I build systems whose claims can be inspected.
Software, open-source contributions, and technical writing centered on verification, provenance, explicit limits, and reviewable evidence.
Current work & evidence index · Eight merged upstream contributions · Attributed AgentMeasure collaboration
|
Portable coding-agent trajectories without pretending portability is native resumption. ATIF v1.7 reference implementation with namespaced fidelity/loss accounting, Claude Code normalization, OpenInference projection, and explicit boundaries around what cannot survive an agent handoff. Python 3.11+ · Pydantic · pytest · mypy · Ruff · OpenTelemetry |
A traceable decision system built around evidence, counterevidence, and inspectable scoring. Local-first evidence → hypothesis → decision workflows with an independent-oracle differential test suite designed to expose scoring divergence instead of hiding it. TypeScript · React · Dexie · Vitest · Playwright · axe-core |
|
Visual workflows where claims retain a path back to their sources and human review gates. Local-first workflow composition with claim provenance, portable exports, browser verification, accessibility checks, and acquisition-package validation. TypeScript · React · XYFlow · Vitest · Playwright · jsPDF |
Deterministic graph simulation for cascades, sensitivity, and reproducible scenario analysis. A visual simulation lab with graph-based propagation, stability heuristics, reproducibility metadata, and exportable scenario reports. TypeScript · React · Vitest · Playwright · Vite |
|
An offline astronomy atlas with source-level provenance and explicit scientific limits. The current main branch adds the v6 scientific workspace: offline JPL Horizons/DE441 states, numerical propagation, Lambert transfers, and local HEALPix catalogs. The frozen v5.1.0 package remains the latest formal release. JavaScript · WebGL · Python · Gaia DR3 · JPL Horizons · HEALPix |
|
My strongest credibility signal is not repository count or self-reported expertise; it is work that survived outside review.
Eight pull requests merged into two external projects, checked October 15, 2026. The scope of each accepted contribution is linked below.
| PR | Project | What changed | Result |
|---|---|---|---|
| #320 | Grid Dynamics Rosetta | Removed quadratic backtracking from a dangerous-command matcher | Merged after independent equivalence, scaling, and mutation checks |
| #319 | Grid Dynamics Rosetta | Removed unreachable authorization behavior and hardened dataset lookup | Merged after substantive review and correction |
| #322 | Grid Dynamics Rosetta | Added regression coverage for dataset-name resolution branches | Merged after fixture gaps were identified and corrected |
| #299 | Grid Dynamics Rosetta | Extended a dangerous-action guard to equivalent force-delete forms | Merged |
| #338 | Grid Dynamics Rosetta | Bounded command evaluations and terminated descendant processes on timeout | Merged September 29, 2026 |
| #339 | Grid Dynamics Rosetta | Restricted the MultiEdit review override to content actually being written | Merged September 29, 2026 |
| #9619 | super-productivity | Kept visible task order synchronized with persistent movement actions | Merged |
| #10126 | super-productivity | Stopped the schedule day panel from repeatedly snapping back to the current time | Merged September 25, 2026 |
Rosetta #320 is the clearest proof point. A reviewer independently checked equivalence across roughly 5.5 million inputs, then re-verified the corrected head across 1.6+ million differential cases after review changes. The review also reproduced the O(n²) → O(n) scaling and mutation-checked the verification harness.
Read the external-validation case study →
AgentMeasure publicly credits David Turner (@atomicdjt) in DR-005 for the provider-inference boundary: client telemetry can establish what was serialized as sent while leaving what reached provider inference unresolved. Its evidence directory also contains my attributed synthetic fixture for preserving ambiguity between measured and replayed aggregation.
Read the contribution and its limits →
Transparency: OpenClaw #125740 was closed without merge or recorded human approval. I do not present it as accepted upstream work.
Repository taxonomy: the repositories presented as original systems use the shared portfolio identity. Public forks used for upstream contribution work intentionally retain their upstream branding and history rather than being styled as original projects.
The recurring pattern is deliberate: define what must remain true → make the smallest defensible change → build checks capable of falsifying the claim → invite review → narrow the final claim to what survives.
|
What upstream review actually established → A compact evidence record for the merged Rosetta and super-productivity contributions. |
From Claude Code JSONL to ATIF v1.7 → What survives an agent handoff, what is lost, and why interchange is not native session resumption. |
|
Your AI agent finished the task. What did it actually prove? → Separating artifact, behavioral, provenance, boundary, and independent evidence. |
Pull request descriptions that survive review → PR structure designed to expose assumptions, verification, and limitations to scrutiny. |
AI-assisted authorship & claim boundaries
This is an AI-assisted portfolio. I direct product strategy, requirements, scope boundaries, acceptance criteria, verification expectations, and public claims. AI systems assist with implementation, research, debugging, testing, and drafting; I review, revise, reject, validate, and take responsibility for what is published.
Commercial availability does not imply verified revenue, customers, active users, or completed acquisitions. Deterministic scores are heuristics, not certified predictions. Local-first storage is not automatically encrypted, durable, synchronized, or compliant.
If you find a broken assumption, weak verification step, misleading claim, confusing interface, or a simpler design, open an issue.
Portfolio · Repositories · Writing · LinkedIn · Email
