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:
walkaroundasks whether the session entered a frozen contract. Done without entry isBYPASSED. Receipts are unsigned; noVERIFIED.charterlockasks whether that journey was allowed to be the exam. Same key writing and sitting it isCHARTER_COLLAPSED. Two MAC keys do not prove two people.trust-meterscores a checkout from evidence.phaseledgerrefuses to advance a phase without a fresh measurerPASS.nullbenchpre-registers decisions and scores them against chance — never backfill.
No GitHub Release for walkaround / charterlock / trust-meter / phaseledger.
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.
| Stage | Project | Question |
|---|---|---|
| Admit | walkaround | Did this session enter a frozen task contract? |
| Charter | charterlock | Was the journey allowed to count as an exam? |
| Measure | trust-meter | What does this checkout actually score from local evidence? |
| Gate | phaseledger | Can this phase advance without a fresh deterministic PASS? |
| Detect | greenwash | Did an agent make CI green by weakening verification? |
| Baseline | nullbench | Did the decision beat chance, with the claim pre-registered? |
| Re-run | RepoPassport | Did the declared journey work within its capabilities and clean up after itself? |
| Reproduce | stateweaver | Can an independent verifier replay the finding against a clean state? |
| Forecast | tomorrowci · tomorrowci-lab | When will dependency or runtime drift invalidate today’s evidence? |
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
- branchback — decision replay lab; belief-at-the-time vs knowledge-now
- constraint-deck — session-first authorial constraint deck; measure first
| Project | Current public status |
|---|---|
| cell-shift | Public subject, not the flagship. Deterministic 3D tissue chamber (CELL//SHIFT). Maintenance-only. Host CI is green; spine verify is not claimed. Not biology. |
| greenwash | v0.1.41 — deterministic diff-level verification-tampering detector. At the 2026-08-15 reconciliation, main was one README-only Action-pin commit ahead of that tag. |
| nullbench | v0.7.0 — pre-register decisions; score against chance; never backfill |
| tomorrowci-lab | v0.2.0-alpha.1 — project-operated prerelease; CANDIDATE_ONLY_NOT_RELEASE_AUTHORIZED. GitHub “Latest” still points at rejected v0.1.0-grok-session |
| unasked | v0.4.0 — authenticated trust plane; public result remains M0_NOT_DEMONSTRATED |
| null-city | v0.1.0-alpha.1 — playable deterministic agent-evaluation sandbox (prerelease; no GitHub “Latest”) |
| receiptradar | v0.1.0-cli.34 — four-platform CLI release with packaged checksums |
| md-brain | v0.2.0 — public prototype for model-independent Markdown continuity (prerelease; no GitHub “Latest”) |
| aurora | v0.1.47 — alpha evidence engine; observations, not investment claims |
| github-radar | v0.1.0 — reproducible stdlib-only alpha with bounded coverage claims (prerelease; no GitHub “Latest”) |
| nelson-release-studio | v1.0.0 — verified Windows-first local release workbench |
| FutureShow-pet | v0.1.0 — personal Windows alpha; Loop 10 and long-soak automation remain open (prerelease; no GitHub “Latest”) |
| tw-stock-lab | v0.2.0 — local TW stock research lab; paper simulation only, not investment advice |
| branchback | v2.0.0 — local-first decision replay laboratory |
tomorrowci still has GitHub
“Latest” pointed at
v0.1.0-grok-session.
That tag is a rejected historical candidate, not an acceptance claim.
Current lab publication lives in tomorrowci-lab as the v0.2.0-alpha.1 candidate-only prerelease.
| Project | Honest boundary |
|---|---|
| walkaround | No release. M4 local kernel; receipts unsigned; ADMITTED is not verified work |
| charterlock | No release. independence_claim is always not_claimed; two MAC keys do not prove two people |
| trust-meter | No release. Local scorer with batch/compare/API surfaces; self-audit only |
| phaseledger | No release. Phase advance requires a fresh measurer PASS; reclaim invalidates later phases |
| tomorrowci-lab | Newest prerelease is v0.2.0-alpha.1 and remains candidate-only. macOS / Windows clean-machine and independent authorization remain BLOCKED |
| RepoPassport | Working v1alpha1 vertical slice; 37-row acceptance registry is machine-checked; observer coverage remains incomplete, so healthy runs stay inconclusive |
| stateweaver | Source-only pre-alpha; M6–M8 implementation gates exist; trusted Reality proof does not |
| NormShift | M0 implemented; production and release remain blocked pending external audit |
| smallestlie | Authorized adversarial harness; no release |
| constraint-deck | Public source; measure-first voice contract; no release yet |
| editorial-doll-engineering-preview | Public M0–M3 engineering preview of a deterministic styling engine; no release yet |
| universe-explorer | Public epistemically honest science knowledge system; no release yet |
| vibe-oracle | Explicitly not evidence — vibe theater that admits the theater |
| why-ledger | Justified sovereign decisions notebook (WJSD); documentation-first; no release yet |
| Project | What it does |
|---|---|
| nelsoncode-ide | Off-mainline optional personal AI coding preview; external security audit remains NO-GO for untrusted use |
| receiptradar | Receipt-to-ledger CLI with no cloud account |
| md-brain | Model-independent continuity runtime for Markdown memory |
| github-radar | GitHub research with measured uncertainty and zero runtime dependencies |
| aurora | Finds unnamed industries from evidence, with no LLM at runtime |
| music-lab | Deterministic local music toolkit; analysis first, no cloud account |
| nelson-release-studio | Windows-first music, asset, lyric-video, and release-package workbench |
| FutureShow-pet | Personal Windows desktop pet with Taiwan and GitHub information loops |
| tw-stock-lab | Local TW market research desk (charts, sector map, signals + Ollama); paper only |
| branchback | Preserve belief-at-the-time vs knowledge-now for decision replay |
- Deterministic first. The same evidence should produce the same verdict.
- Evidence over vibe. Claims point to tests, diffs, logs, artifacts, or an explicit insufficient-data result.
- Fail closed. Missing observation is not a pass; capability violations outrank functional success.
- Local first. Prefer loopback services, offline-capable CLIs, and local runtimes over mandatory cloud accounts.
- 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: 2026-08-17. Aligned with nelson-stack.