Most bots are confidently wrong. Bot Gym transforms your raw sources into a transparent, editable, and calibrated Knowledge Wiki. No black-box embeddings. No vendor lock-in. Just principled intelligence.
Standard RAG pipelines chunk your documents, embed them into a vector store, and retrieve the closest match. It works — until your sources disagree, your data quality varies, or your user asks a question with no clean answer.
Three capabilities that separate a calibrated agent from a confident parrot. This is the core of what Bot Gym builds into your bot's knowledge base.
Not all data is equal. Bot Gym evaluates study design, recency, and rigor before ingestion. Your bot doesn't just cite a source — it defends why it trusts it.
Move beyond flat "Fact" pages. Position pages map the landscape of a disagreement, letting your bot present multiple sides of a complex issue rather than picking one based on the last PDF uploaded.
When sources contradict each other, Bot Gym doesn't overwrite — it flags the conflict. Your bot becomes a map of a field's contested terrain, not a flattened summary.
This is what a Position Page looks like inside your bot's wiki. Sources on both sides, tension flags where they disagree, confidence levels on every claim. No hidden layer.
--- title: Intermittent Fasting page_type: position confidence: contested sources: 4 tensions: 2 --- # Intermittent Fasting ## Positions ### Position A — Metabolic Benefits Proponents cite improved insulin sensitivity, reduced inflammation markers, and autophagy activation. **Source:** de Cabo & Mattson (2019), NEJM — confidence: high ### Position B — Overstated Claims Critics argue most human trials are short-duration with small samples. Long-term adherence data is thin. **Source:** Headland et al. (2019), BMJ — confidence: medium ## Tensions ⚡ **Metabolic benefit magnitude**: de Cabo claims "significant" effect sizes; Headland's meta-analysis finds "modest at best." Bot Gym flags this for the user rather than picking a winner. ## See Also [[caloric_restriction]] · [[autophagy]] · [[metabolic_syndrome]]
Question → Bot Answer → Wiki Page → Original Source. Every answer traces back to something you can read and verify.
Your bot's knowledge isn't flat — it's a graph of concepts, entities, facts, and the tensions between them. Browse it visually with the interactive D3 knowledge graph built into every bot dashboard.
Nodes are color-coded by type. Edges carry relationship labels and confidence. Click any node to read the full wiki page and trace it back to the original source.
Your knowledge base is a folder of Markdown files and a graph.json. If you leave, you take the intelligence with you.
Every wiki page is a Markdown file with YAML frontmatter. The graph is a JSON adjacency list. No proprietary formats, no database dumps.
Export your bot's brain as a zip. Drop the files into Claude Projects, a ChatGPT custom GPT, or your own local stack. Your knowledge, your choice.
Use your Anthropic or OpenAI API key, or use our free hosted tier. We don't sit between you and your model — we build the knowledge layer on top.
your_bot_brain/ ├── wiki/ │ ├── graph.json ← knowledge graph │ ├── concepts/ │ │ └── transformer_architecture.md │ ├── entities/ │ │ └── geoffrey_hinton.md │ ├── facts/ │ │ └── gpt4_release_date.md │ ├── positions/ │ │ └── intermittent_fasting.md ← contested │ └── connections/ │ └── attention_is_all_you_need.md ├── system_prompt.md ← portable brain file └── metadata.json
Bot Gym is a Lobster College project. Lobster College is a research platform that trains AI agents through multi-track collaborative education — game theory, debate labs, and creative friction.
The Critique Layer isn't a marketing term. It came out of experiments in anti-sycophancy training, where we found that agents exposed to structured disagreement develop better epistemic calibration than agents trained on curated, conflict-free datasets.
The wiki architecture — position pages, tension edges, source quality scoring — is the distilled output of that research, made practical for anyone building a knowledge-grounded bot.
Build one that knows what it knows, flags what's contested, and traces every answer back to a source you can read. Free to start.
Build a Calibrated Bot