Question 1

What is RCT Labs' hallucination rate?

Accepted Answer

RCT Labs measures a 0.3% hallucination rate on controlled enterprise workloads, compared to an industry average of 12–15%. This is achieved through SignedAI multi-model consensus verification and the FDIA constitutional gating system. The measurement methodology is: content verification across controlled test workloads, cross-referenced against SignedAI consensus disagreement logs and manual validation sample (n=500).

Question 2

What is the FDIA accuracy score of 0.92?

Accepted Answer

The FDIA accuracy score of 0.92 measures how accurately the FDIA equation predicts output quality versus human-evaluated ground truth, measured on a factual question-answering benchmark (n=1,000). The industry baseline of approximately 0.65 is an approximation based on standard LLM accuracy measurements across comparable enterprise workloads.

Question 3

What does the 4,849/0/0 test result mean?

Accepted Answer

4,849 tests passed, 0 failed, 0 errors — measured on RCT Ecosystem v5.4.5 (March 21, 2026). Tests run across 8 levels: Unit, Integration, Service, Contract, Performance, Security, Chaos, and Property-based tests. The test suite runs on GitHub Actions CI/CD pipeline on every commit.

Question 4

What is warm recall and how fast is it?

Accepted Answer

Warm recall is when the Delta Engine serves a response from its hot-zone semantic cache (similarity threshold 0.95) instead of calling an LLM. Measured from request receipt to response delivery, warm recall achieves under 50 milliseconds. Novel queries always take the cold start path (3–5 seconds). Hot zone capacity is finite; entries migrate to slower zones based on frequency.

Question 5

How does the Delta Engine achieve 74% memory compression?

Accepted Answer

The Delta Engine stores only incremental state changes (deltas) rather than full state snapshots. The 74% compression rate was measured as the average reduction versus full-state storage across 10,000 sequential query sessions. Compression is lossless — full state can be reconstructed with sub-1ms overhead. Short or highly novel sessions may show lower compression ratios.

Question 6

How does the 3.74x cost reduction work?

Accepted Answer

The RCT HexaCore router uses intelligent routing to select the most cost-efficient model appropriate for each task rather than always routing to a premium model like Claude Opus. The 3.74x figure was measured by comparing HexaCore routing versus always routing to Claude Opus across a 10,000-query production-equivalent mixed workload. Actual savings depend on query mix — complex workloads requiring premium models will show lower savings.

สรุป benchmark แบบ public-safe ที่อ่านคู่กับ method และ caveat

ตัวเลขหลักที่สื่อสารได้โดยไม่ตัด method ออก

Hallucination rate 0.3%

FDIA accuracy 0.92

Warm recall ต่ำกว่า 50ms

ข้อจำกัดและ caveat ถูกเปิดเผยชัด

สภาพแวดล้อมการทดสอบที่เปิดเผยสาธารณะได้

หน้าถัดไปที่ควรใช้ประกอบการตีความ benchmark

อ่าน methodology ต่อ

ไปหน้า evaluation

benchmark summary ควรใช้คู่กับ methodology และ evaluation เสมอ