Zero AI-Discovery Frameworks Won Everywhere in a 3.1-Million-Run Study. Here's Everything Else That Shipped.

Zero. That's how many of the 30 AI-discovery frameworks won across every problem in a study published July 20, built on 3.1 million-plus LLM rollouts — not even OpenEvolve, the field's default choice. We go deep on that, and two papers making the same case from different angles, in today's feature. Here's everything else that landed this week.

The framework study, in one line

Automated Discovery Has No Universally Superior Harness (arXiv 2607.18235) decomposed OpenEvolve and TTT-Discover into their constituent design choices and tested 30 resulting variants across 12 model-problem pairs. No fixed harness reliably wins everywhere, and OpenEvolve variants generally underperform simpler alternatives. The fix that beat picking a framework and sticking with it: start several at once, and reallocate compute toward whichever is winning early. Full breakdown in today's feature.

A materials-discovery agent whose gains survive real holdout data

Auto Research for Materials(arXiv 2607.17100) ran seven closed-loop searches producing 701 evaluated changes across ten Matbench endpoints, then froze each selected change and tested it exactly once on data the search loop never saw. 9 of 10 held up, with held-out MAE reductions of 17.4% on band gap and 18.6% on steel strength — and 26.3% when separately found feature and model changes were combined. The one that didn't survive is the more interesting number: without that holdout split, it would have been reported as a win.

One planted document cuts research-agent accuracy by up to 88 points

DRNOISE(arXiv 2607.17291) is a 100-task benchmark that seeds one ordinary-looking, plausible-but-wrong document into an otherwise well-supported evidence set. Agents that score well on the clean version of the same task drop 66–88 percentage points once that single document is present. The dominant failure mode, per the trace analysis: agents retrieve the correct records, then stop short of reconciling them against the conflicting claim instead of resolving the contradiction.

NVIDIA evolves a 3.6x-faster wireless equalizer, no neural net required

The AI Telco Engineer(arXiv 2607.17762), from an NVIDIA research team, applies LLM-driven evolutionary search — the AlphaEvolve-style paradigm — to physical-layer wireless algorithm design. It produces an OTFS equalizer with 3.6x lower computational latency than the strongest known baseline, and derives explainable, non-neural OFDM receiver algorithms that match state-of-the-art neural receivers without using pilots.

Cost becomes a first-class variable in discovery loops

CADAQUES(arXiv 2607.16127), an MIT-licensed open-source framework, charges every query and every planning decision against a shared budget spanning wall time, CPU hours, dollars, and LLM tokens, with an append-only ledger recording declared versus settled cost per transaction. Metering adds "tens of microseconds per iteration, three orders of magnitude below the cheapest oracle query" — overhead cheap enough that cost-aware scheduling stops being a trade-off.

Anthropic opens $50K rare-disease research grants

Anthropic's AI for Science program is accepting applications through August 2, 2026 for up to $50,000 in Claude API credits over six months, across a basic-science track and a biotech track. Existing grantees cited include Every Cure, using Claude to screen drug-repurposing candidates across millions of compounds, and the Centre for Population Genomics, which built a Claude-based system that drafts variant classifications for expert review.

What this means for reproducible, local-first science

Six unrelated releases this week, and every one of them is about rigor: cost accounting, holdout separation, adversarial stress-testing, and honest framework selection all showing up as first-class concerns in the same seven days. That's a healthier direction for the field than another leaderboard topper — and it's the same direction a reproducibility record and a held-out check push toward at the scale of a single lab bench.