10× token reduction with deino-loom
Context curation via attention budgets that reduces tokens without losing fidelity — a PROJECTED number, and the mechanism behind it.
You do not need the whole graph to answer a query. You need the neighborhood that matters. deino-loom finds it before the model ever sees a token.
The expensive way to ask an LLM about a world model is to dump the whole graph into the context window and hope. It is expensive, it is slow, and it is the thing that breaks the moment the graph is planet-scale. deino-loom is the context-curation layer that sits between the GNL and the sisters: it decides what the model needs to see before the model sees anything.
A number, then the honesty label: we project roughly a 10× token reduction versus a full-recompute baseline. That number is PROJECTED, not measured — it will be labeled measured only once it is measured per workload. The mechanism, though, is real and shipped.
Bounded awakening, not full dump
The core idea is bounded awakening: a query awakens a budgeted neighborhood of the graph, not the whole thing. The budget is the attention window the model would have spent anyway, but spent on a neighborhood that actually bears on the question. Per-query cost is bounded independent of total graph size — which is the property that makes the oracle tractable at planetary scale, where a full dump is not just expensive, it is impossible.
Curation preserves the manifest
The trap with context curation is that it can make the output irreproducible: if the model saw a curated slice, an auditor replaying the query needs to see the same slice. deino-loom records the exact NodeVersions it selected into the manifest, so the curation is part of the provenance. You can replay not just "the model said X" but "the model said X after seeing exactly these versions." Curation and reproducibility are not in tension here; they are the same record.
Curation and reproducibility are not in tension. The manifest records the exact slice the model saw — curation becomes part of the provenance.
- Bounded awakening: per-query cost flat in total graph size — the tractability property.
- Curation is logged: the selected NodeVersions are part of the manifest, so replay sees the same slice.
- Token reduction is PROJECTED (~10× vs. full-recompute) — labeled measured only when measured per workload.
What fidelity means here
"Without losing fidelity" is the part that earns the word "loom" and not "filter." A naive filter drops nodes the query needs and the answer drifts. deino-loom’s selection is driven by the query and the graph structure together, so the neighborhood it awakens is the one the answer depends on. Fidelity is measured the same way everything else is — against the realized timeline, via the calibration loop. We do not claim the reduced-context answer is as good as the full one; we measure whether it is, and the gate (Gate 1b) is what lets us say so.
The number is projected; the mechanism is shipped. We will not call the 10× measured until it is — and until then the label stays.
— deino-loom note