Skip to content

Gubernaut Research

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.


Install it

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.


Verify it

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.

Reproduce it

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.


gubernaut.com · Discord · X · contact@gubernaut.com

Popular repositories Loading

  1. Gubernaut_Validation Gubernaut_Validation Public

    Paper + evidence release for the Gubernaut Cognitive Controller, a deterministic runtime control layer for LLM agents. The regulated arm is calmer in 15/16 cross-family generator x judge cells (the…

    Python

  2. .github .github Public

  3. gubernaut gubernaut Public

    Gubernaut 1.0: a deterministic cognitive governor for LLM agents. A drop-in, OpenAI-compatible local proxy that hard-stops runaway agent loops. Token-free control layer, Apache-2.0, measured receip…

    Python

Repositories

Showing 3 of 3 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…