Eighteen generations of refinement. Every node below is the actual voxel layer the
agent admitted to its knowledge graph — seeds distilled from exploration reports, then crossbred
pair-by-pair into deeper synthesis layers.
Hover any tile to trace its bloodline; click to read the hypothesis behind it.
200×200×8 voxel grid · lon 117.832–117.973 · lat −27.441–−27.300 · 0–80 m depth · EPSG:4326
Geologists find gold not by looking for it directly, but by what accompanies it: faults, altered rocks, chemical halos. That evidence is scattered across decades of paperwork — exploration reports, assay tables, and drill logs.
We fed our AI agents all that paperwork and had them figure out what predicts the gold.
Every hypothesis an agent makes takes a physical form: the terrain is divided into a voxel grid, and the agent writes values and geometry to it. Each value indicates how strongly a feature is expressed at a specific location — probabilities, expected concentrations, relative distances. This converts an idea like “gold occurs near the contact of these two rock units” into a representation that can be tested.
Starting with eleven sets of hard data lifted straight from the reports, everything since is the agents building on those observations: taking two existing clues as inspiration, combining them into a sharper guess, and testing whether the new feature explains all observations so far. If it passes, it joins the graph as material for the next iterations.
The system prioritizes optimal parents. It pairs hypotheses that score well with the hypotheses least similar to them, encouraging agents to reconcile unlike ideas. This keeps the graph from collapsing into variations of a single guess.
Not every pair works: of 1,137 attempts to sharpen the hypotheses, only 60 stuck. Each node is one of the survivors — the agent’s reasoning and the voxel layer it proposed.
And the graph is not the only thing that accumulates. Every curated episode also becomes training data — about a thousand rows this run — used to finetune the underlying model on a sliding window of its own recent experience (a LoRA adapter on gemma-4-31B, retrained as fresh successes arrive and stale ones age out). The agents that build the next graph start from the experience of the agents that built this one: the graph below is both the output of that loop and its curriculum. See the loop animated below.
Together, the graph is the agents’ working theory of the terrain, and of what accompanies the gold within.
Admitted episodes don’t just join the graph — they become training rows. A LoRA adapter (r64) on gemma-4-31B-it-AWQ is retrained on a sliding window of the agent’s own recent successes, so each generation of episodes runs on weights shaped by the last. Schematic — cadence and window size are illustrative; the funnel is real: 1,137 episodes → 110 successes → 1,905 raw → ~1,004 curated SFT rows (target 4,500).