You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I started Corridor while I was working on a separate developmental AI architecture. I was trying to understand why certain internal mechanisms were not behaving the way I expected and I needed better instrumentation to see what was actually happening inside the system over time. That diagnostic work eventually became a separate project called Corridor.
Corridor is an independent AI diagnostic system. It’s not supposed to determine whether an AI is conscious or sentient. It’s designed to examine the telemetry produced by an AI system and determine what can actually be supported by that evidence, what cannot be observed and whether important failure patterns are developing over time.
At this time, Corridor has two related layers. The first works at the practical runtime level and can observe things such as agent actions, tool calls, arguments, results, retries, provenance, blocking decisions and task completion. This part of Corridor has already been tested successfully against external agent frameworks, including experiments where Corridor blocked an action and observed the agents recovery behavior.
The second layer goes deeper. Corridor currently has three implemented detectors which I call FS1, FS2, and Invariant Persistence. They are designed to identify structural problems that may develop across many system cycles rather than appearing in a single output. They look at things such as unresolved prediction tension, dominance among competing predictive models, falsification debt, reinforcement decisions and persistent violations of structural invariants.
The difficult part has been applying these deeper diagnostics to independently developed systems.
Across six external validations, I repeatedly found that many existing systems do not preserve or expose the kind of internal telemetry these detectors require. A conventional model may expose token probabilities, loss, hidden states, tool traces or attention values but those measurements are not automatically equivalent to the cognitive quantities Corridor is designed to evaluate. I have deliberately avoided relabeling convenient signals just to make an integration look successful.
That negative result has changed the direction of the project.
Instead of focusing only on retrofitting mature systems, I want to investigate whether advanced cognitive or adaptive architectures can be built with the necessary observability from the beginning. This could involve working with researchers who are still developing cognitive architectures, creating an architecture neutral telemetry specification, and testing whether Corridor can independently observe genuine internal state without requiring every system to share the same implementation.
Funding would support that next stage of research, defining the telemetry and evidence requirements more formally, evaluating additional independent architectures, developing observer instrumentation, working with potential research partners and testing whether Corridor’s deeper diagnostic approach can generalize beyond its controlled environment.
The broader goal is to improve how advanced systems are evaluated. As AI becomes more autonomous, persistent, and adaptive, I believe it will become increasingly important to understand not only what a system outputs, but whether the evidence beneath that behavior supports the conclusions being made about it.
Corridor is an attempt to build that kind of independent diagnostic capability.
The main goal of this project is to determine whether Corridor’s deeper diagnostic system can operate reliably on independently developed architectures rather than only on controlled Corridor telemetry.
The first objective is to define an architecture neutral telemetry
and evidence specification for what Corridor’s deep detectors actually require. This includes longitudinal unresolved prediction tension, influence among competing predictive models or hypotheses, falsification debt, reinforcement decisions, unresolved question state, structural invariants and the provenance needed to understand how those values change over time.
The second objective is to test that specification against independent cognitive or adaptive architectures. Each candidate system will be examined to determine which Corridor quantities are directly supported, which can be legitimately derived from native telemetry and which are simply not represented. Corridor will not substitute unrelated measurements merely to make an integration succeed.
A third objective is to investigate observability by design. Instead of relying only on mature systems that were never built to expose this type of internal state, I want to work with researchers developing new cognitive architectures where meaningful telemetry can be exposed from the beginning. I have already begun outreach to machine consciousness and cognitive AI research groups to explore this possibility. If an appropriate partner is not available, a bounded experimental architecture could be used to test the telemetry requirements without designing the system merely to make Corridor’s detectors fire.
The work would be carried out through compatibility audits, observer only instrumentation, controlled experiments, longitudinal evidence collection, and repeated validation against Corridor’s existing FS1, FS2, and Invariant Persistence detectors. Results would include both successful detections and cases where insufficient evidence prevents a valid conclusion.
Success would mean demonstrating that Corridor can obtain semantically valid evidence from at least one independently designed architecture and showing that its deep diagnostics can operate without changing the underlying meaning of the candidate system’s internal state. Even a negative result would be valuable if it clearly establishes which telemetry requirements cannot be generalized.
The longer term goal is to turn this research into a repeatable diagnostic and assurance capability that can be used by organizations developing increasingly persistent, adaptive, and cognitively complex AI systems.
