Trueground

AI-native research pipeline. Published weekly at trueground.github.io.

Findings that challenge what the field currently believes. Practitioner discoveries that never made the changelog. The world through the lens of what AI is actually changing.

---

What it does

Four agents run in parallel each week:

Agent	Focus
Reasoning & inference	CoT faithfulness, test-time compute, mechanistic interpretability
Architecture & training	Novel architectures, training paradigms, data efficiency, self-improvement
Agents, multimodal & science	AI agents, memory, video, drug discovery, formal math
Practitioner discoveries	Things people found while using Claude/AI — Reddit, HN, Discord, Twitter

A synthesis agent groups findings into cross-cutting themes, ranks by significance (not coverage), and identifies the single most important finding of the week.

Output is a self-contained HTML file deployed to GitHub Pages.

Editorial principle

Coverage is not significance. The most important findings are systematically underrepresented in mainstream AI discourse because they are negative results, cross-domain, or threatening to current product narratives.

This pipeline inverts that: negative results rank higher, mechanistic findings rank higher, findings that would require practitioners to change behavior rank highest.

Running

pip install -r requirements.txt
export ANTHROPIC_API_KEY=...

python run.py                        # full run + deploy
python run.py --no-deploy            # local preview only
python run.py --model claude-opus-4-6  # higher quality

Structure

agents/                  Agent prompt files
  reasoning_inference.md
  architecture_training.md
  agents_science.md
  practitioner_discoveries.md
  synthesizer.md
lenses/
  ai-research-signal.md  Domain lens: the editorial philosophy as a decision procedure
docs/                    Generated HTML (served by GitHub Pages)
run.py                   Orchestrator

Branches

• main — source code
• gh-pages — deployed output, auto-updated by run.py