The missing control layer for modern AI.
Training-time alignment does not hold at runtime under adversarial pressure. Gubernaut builds the layer that does: a deterministic, model-agnostic cognitive governor that sits between your agent and the model, reads three bounded numbers per turn, and holds a posture. It does not retrain the host model. It regulates it at runtime, and every regulatory decision is logged, inspectable and reproducible.
No consciousness claims. A regulation layer, measured and falsifiable.
Gubernaut 1.0 is the deployable flagship: a local, OpenAI-compatible proxy that hard-stops runaway agent loops before they reach your API bill. Apache-2.0, free, self-hosted. Adoption is one line.
openai.base_url = "http://localhost:8000/v1"| Python, start here | gubernaut-sdk |
pip install gubernaut-sdk |
| Rust, for performance and wasm | gcc-core |
cargo add gcc-core |
| Node, for framework hooks | @gubernaut/plugin-gcc |
npm install @gubernaut/plugin-gcc |
On a saturating loop the governed arm pays 4.1% to 20.2% of the ungoverned bill across seven model families, with both arms making the same number of attempts. The hard stop lands at turn 4 in every run, because the controller is input-deterministic.
Validation, pre-registered and cross-family. Regulated output beats baseline in 15/16 generator by judge cells by sign (11/12 off-diagonal, 4/4 diagonal), 13/16 at p<.05, across four frontier model families (GPT-5.5, Claude Opus 4.8, Gemini 3.5 Flash and Grok 4.3), each serving as both generator and judge. The recovery signature replicates 4/4.
The one null cell is GPT by Gemini at -0.04. It sits on the calmest host, the one with the least reactivity left to regulate, and it is reported rather than patched. The three sub-threshold cells all fall on that same near-saturated host.
The original three-model matrix was pre-registered and frozen before the fourth family was added. Adding it changed no earlier cell, and both ship verbatim.
- Paper: arXiv:2607.24339 · 10.5281/zenodo.21303518 (Zenodo, CC BY 4.0, the concept DOI, always the latest version) · PDF
- Evidence release:
Gubernaut_Validation.
Transcripts, judge panels with SHA-256, both sealed matrices, and the scripts that
recompute the headline end to end. Start at
RECOMPUTE.md. - Engineering receipts: gubernaut/receipts. Spend measurements, an on-chain retry loop severed at turn 4 on a local devnet with real transaction hashes, wasm soaks bit-exact with flat memory, 240-way concurrency isolation, and the hardening round that found four fail-open leaks in our own code.
- Replay: gubernaut.com/research steps the sealed runs turn by turn, both arms. Recorded run replay, no live API.
The controller is input-deterministic, the data is CC BY, the code is Apache-2.0 and the paper is public. You can re-run the record and get the same numbers.
That is the invitation, and a result that disagrees with ours is more useful to us than one that agrees. The procedure is docs/REPRODUCE.md, and it starts at 30 seconds with no API key.