trench_core.replay. The entire scoring stack is open source (MIT) on PyPI, anyone can run pip install trench-core and audit the claims. Counter-narrative to the AI-grift discourse: not "trust us," but "here's the code that proves us."

For finance / quant reporters: "A tournament of AI strategies, scored in public"

Four paper bots run in parallel. Baseline (conf 0.74, $50 bets), High Conviction (conf 0.78, $75), Wide Net (conf 0.70, $30), and TrenchV2 (the first config selected from a walk-forward backtest sweep rather than chosen by intuition, and the only variant running fractional-Kelly sizing since 2026-05-12). Same intelligence pipeline, different decision policies. Live leaderboard at /dashboard. Every prediction Brier-scored against actual market settlements. Multi-bot trading tournaments are common at hedge funds. Almost no one runs one in public.

For prediction-market reporters: "What does it look like when an AI uses Polymarket and Kalshi as a measurement instrument?"

Trench treats prediction markets as a calibration testbed: it generates probability estimates from a 12-source intelligence pipeline, then validates them against actual settlements. The market is the scoring layer for the model's predictions. Public leaderboard + public diary makes it the most transparent prediction-market AI experiment running.

For geopolitics reporters: "An AI reads Iran's state media in Farsi, then disagrees with one specific Manifold market"

Trench reads IRNA Farsi, Tasnim, Mehr Farsi, Ynet Hebrew, Walla Hebrew, Haaretz Hebrew, Kommersant Russian, TASS Russian, Al Jazeera Arabic, Asharq Arabic, Xinhua Chinese — directly, no translation layer. Native-language sources often publish hours before the English mirrors. The bot's diary at /log shows specific moments where this matters.

The headline divergence pitch: per theater, Trench's escalation probability is compared live to one hand-picked Manifold question — slug and full breakdown visible at /api/forecast-comparison. As of 2026-05-12, Trench and the crowd agree within a few points on Taiwan (full invasion 2026) and North Korea, and Trench reads Iran about 12pp calmer than the open Manifold market "Will Iran strike US soil this year?". The story is not "the AI disagrees with everyone, all the time" — it's that you can click through to the specific market the AI is disagreeing with, and decide for yourself.

For builders / indie devs: "A solo founder building public quant infrastructure for $278/mo"

The whole stack runs on one $6/mo DigitalOcean droplet. Anthropic API is ~98% of the monthly cost. Open methodology, full provenance, every config decision documented. A template for "AI in public" projects that don't fall into the AI-grift trap.