Trade inference compute for quality: rather than a bigger model, the model does more work per query, along a smooth cost-versus-quality curve.

Having a bot think for just 20 seconds in a hand of poker got the same performance boost as scaling up the model by 100,000× and training it for 100,000 times longer.Noam Brown, OpenAI, 2024

An agent runs the research loop itself: change one file, run a short fixed-budget experiment, keep the change if the metric improved, otherwise revert. Repeat, overnight. It is hill-climbing, with an LLM as the mutation function.

You're not touching any of the Python files like you normally would as a researcher. Instead, you are programming the program.md files that set up your autonomous research org.Andrej Karpathy, Anthropic, 2026

In-domain mean ΔnDCG@10 climbs from +0.07 to +0.24. This looks exactly like test-time-compute scaling.

In-domain only. The held-out test comes next.

Discovered on j-v5-nano. Mean ΔnDCG@10 is positive on all four encoders and largest on the two families never seen during discovery; the medians sit near zero on the jina-family encoders, so the improvement is concentrated in a positive tail, not broad. The transfer follows general embedding geometry, not artifacts of the discovery encoder.

Trained on the same 14 discovery tasks, a linear, low-rank, or MLP head improves in-domain retrieval by +0.20 to +0.25 ΔnDCG@10, yet falls below baseline on every held-out encoder.

Adding parameters at the same data budget does not transfer. Recombining the frozen geometry does.

Given a token budget, spend it on local small models instead of a frontier model: run search → read → write, repeat, until the knowledge is dumped into one cited .zip. That dump is the dataroom - the open web distilled down to a small, local corpus a machine can consume.

Stage one of two: the grounded .zip then goes to searchbox (next) or a frontier model for the expensive second stage.

An airgapped testbed for search as test-time compute: lock an agent in the box with one .zip dataroom and no web, so it can only answer by composing its own pipeline from local tools - grep, embed, rerank, similarity, cluster, select_diverse. Nothing leaks in; the search has to exhaust what is in the box.

Trivial questions are useless for agentic search: if one grep finds the answer, every method scores the same. To evaluate searchbox you need questions that force the search to actually work.

Turn the corpus into a knowledge graph - each fact a (subject)-[predicate]->(object) edge - then walk its longest paths. Those chains become hard multi-hop questions no single passage answers: a private, corpus-grounded eval, grown from the same corpus searchbox is locked inside.