A library of explorations.
Most of these repos are methods, experiments, protocols, or architectural patterns. What they share is that each one outlived the investigation that produced it, and turned out to be useful somewhere other than where it started. Sometimes, the most useful outcome had little resemblance to the question that started the investigation.
Curiously, none of this was planned. They began as long, meandering investigations; most went nowhere and were never written down. The ones here are simply the explorations that left something reusable behind.
Most of the work stays private — working notebooks, where ideas are allowed to be wrong. What's public is just in case someone outside its original field could pick it up and use it.
These repositories span different domains because the questions happened to lead there. Some readers will only care about one, and that's perfectly fine. Others may notice similar structural patterns appearing across unrelated problems. Neither is the "right" way to read the collection — take what's useful and leave the rest.
📫 leenathomas01@gmail.com · Medium
These repositories aren't meant to be read in order, and they don't add up to a single framework. Each is meant to stand on its own. Some later turned out to rhyme with ideas from very different domains — but those connections emerged during the work; they aren't prerequisites for understanding any individual repository.
Listed in no particular order. The tag beside each name is only there to help you spot your own field.
connector-os · control systems · human-AI
What it is: A control-theoretic wrapper that puts a raw model inside a regulated feedback loop (sensors → control → actuators → feedback), with a no-code example that runs on a live physiological signal and needs no model at all.
Why it travels: It's a general recipe for making an unpredictable component behave inside a closed loop without modifying the component.
Borrow: Gate behavior by threshold bands on how far a live signal has drifted from its baseline.
designing-for-failure · reliability engineering
What it is: A pattern language arguing the real failure mode in high-stakes systems is recovering too early, formalized as five primitives and a five-state model with enforced hysteresis.
Why it travels: Any system that must decide "are we out of the woods yet?" faces the premature-recovery trap.
Borrow: Enter crisis states instantly, but require sustained, multi-signal evidence before you're allowed to exit.
gii-protocol · power systems
What it is: A protocol (Grid Inertial Interface) that lets edge devices emulate the physical inertia spinning turbines once gave the grid for free — now missing as inverters and batteries replace mechanical generators.
Why it travels: It's a clean case of restoring a stabilizing property a system used to get "for free" from physics, once you swap the physical thing for a digital one.
Borrow: Separate a fast layer from a slow layer and let them interact only through a physical quantity, never through control signals.
phase-aligned-inquiry · research methodology
What it is: A protocol that separates the generate / execute / falsify / interpret roles and pre-commits pass/fail criteria, shown on a case where the hypothesis was cleanly disproven.
Why it travels: Confirmation bias afflicts any investigation, solo or team, in any field.
Borrow: Decide what would prove you wrong before you start, and keep an idea's authors away from judging it.
Spectral-Storage-Systems · storage · associative memory
What it is: A simulated model that stores identities as interference patterns retrieved by correlation, and treats reachability (can you get to it?) and correctness (is it the right one?) as two independent conditions.
Why it travels: Any retrieval system that ranks by similarity quietly assumes "closest = correct" — this shows why that assumption breaks.
Borrow: Check separability and correctness independently; high similarity alone never guarantees a correct match.
Stability-Before-Alignment · AI safety architecture
What it is: An architecture for keeping a self-modifying system coherent regardless of whether its goals are correct, built on layered invariants and revocable ("leased") authority.
Why it travels: Any system that can rewrite parts of itself faces the question of what must stay fixed for it to stay trustworthy.
Borrow: Non-Reflexive Evaluation — never let the thing being evaluated modify its own evaluator.
Unlearnable-Interference · signal processing
What it is: A runnable demonstration that adaptive interference cancellation fails against signals whose structure resets faster than any learner can align — not because the model is too small.
Why it travels: Any adaptive system chasing a moving target can hit this wall; the limit is the target's rate of change, not your capacity.
Borrow: When a problem is discontinuity-limited, stop scaling the model and measure the rate of structural change instead.
Transition-Grammar-for-Reasoning-Systems · reasoning systems · telemetry
What it is: A telemetry layer for reasoning systems that monitors the trajectory of reasoning rather than judging the final answer. Using simple geometric signals derived from embeddings, it detects failure modes such as Stall, Sinkhole, and Fog before an output is committed.
Why it travels: Many systems only evaluate outcomes. TG asks whether the process itself is healthy. That distinction extends beyond LLMs to any adaptive system where state transitions matter more than isolated decisions.
Borrow: Separate observability from governance. Before deciding how to intervene, build instrumentation that can tell whether the system's trajectory is stable, uncertain, or collapsing.
hyperloop-fxso complex systems · emergence
What it is: A runnable exploration of whether simple overlap-based interactions between locally evolving agents can produce stable collective behaviour without explicit message passing. The repository develops the idea from an architectural hypothesis into a minimal simulation, empirical phase map, and a set of bounded claims.
Why it travels: Many systems—from swarms to distributed optimisation to multi-agent AI—ask whether coordination requires communication, or whether it can emerge from local interaction alone. The repository is less about answering that question than about turning it into something testable.
Borrow: Start with the smallest dynamical system that can falsify your intuition. Keep the negative results. Let the hypothesis shrink before it grows.
Alongside these, FieldNotes is an open notebook — thoughts and analysis in progress rather than finished ideas. It's the one notebook I keep public.
These are architectural reasoning artifacts, not production systems. Evaluate independently and apply your own domain risk assessment.
Note: Updates may be retrospective.