Prompting and steering AI
I turn an intention into clear questions, instructions, and constraints. I give a model the context it needs, challenge its assumptions, and refine the approach when it misses the point.
Back to home
Curated capabilities
More than I could do alone.
I curate and develop capabilities by combining human judgment, AI, formal methods, and custom tools into workflows that produce useful, checkable results.
These are human and machine capabilities: ways to think, learn, create, explore, and act with the help of AI and other tools. Prompting, research, software, and visual explanation all play a part. My projects show some of these combinations in use; others are still taking shape.
Intelligence augmentation
Human and machine, working together.
I think of this as a practical form of cyborgization: extending my thinking and my ability to act through machines. I bring purpose, context, and judgment. Models help me explore possibilities, and tools help me calculate, build, remember, and check.
I use these combinations to understand, discover, create, and act. I want to keep expanding those abilities as AI advances and changes how people live.
| Part | What it brings |
|---|---|
| Human | Purpose, judgment, context, and responsibility. |
| Machine | Generation, search, memory, calculation, and execution. |
| Together | New ways to think, learn, create, and act, with checks suited to the task. |
Capability map
Showing is better than telling.
Here are some of those capabilities in practice, with examples you can explore. The labels distinguish current use, public evidence, and experiments still to be completed.
U: unaided, based on my account unless independently proctored. A: assisted by AI, with the tools and process disclosed. V: finding and correcting errors in results generated by AI. Benchmark scores are reported only after a completed run has been reproduced.
Learning and connecting ideas
I explore unfamiliar subjects, compare explanations, and connect ideas across philosophy, computation, governance, and economics. Questions and independent checks help me find out where an explanation holds up.
Creating and explaining
I use writing, diagrams, interactive examples, and software to make difficult ideas easier to understand. I show how an idea works, what supports it, and where it falls short.
Neurosymbolic research workflows
I develop research questions with LLM assistance. Models help me explore and recombine possible hypotheses, witnesses, proofs, and designs. I then test the resulting claims with symbolic tools and reproducible experiments.
I guide the investigation using my custom Research Kernel Protocol MCP, LEAP, and PopperPad. Lean, Z3, Tau, and other deterministic tools check formal claims. My scientific practice emphasizes attempts to falsify competing hypotheses. Skills include hypothesis development, abstraction design, experiment design, and evaluation of results generated by AI.
Research verification & fact checking
I trace claims back to their sources, look for evidence against them, and record what they depend on. Research Kernel and PopperPad help me keep that record so I can revisit the conclusions.
Formal verification
I use Lean, TLA+, deterministic verifiers, and replay to check precisely stated properties. I explain the assumptions and what those results establish about the system.
AI workflow & agent engineering
I build systems that let models suggest actions, with explicit rules and checks controlling what can run.
Software QA & adversarial testing
Define invariants, construct failure cases, reproduce defects, test tamper/replay boundaries, and distinguish passing tests from justified system claims.
Cryptographic implementation security
I investigate post-quantum signature verification boundaries through reproducible experiments, sanitizer testing, and explicit API contracts. My published SQIsign study documents undersized-input decoding and overlong detached-signature acceptance at a pinned source revision.
The SQIsign team had independently developed equivalent fixes before my report. Later Round 3 validation was reported by yui9696; trancee documented equivalent protections in an independent Rust implementation. The findings concern implementation behavior, not a mathematical break in SQIsign.
AI evaluation & model QA
Turn ambiguous quality requirements into testable criteria; find hallucinations, regressions, instruction failures, brittle scoring, and evaluator mistakes.
EVAL-001: build an LLM Regression Sentinel with reference labels reviewed by a person, deterministic checks, failure taxonomy, adversarial tests, and a memo explaining whether the result is ready to use.
Data analysis & decision support
I plan to demonstrate how I check messy data, reproduce an analysis, explain a decision using numbers, and identify misleading averages or totals.
Search relevance & rubric application
I plan to test how consistently I can rate search results, resolve disagreements, and compare my judgments with a reviewed reference set.
Open source · Accepted upstream work
Open Source Contributions.
From digital rulemaking to compiler correctness and formal verification, I investigate failures against intended behavior, identify their causes, and work toward checkable corrections. The entries below include only merged pull requests or contributions that upstream maintainers explicitly recognize as resolved in their main branch.
