Project summary

I built Stratum, a programming language that treats AI safety as a property of the substrate rather than as patches on trained behavior. The shipping version already implements capability-gated effects (every side effect explicitly declared and granted), actor isolation with no shared mutable state, runtime anomaly detection through a physics-based signal processing pipeline, behavioral attestation through HMAC-signed state snapshots, automatic quarantine with privileged handler whitelist, and watchdog protection against runaway execution. The language runs on $4 Raspberry Pi Pico hardware and has caught real thermal anomalies through automatic self-healing response.

AI safety today relies on training-time alignment that can fail in ways the deployed AI cannot detect. Structural constraints enforced at the runtime level cannot be bypassed by prompt manipulation, training drift, or deceptive alignment because they are properties of the substrate rather than learned behaviors.

What are this project's goals? How will you achieve them?

The goal is to extend Stratum's existing safety properties into a production-ready AI hosting environment. I will achieve this through three specific deliverables:

Implement biological-logic hardening primitives: Extend safety properties across fleets of actors via stigmergy-based threat signaling, cluster-state topology defenses, and metabolic response scaling.

Implement the cognitive_tick layer: Extend structural constraints specifically to AI cognition through capability-gated output channels resistant to fiction-frame jailbreaks, runtime introspection grounded in actual capability state, and council voices with consonance enforcement.

Empirical Validation: Complete the FPGA conditioning experiment to test substrate memory persistence across cold reboot, validating a key empirical prediction underlying the anomaly detection pipeline.

How will this funding be used?

The funding dictates the speed and scope of the development.

At the $15,000 minimum tier: Funds focus strictly on the FPGA conditioning experiment execution ($300 hardware, $8,000 focused implementation time, $700 paper submission, $1,000 contingency).

At the $30,000 tier: Converts me to full-time focus for six months. Covers stipend ($22,000), self-employment tax obligations ($5,000), FPGA hardware ($1,500), software/submission ($1,000), and contingency ($500).

At the $50,000 full funding goal: Six months full-time with an expanded research scope. Includes stipend ($35,000), self-employment tax ($4,500), FPGA experiment ($1,500), compute/software ($3,000), paper/conference ($2,000), publications/documentation ($3,000), and contingency ($1,000).

Who is on your team? What's your track record on similar projects?

I am Alexander Kalyniuk, an independent researcher operating solo in Edmonton, Alberta, with no formal training in physics or computer science. I built Stratum from concept to working production deployment in six months.

I operate FSME Logic commercially with documented validation on public datasets: 9-month lead time on ESA spacecraft telemetry, 6.3-hour warning on NASA Curiosity Rover actuator binding, and a 78% detection rate with a 126-cycle average warning across 509 NASA C-MAPSS jet engines.

My cross-domain pattern recognition comes from years of careful attention to physics, willingness to test ideas against actual data, and a demonstrated experimental discipline. I recently completed a professional-grade hardware falsification study achieving 186 times signal-to-noise margin on consumer hardware.

What are the most likely causes and outcomes if this project fails?

The FPGA conditioning experiment has a sealed prediction protocol with falsification criteria documented in advance. If the experiment yields negative results (e.g., signal collapses on shuffle audit, or no persistence across cold reboot), it informatively constrains the theoretical framework underlying the anomaly detection pipeline. For the software engineering side, the risk is implementation delay rather than fundamental failure, as the capability-gated architecture and structural constraints are already operational on physical hardware.

How much money have you raised in the last 12 months, and from where?

I have raised $0 in external funding. The project has been completely self-funded, with $2,825 CAD in equipment, $200 CAD cash invested, and approximately $58,000-$64,000 CAD in documented sweat equity over 780 hours.