# Research method: Grok consultations and source verification

Published 7 October 2026. These consultations support the source reviews, not robot-performance or product benchmarks.

## Observation-to-skill review

Three parallel public research prompts ran through Pi 0.87.1 and the exact xAI model `grok-4.7`. Native response metadata recorded 32 X search calls, 157 fetched post items and 52 web search calls. An additional probe recorded two X calls, ten items and three web calls. Items can repeat; these totals do not count unique or independently verified posts.

Pi misclassified hosted search results as local tools after native searches completed. Two normalization passes finished; the simulation normalization pass reached its 480-second deadline after the original completed native answer had been preserved. The report uses those completed native answers and independent primary-source checks.

## Fast simulation review

Three further completed native consultations covered engines, physical contact/deformation and product architecture. Native response metadata recorded 23 X search calls, 95 fetched post items and 45 web calls. A scoped adapter preserved each first completed native response and stopped the redundant normalization request.

## Evidence boundaries

Research inputs contained public prompts and sources. Local model tools and automatic workspace context were disabled. Credentials and request headers were not recorded. Raw sessions and local execution paths are excluded from this public research archive. The counts above summarize local execution records; those raw records are not independently reproduced here.

X results were discovery leads. Exact post wording, author and dates were often unavailable for independent verification. Claims in the publications are tied to checked primary papers, project documentation and source code; cited external results remain author-reported. Code availability inspection does not establish successful reproduction.

The initial spatial-skills synthesis predates successful Grok access and records that limitation. The later review records the successful CLI research. No training or physical robot trial ran during either literature review. The separate contact-fixture publication explicitly reports its 54 native CPU simulation trials.