Independent replication
Qwen3 4B
84.20%LLM call reduction
84.18%Prompt token reduction
83.69%Wall-time reduction
81.71%Energy estimate reduction
Results
Measured results.
Selected completed experimental outcomes from controlled RHEA Kernel validation work. Published figures are benchmark-specific and are not universal performance guarantees.
Public-results policy. This page reports measured outcomes already disclosed in RHEA Kernel's public validation record. Proprietary implementation details and unpublished supporting material remain private.
83.88%
fewer complete LLM calls
83.94%
fewer prompt tokens
83.53%
less end-to-end wall time
81.16%
lower measured system-wide energy estimate
100%
coverage and accuracy for both baseline and RHEA
Controlled benchmark
Llama 3.2 3B · 6 × 600 queries
RHEA Kernel was measured against a controlled baseline across six rounds of 600 queries while both arms maintained 100% coverage and accuracy.
| Metric | Mean reduction |
|---|---|
| Complete LLM calls | 83.8779% |
| Prompt tokens | 83.9446% |
| Output tokens | 83.5435% |
| End-to-end wall time | 83.5298% |
| Measured system-wide energy estimate | 81.1594% |
Quality preserved
The benchmark recorded the same required quality outcome for the controlled baseline and RHEA Kernel.
100%Baseline coverage
100%RHEA coverage
100%Baseline accuracy
100%RHEA accuracy
Energy measurement note. The published energy figure uses sampled macOS system-wide telemetry integrated over the measured intervals. It is not a calibrated external power-meter measurement and is not process-exclusive.
Cross-model replication
Reproduced across additional model families.
The six-round × 600-query validation pattern was repeated with Qwen3 4B and Gemma3 4B. Output parity and quality were 100% across the recorded rounds.
Independent replication
Gemma3 4B
84.20%LLM call reduction
84.18%Prompt token reduction
81.81%Wall-time reduction
80.38%Energy estimate reduction
These replications support reproducibility of the measured benchmark effect across more than one tested LLM family. They do not establish universal behavior outside the tested models and workloads.
Preregistered paired energy validation
Three controlled conditions · 24 valid paired measurements.
A separate preregistered validation evaluated three distinct workload conditions with 100% quality and output-parity requirements in the recorded valid measurements.
24valid paired measurements
3controlled workload conditions
100%quality and output-parity requirements
5.15–40.21%mean paired energy reduction across the separate conditions
Do not combine the conditions into one universal percentage. They are separate experiments. Each condition recorded 8 positive and 0 negative pairs, with the bootstrap 95% confidence interval above zero and an exact two-sided sign test of p = 0.0078125.
Functional validation
Later functional validation.
On 25 August 2026, a later frozen private runtime received a formal Functional Freeze decision. These are functional-validation results, not energy claims.
171/171hosted validation tests passed
2/2cross-model cases passed
9/9endurance cases passed
5,220/5,220recorded raw events validated
Functional evidence is not an energy claim. No energy claim is transferred to the later frozen private runtime by the dated public validation record.
Interpretation
What the public evidence supports.
Within the stated experiments, RHEA Kernel has demonstrated reductions in complete model invocations, token use, elapsed workload time and measured system-wide energy while preserving the required output quality.
Scope matters.
The project does not claim that the same percentages apply to every AI system, model, workload, company or hardware platform. Results on this page are tied to the specific experiments and measurement procedures described in the public record.
Commercial measurement boundary
Historical results and customer savings are different measurements.
The percentages on this page are historical experimental results, not forecasts of customer savings. Any future customer's Verified Net Savings would be calculated separately from that customer's own equivalent workload, agreed quality gate, RHEA-enabled workload cost and attributable incremental RHEA costs.
No benchmark-to-customer guarantee. A measured reduction in a published experiment must not be treated as a promise that a future customer will obtain the same percentage of monetary savings.