Corrections
The figures below appeared on our site, articles or listings and were withdrawn on 28 September 2026 after we checked them against evidence. Older articles are kept as the historical record, with a notice pointing here. Evidence last updated 2026-10-12.
Withdrawn
The cited benchmark suite and 1,000-prompt dataset are not in any repository, so the figure cannot be reproduced. No hallucination rate is claimed.
A private suite nobody outside can run. We now quote only CI-verified counts from the public repositories.
There is no uptime history or independent monitoring yet, so no SLA is offered. The live status page shows current health only.
74% was a design floor. 91.5% measures agent-state memory in a synthetic simulation, not context. Measured context compression is in the table below.
zk_fdia.py is a hash commitment. A verifier cannot check that the sealed score equals D^I × A, so it is not a zero-knowledge proof.
Logs are hash-chained and Ed25519-signed (tamper-evident). Publishing chain heads to an outside witness (tier A3) is being rolled out.
No LLM was run; the benchmark hardcoded the baseline.
Measured instead
Sources: Whitepaper 3.0 §9 · Claim Registry