Funding would primarily be used to give Corridor the time and technical resources needed to move from an independent development into a structured external research and validation effort.
A portion of the funding would support my ability to devote substantially more of my time to Corridor rather than developing it around other work obligations. I have taken the project from its original concept through implementation and multiple external validation efforts, completely on my own but the next stage requires more sustained research and testing than I can support without funding.
Funding would also be used to bring in specialized help where needed. This could include software engineering, AI research, cognitive-architecture expertise, telemetry and instrumentation work and independent technical review. I do not intend to build a large team immediately, the goal would be to add targeted expertise as needed.
Other expenses would include compute and model/API access, cloud infrastructure, storage of experimental evidence, development of observer and telemetry tooling, compatibility testing against additional architectures and reproduction of experiments across multiple conditions.
If an external research group agrees to collaborate, funding could support the engineering work required to instrument its system without interfering with the underlying architecture as well as the analysis and validation of the resulting telemetry.
Funding would also support development of a formal Corridor telemetry specification, documentation, reproducible test procedures and an evidence package that outside researchers can independently examine.
The goal is not simply to add features to Corridor. The funding would be used to answer the central research question, whether Corridor’s deeper diagnostic methods can operate on genuine evidence from independently designed advanced AI systems and whether that process can be made repeatable enough to become a practical assurance technology.
Corridor is a founder led project. I am the primary person responsible for the project’s direction, research questions, system design, validation strategy,, and commercialization work. I use AI assisted engineering tools together with GitHub, Replit, Python-based testing, and structured review workflows to develop and evaluate the system.
My background is not traditional academic computer science or software engineering. My strength has been identifying the problem, defining what the system needs to test, directing development, reviewing results and maintaining a disciplined validation process rather than just accepting results that seem favorable.
My track record is Corridor itself. I have taken it from an internal diagnostic tool for a developmental AI project into an independent diagnostic system with implemented detector logic, evidence schemas, test harnesses, external framework integrations, and multiple validation efforts.
Corridor has been tested across six external AI systems or architectures. Those experiments produced both positive and negative results. The runtime layer has successfully observed and intervened in external agent behavior, including blocking an action and observing recovery through an alternative path. The deeper diagnostics have also been tested in controlled environments, while external experiments exposed a major limitation which is, many existing systems do not provide the longitudinal internal telemetry required for those diagnostics. I chose to preserve that result rather than forcing incompatible signals into the system.
I have also published an SSRN paper that develops part of the theoretical basis behind Corridor, including the concept of structural false relief in agentic systems.
I do not currently have a large formal team. If funded, I would add targeted technical and support where it provides the most value, including software engineering, cognitive architecture expertise, instrumentation, and independent technical review. I am also pursuing research partnerships with groups working on early stage cognitive and machine consciousness architectures.
The most likely way this project could fail is if Corridor’s deeper diagnostics do not generalize beyond the controlled environments where its detectors have already been tested.
The central risk is that independently developed cognitive AI systems may not maintain the kinds of persistent internal state Corridor needs or that superficially similar measurements may not have the same meaning. In that case it may be impossible to create a reliable architecture neutral telemetry layer without changing the system being evaluated.
A second failure mode is that observability by design may prove too intrusive or expensive. Instrumentation could add excessive engineering complexity, create performance overhead or interfere with the internal processes Corridor is supposed to observe.
Another risk is access. The most relevant cognitive or machine consciousness systems may be private, immatureor unavailable for outside testing. If research partners cannot provide sufficient telemetry or controlled access, external validation could be slower than expected.
Commercially, the project could also fail if the technical capability proves useful to researchers but does not solve a problem organizations are willing to pay for.
If the deeper approach does not generalize, Corridor would not simply treat that as a successful result. The outcome would be documented as a limitation. The project could then narrow its focus to the runtime assurance capabilities that have already worked with external agent systems while preserving the deep diagnostic work as a specialized research tool for architectures that genuinely support the required telemetry.
Even in a negative outcome, I expect the work to produce useful results, a clearer telemetry specification, evidence about which internal quantities can and cannot be generalized, compatibility findings from independent architectures and a better understanding of the technical limits of deep AI observability.
I have not raised outside funding for Corridor in the last 12 months. The project has been self funded and developed, completely independent to date.