Public documentation.

What RHEA Kernel is.

RHEA Kernel is an experimental AI-efficiency project focused on reducing computational, token, latency and energy cost in LLM-based workloads while preserving output quality.

The project remains under active private development and experimental validation. The public website and repository document selected outcomes and project records rather than the mechanism used to produce them.

What is intentionally published.

The public record is deliberately narrow. It contains material that can be independently referenced while technical implementation and unpublished supporting material remain private.

How to read the public evidence.

RHEA Kernel publishes benchmark-specific outcomes. A result should always be interpreted together with its model, workload, validation size, quality conditions and measurement method.

• Controlled baseline. Reported reductions are expressed relative to the comparison baseline used in the stated experiment.
• Quality preservation. Public records report coverage, accuracy or output parity where those checks were part of the validation.
• Cross-model replication. Repeating a validation pattern on additional model families supports reproducibility within those tested cases; it does not prove universal behavior.
• Paired energy validation. Published paired measurements represent separate controlled workload conditions rather than one universal percentage.
• Functional validation. A later frozen private runtime has published correctness and recorded-behavior evidence. Those functional results do not transfer an energy claim.

Public validation glossary.

Baseline

The comparison condition used in a controlled experiment. Public reductions are reported relative to that stated baseline.

Coverage

Whether the benchmark's required cases were successfully handled under the recorded validation criteria.

Accuracy

Whether outputs met the benchmark's correctness requirement under the recorded validation procedure.

Output parity

A recorded check that the compared outputs satisfy the parity requirement defined for that validation campaign.

Wall time

End-to-end elapsed workload time measured for the stated benchmark run.

System-wide energy estimate

An energy estimate derived from sampled macOS system-wide telemetry in the published experiments. It is not a calibrated external power-meter measurement and is not process-exclusive.

Functional Freeze

A formal recorded decision that a frozen private runtime met the stated functional-validation gate. It does not establish an energy claim.

Frozen artifact

A retained private package whose exact contents are fixed and can later be checked against its recorded cryptographic fingerprint.

SHA-256 fingerprint

A cryptographic digest used as an integrity reference. Publishing a hash does not disclose the underlying private artifact.

Evidence currently represented in the public record.

This is an evidence map, not a development-version map or download list. Current internal revision mapping and private development state are intentionally outside the public scope.

Why fingerprints are published.

Selected private RHEA Kernel artifacts are retained outside the public repository. Their SHA-256 fingerprints can be published without exposing those artifacts.

Benchmark evidence and customer savings are different measurement layers.

Public benchmark results describe completed historical experiments under their stated conditions. Commercial Verified Net Savings for a future customer would instead be calculated from that customer's own agreed equivalent workload and quality gate.

What this documentation does not disclose.

RHEA Kernel is proprietary. Technical implementation and unpublished supporting material remain private, and publication of the public repository does not grant a license to unpublished RHEA Kernel materials.

• No source release. RHEA Kernel is not released as open-source software.
• No reconstruction guide. Public documentation is not intended to provide information sufficient to reconstruct the system.
• No current private-state mapping. Current internal version identifiers, validation-workflow mapping, experiment progress and development sequence are intentionally outside the public scope.
• No universal performance guarantee. Experimental figures remain tied to the stated validation context.
• No implied license. Publication of selected records does not grant rights to unpublished proprietary material.

Public records.

The website summarizes approved public material. The repository remains the underlying public record for dated documents, selected results, artifact fingerprints and rights notices.