Six years of building at the edges.
From cryptographic provenance and micropayments to AI agents, post-quantum signatures, axiomatic model training, and pairing and SNARK verifiers running in BSV Script on mainnet.
Each date is when the repository was created. Public repositories are linked; the rest are private and available to discuss on request.
How the work progressed
Token and stablecoin experiments, then a durable line of hash notarization, carbon-data anchoring, and IP registration on BSV.
Pay-per-answer LLM micropayments and retrieval-driven smart-contract generation.
LLMs that plan a project’s file tree, then write each file.
OS-level enforcement with blockchain keys, post-quantum hash chains, signed LLM output, and the start of verifier-gated model work.
Ternary-precision weights, cryptographically audited agents, certificate-producing solvers, falsification-first high-dimensional research, and pairing and SNARK verification in layer-1 Script.
Every project
Hash notarization, carbon-data anchoring, and IP registration on BSV. Developed into a durable product line.
Each LLM answer triggers a sub-cent HandCash BSV micropayment, weeks before the ChatGPT API launched.
GPT-3.5 writes BSV smart contracts using retrieval over the entire sCrypt codebase, early in the LangChain wave.
Plans a project’s file tree, then generates each file with an LLM.
A Linux fanotify gate that allows only file operations signed with BSV keys. An unusual pairing of OS-level enforcement with blockchain keys.
Hash chains signed with SPHINCS+ and ML-DSA within months of NIST finalizing FIPS 204 and 205. pq-hashchain followed in July 2025.
Every structured LLM response signed with ML-DSA, when post-quantum signing of AI output was rare.
A transformer trained with Direct Feedback Alignment instead of backpropagation, using Performer FAVOR+ attention and an orthogonality loss, with diagnostics that measure how closely its updates track true gradients. 221 passing tests.
Plain English becomes an LLM-generated, schema-checked program that runs in a restricted interpreter with an operation whitelist and a step limit. Every output is dual-signed with ML-DSA-87 and ECDSA.
A capability-based runtime: Ed25519-signed capabilities, a persistent object store and a small stack VM with its own parser, covered by 117 passing tests.
An LLM writes code in a tiny stack language, and a static verifier checks it before it runs. Using a static verifier as a gate on LLM output was new, and it started my axiomatic-model work.
Litecoin fork with ML-DSA-65 at consensus plus post-quantum multisig, while Bitcoin was still debating how to migrate.
Axiom curriculum with stacked LoRA layers, Forth generation with a verifier repair loop, and Z3 checks.
Delta-sigma trit-stream weights that make ternary precision a dial set at inference time. Extends the fixed-ternary idea of BitNet b1.58 (Feb 2024).
Agent CLI that signs every tool call with ML-DSA, hashes it into a Merkle tree, and anchors it with OpenTimestamps.
QUBO / Ising solver that emits optimality certificates checkable with plain arithmetic, applied to protein side-chain packing.
Concentration-of-measure research, exact-certificate audits, and computational no-go tests. Falsification-first throughout.
Blockchain-timestamped proof of process for human–AI collaboration, early in a new category.
RSA, BLS12-381 pairing, Groth16, and SLH-DSA verifiers in BSV Script on mainnet, plus an equivalence-proven superoptimizer that cut a 333 KB pairing script by 40%.
Blocks 966795 and 966923, ~1.25 MB transactions each. 83 commits, 32 test paths.