About
An independent research lab.
Gubernaut builds and measures runtime control layers for LLM agents. What it publishes is a validated result and the discipline that produced it: pre-registration, cross-family evaluation, and a public record anyone can recompute.
01 · The name
The steersman.
Gubernaut is a compound of two old words for the same job. The Latin gubernare descends from the Greek kybernan, to steer: the root Norbert Wiener reached for when he named cybernetics in 1948. The lineage this work claims runs back to that founding: closed-loop regulation under pressure, measured in the open.
The name states the position. An external regulatory navigator sits above a powerful engine; it steers while the engine rows. The visual identity follows the same law: geometric, structural, instrument-grade.
gubernare
Latin
To steer, to govern. The root of governor: the small device that keeps an engine from running away with itself.
nautēs
Greek
Sailor, voyager. The one actually at sea in the system while the weather does what it does.
02 · Stance
Evidence first.
Research first
The site reports a measured result and publishes the record behind it, method included.
Pre-registered
Criteria frozen before the run. Failures stay in the record as data: five of them, documented.
Reproducible
Every number traces to a sealed table; every reply ships with SHA-256 provenance and can be re-judged.
03 · Vision
Build the driver.
A hardware-style governor for software, measured in the open.
The vision is a control layer worth trusting: an external regulatory architecture that makes an agent's conduct inspectable and measurable, whatever engine it runs on. Cognitive agency lives in the gap between impulse and response, and the bet here is that engineering that gap matters as much as scaling the engine under it.
The goals are narrow on purpose. Show that runtime self-regulation is a separately measurable layer. Keep the record open: replay, transcripts, judge panels, scripts. Then extend it through a pre-registered follow-up program, one falsifiable run at a time, starting with a static-instruction ablation and a cross-family swap of the appraisal layer.
The hardening track is declared alongside it: a governor-bypass (posture-defiance) battery, per-call Δt and token-overhead accounting, and a human validation panel.
04 · The house plan
One principle, a branded house.
Every future layer inherits the discipline of this one: pre-register, measure, publish the record. One tier is real today; the rest are declared research directions.
Cognitive Layer
Live preview
The governor this site reports: a deterministic control loop, evaluated across four model families, with the sealed record and replay open for audit.
Memory Engine
Roadmap
Durable episodic memory under the same governance, gated by its own pre-registered battery.
Reflection Core
Roadmap
A background reflective loop over that memory, the designated carrier for planning.
Actuator Layer
Research direction
The embodiment question, deliberately held at arm’s length: one roadmap sentence, no capability claims.
05 · Founder
The engineer.
Dushyant Sharma, M.Eng.
Principal Architect & Founder
Gubernaut is built by one engineer. Dushyant Sharma trained in mechatronics: the discipline where mechanical systems, electronics, and software have to be understood as a single machine rather than three separate subjects. That training runs through machine learning, robotics, and microcontrollers, and it leaves a habit this project inherits: treating technology as a whole, because the interesting behavior lives where the layers meet.
Cognitive architecture entered from the other side, as a personal inspiration rather than a career plan: a long-running study of human psychology, of how one regulates attention and impulse under pressure, pursued because the questions would not leave. That origin is kept on this site deliberately. Nothing here was productized first and justified later; the project happened the way it reads.
The governor is where the two threads met. Regulation principles drawn from that study of human cognition were restated as a deterministic control loop an engineer could build, instrument, and measure, then put through the pre-registered, cross-family evaluation this site reports, including all five documented failure modes.
The framing here is secular control theory throughout: monitoring and control, postures and recovery windows, measured effects with stated uncertainty. The architecture's design provenance is documented in the paper's provenance table; it does not change how the system is described or what it is claimed to do.
06 · Principles
Measured, with receipts
Claims carry a pointer: a table, a trace, a log. Effects are reported with their t-statistics and their nulls.
Replay-labeled demos
The cockpit is a recorded run replay with no live API. The audited record is the product.
Weak cells carry weight
The strongest secondary finding came from the weakest cell: the regulation effect is bounded by the host’s intrinsic reactivity headroom.
Scoped to the controller
Injection-resistance is claimed for the token-free controller only; the arbiter reads text by necessity.
Contact
Request a live audit session.
Serious auditors, researchers, and collaborators can request a supervised walkthrough of the live system against their own probes.