Evidence

What has been tested, and what has not.

Plenty of tools in this category are sold on a demo and a promise. This page is the honest state of the testing: what has been demonstrated, the actual numbers from controlled testing, what has been tried and rejected, and what real-world testing still means and when it happens.

Content exposed to the model

In a synthetic routing comparison, opening the specific note that answers a question exposed 4,755 characters across 12 targeted reads — about 65% less text than the 13,419 characters across 6 reads it took to answer the same question by loading the broader source material directly. Measured in a controlled test environment built for the comparison, not on a live installation.

What that costs locally

In warm-cache local testing, doing those extra targeted reads took roughly 0.6 milliseconds longer than reading one large file directly (five repeated runs, 0.518–0.616 ms; 0.524 ms for the single large read versus 1.148 ms for the targeted reads). That is the trade: about half a millisecond more local read time for roughly two-thirds less content reaching the model. This is a synthetic-environment measurement of local file reads, not a claim about real Google Drive latency, which depends on your network and Google’s own service.

A cloud-routing approach we tested and did not ship

We measured a version that re-checked Google Drive directly on every question instead of reading from a local cache: it saved roughly 1,743 tokens of context per question but added about 12.5 seconds of latency, against a live index that itself ran about 4,891 tokens and took roughly 2.08 seconds per round trip. By our own cost estimate that is roughly 24 times more latency than the token savings were worth, so it did not ship. The numbers elsewhere on this page describe the local-cache approach that did.

Automated tests before a release ships

The current release candidate passes 65 automated tests and 5 payment-integration contract tests, a file-by-file integrity check across every distributed core file (58 files, matched by hash), and a strict internal review pass that currently reports zero errors and zero warnings. A delivered package archive was independently confirmed to contain all 59 manifested files.

Updating an existing installation

A rehearsed update from one release to the next applied 10 tracked changes with zero blockers, left existing personal memory files untouched, verified a backup before writing anything, and correctly reported no changes needed when run a second time against the same, already-updated installation.

Catching contradictions in memory

An early internal test set of 7 crafted cases — duplicate notes, direct contradictions, information that had gone stale, and similar problems — produced the 5 expected findings with no false positives and no false negatives. This is a small first test set built to exercise the reviewer, not a broad accuracy claim.