Home / Projects
Open code and specifications. Each description says what exists today, including what doesn’t yet.
A 1.7B model (Qwen3-1.7B + LoRA) trained to query a verified oracle built from Minecraft’s own game data instead of recalling facts, and to decline when the oracle has no answer. Recovery-trace training plus a harness fix moved live multi-step crafting from 19/50 to 38/50. An early base-vs-adapter comparison was retracted as unfair; the paired rerun and claim ledger are published.
A verifier-as-reward pipeline for JSON-RPC tool calls on a small invented support-ticket API. One deterministic verifier checks parse, schema, method, preconditions, state transitions, and sequence; the same function grades evaluations and produces the training reward. A strict-C11 port is held in parity with the Python reference. The GRPO training loop is not implemented yet.
A standalone, general-purpose video package format: ordinary media kept as-is, linked to exact sample references, timed annotations, and optional visual summary sheets. Draft 0.1.0 includes a format specification, JSON Schema, a partial conformance checker, and reference packages — not an encoder or an established standard.