Evidence standard: An open proposal, closed issue, or passing test alone is not enough for this list. Each entry links to a merge record or the project's published contribution record. Issue-based recognition is kept distinct from authorship of merged code.
Tau Language (IDNI)
Contributed diagnoses, reproducible failures, and proposed repairs spanning solver correctness, quantifier elimination, exact rational arithmetic, performance, and interpreter reliability. IDNI's official record credits 30 issue-based contributions in Q3 2026 and two in Q1 2026, each tied to an issue solved in main.
The 32 accepted contributions are issue-based. They do not represent 32 pull requests or a claim that I authored every upstream fix.
Dafny
Contributed proofs that decimal, hexadecimal, and custom-alphabet numeric conversions round-trip correctly, including the required digit-alphabet and parsing conditions. The change preserves the existing public refinement contract and replaces the hexadecimal character-consistency axiom with a proof.
AI generated the code; I reviewed it and Fabio Madge's additions, accepted selected changes, and guided the work toward maintainer preferences and repository conventions.
Verus
My merged fix preserves source function types in SpecFn lambda wrappers and restricts application semantics to valid inputs. The contribution includes regression coverage and addresses the reported source-type preservation issue.
Kontrol (Runtime Verification)
Fixed incremental builds that could reuse stale required lemmas after their source files changed. The merged patch refreshes changed or missing copies by content, invalidates compilation-success markers when necessary, and preserves correct caching for unchanged inputs. Added regressions for source changes, cache reuse, failures, and retries.
Contributed a documentation link to an evolutionary Nomic approach to digital rulemaking in the constitution repository. Merged PR #8.
Scaling · Development direction
How far can this go?
As AI becomes more capable, I want to expand what I can understand, create, and act on. Today I guide the process. Next I want to teach more of it to a coordinated swarm of agents through prompting, tools, and eventually training.
Stronger models and more compute may let me explore more ideas at once. I want to find out which parts of my approach can be passed on to agents, and whether the results hold up when I do.
- Transfer the method
Describe how I choose questions, build abstractions, look for counterexamples, and assess evidence so agents can follow the process. Specify which decisions require my review.
- Coordinate more agents
Let agents explore and implement in parallel, with Research Kernel preserving the research record. Keep deterministic checks and independent evaluation central as the volume of proposals grows.
- Measure the gain
Compare the process run by agents with the process I guide today. Increase the use of approaches that produce better results under the same budget and checks.
What successful scaling would mean
More useful discoveries, creations, and solutions that stand up to independent checks, for the time and compute spent. I also track mistakes, unsupported claims, repeated effort, and what I learn from failed attempts. More text, or more agents agreeing, doesn't tell me that the process is better.
Current status
I already use the tools and guide the process myself. Passing the whole process to an agent swarm, or training it into a model, is a research goal. I still need to test how much can be automated reliably and what it costs.
I share the principles and the evidence here. I keep the detailed orchestration and training recipes private.
Evidence standard
Check your work.
Each kind of evidence answers a different question. A Git commit records a version; it doesn't establish whether AI was used. Repeating an experiment checks whether a result can be reproduced. Independent review asks whether the result supports the conclusion.
Where possible, the evidence chain is: task → submission → rubric → verifier → receipt → live defense.
The required artifact exists.
A clean run reproduced the committed result.
The benchmark met its score threshold with no unresolved critical failure.
An identified third party independently evaluated or proctored the work.
Benchmarks and practice
One step at a time.
I also use a capability ledger to practice and test specific abilities. Some benchmarks prepare me for paid projects. It keeps one benchmark active until the result meets its completion criteria and can be reproduced.
For career training, I choose the next benchmark by looking at demand and abilities I still need to demonstrate. I also pursue questions, experiments, and creative projects for learning and discovery.
- 1. Observe
For career benchmarks, look at current roles and the tasks they have in common.
- 2. Build
Use AI and other tools to make something small enough to finish and substantial enough to test the relevant abilities.
- 3. Verify
I review the important judgments, check the scoring, repeat the experiment, and disclose how AI was used. I test the evaluator itself and explain my decisions in a live review.
- 4. Graduate
I publish the result when it meets the evidence requirements. If it falls short, I keep working on the same benchmark.
Explore together
What could we do together?
A question to investigate, an idea to make tangible, a tool to build, or a new way to combine human and machine abilities. Start with something worth exploring and a way to tell whether we're making progress.