I build local-first, deterministic, auditable tools for AI agents and developer workflows.

我在台灣打造本機優先、可重播、可稽核的 AI agent 與開發工具。

Models may propose. Verifiers decide. Missing evidence stays UNKNOWN / INCOMPLETE.

Explore the full Nelson Stack →

The active line is admit, then measure:

• walkaround asks whether the session entered a frozen contract. Done without entry is BYPASSED. Receipts are unsigned; no VERIFIED.
• charterlock asks whether that journey was allowed to be the exam. Same key writing and sitting it is CHARTER_COLLAPSED. Two MAC keys do not prove two people.
• trust-meter scores a checkout from evidence.
• phaseledger refuses to advance a phase without a fresh measurer PASS.
• nullbench pre-registers decisions and scores them against chance — never backfill.

No GitHub Release for walkaround / charterlock / phaseledger. trust-meter has reached v0.2.1. Lab publication remains a measured pre-alpha with explicit NOT_RUN / BLOCKED rows.

The flagship is that audit spine, not a demo. First CI-backed public subject: cell-shift (CELL//SHIFT). Frozen journey and external receipts live in that repo. The chamber's own tests are not the spine verdict.

Supporting evaluation surfaces:

• unasked — evidence-gated repo investigation; non-certifying
• smallestlie — authorized adversarial harness; smallest accepted lie
• null-city — deterministic crisis-response sandbox
• NormShift — evidence-backed standards diffs
• hgsoc-state-engine — domain-transfer case for evidence-gated naming and path admission; not a medical device
• branchback — decision replay lab; belief-at-the-time vs knowledge-now
• constraint-deck — session-first authorial constraint deck; measure first

The public surface currently contains 25 active project repositories plus two meta repositories (taipei49314 and nelson-stack). Private tools are intentionally omitted from this page.

Neither tomorrowci nor tomorrowci-lab advertises a GitHub “Latest” release. The v0.1.0-grok-session tags are rejected historical candidates, not acceptance claims. Current lab publication lives in tomorrowci-lab as the v0.2.0-alpha.1 candidate-only prerelease.

1. Deterministic first. The same evidence should produce the same verdict.
2. Evidence over vibe. Claims point to tests, diffs, logs, artifacts, or an explicit insufficient-data result.
3. Fail closed. Missing observation is not a pass; capability violations outrank functional success.
4. Local first. Prefer loopback services, offline-capable CLIs, and local runtimes over mandatory cloud accounts.
5. Keep the failures. Negative controls, blocked gates, and NO-GO verdicts remain visible instead of being rewritten as success.

Python · Rust · Go · TypeScript · FastAPI · React · Electron · SQLite · Ollama

Older market, persona, creative-production, and agent-console experiments are kept private or archived when they stop being the active line. Market-related projects are paper-only research simulations, never broker or investment systems.

_{Last portfolio reconciliation: 2026-08-21. Aligned with nelson-stack.}