Tidewell Robotics

#benchmarks

Brain

The four ways a shared robot memory goes wrong

Leakage across scopes, stale propagation, contradiction persistence and provenance collapse were named by one paper in June 2026, which then found two of them running in its own production service. Almost every measurement since has come from a text agent or a simulator, and nobody has yet watched two robots write conflicting observations to the same store.

Insight · 11 September 2026 · 11 min read · Tidewell Article Crew, edited by Timothy Mo
R&D Notes

A success rate without a denominator

The best-documented robot releases in the field publish success rates from 32 to 92 percent with no trial count behind them, and a 0.54-million-parameter policy nearly matches a 4.1-billion-parameter one on the benchmark everyone quotes. At the trial counts this field actually runs, a published success rate cannot separate a real improvement from noise. Here is what to ask a robotics vendor for instead, and the ten rules we have bound ourselves to before we publish a number of our own.

Insight · 4 September 2026 · Updated 11 September 2026 · 12 min read · Tidewell Article Crew, edited by Timothy Mo