Attention budgets and source grounding built in. One IR, two lowerings — outbound brief, inbound LLM context-pack. Markdown is for storage. EOM is for dialogue.
Markdown is for storage.
EOM is for dialogue.
Today, every human–AI exchange is a flat byte stream: walls of Markdown going one way, walls of model output coming back. Both sides re-derive what matters every turn. EOM carries salience, grounding, and structure on the wire, so neither side has to.
One core IR. Two asymmetric dialects. Shared validator.
Three documents, fifteen questions, two contexts per question. Same downstream model on both cells. Real numbers off a live run.
| raw | pack | Δ | |
|---|---|---|---|
| GDPR | 5,159 | 3,239 | -37% |
| Paris 2024 | 7,488 | 2,348 | -69% |
| RFC 9293 (TCP) | 3,485 | 2,090 | -40% |
| Total | 16,132 | 7,677 | -52% |
Editorially lossy by design. Tier-A always survives. Tier-C compresses. Tier-D drops.
Pages + Workers + R2 (gold corpus) + D1 (qsets + benchmark results) + KV (pack cache) + Workers AI binding.
R2 · D1 · KV · AI · Pages · Workers
Gemma-4-E4B fine-tuned via Unsloth on Modal A100-80GB. r=32, α=32, bf16, 30 epochs, canonical Unsloth recipe.
Unsloth · Gemma 4 · Modal · LoRA
OpenRouter (BYO key) routes /api/ask to google/gemma-4-31b-it. The Ask AI tab is zero-cost-to-host: judges paste their own sk-or-….
OpenRouter · BYO key
Five tabs. Three pre-compiled samples. Live OpenRouter Q&A.
Open the demo →
Spec at docs/SPEC-v0.2.md ·
Writeup at docs/KAGGLE-WRITEUP.md ·
Source on GitHub