Curator of open-source infrastructure, research, and tools for the AI era.
The site at heznpc.github.io is the rendered hub. This README is the same map in text.
Brand 5 — the five flagship axes:
| Axis | Repo | What it is |
|---|---|---|
| Product | AirMCP | One MCP server for the Apple ecosystem (272 tools, 29 modules). Core infra — per-call HITL, HMAC-chained audit, OAuth 2.1 + Resource Indicators, allowNetwork policy — is the differentiated layer. |
| Research | ploidy | Cross-session debate protocol — same model, different context depths. PyPI ploidy 0.3.3. Paper + tool monorepo. |
| Trust | canary | Operator-machine observability — git state joined with Claude Code session transcripts. APL / MIP / PLR / UCP push-leakage metrics. |
| Traction | skillBridge | Translates Anthropic Academy lessons across 33 locales with an inline Claude tutor (Puter.js). MV3 v3.5.39. |
| Signal | anvil | MCP tool that bundles commit → PR → CI-wait → merge into one atomic call. Published as @heznpc/anvil on npm. |
5 Research Programs — each with an anchor and companions:
| # | Program | Anchor | Companions / bridges |
|---|---|---|---|
| P1 | Human-Controlled AI Systems | tools side: AirMCP / canary / anvil | statute-of-limitations (bridge 1↔4), oncology, ochre, villagent |
| P2 | Epistemic Failure and Correction | ploidy + meta (reflexive) | narcissus, coup, eddy (DOI), lifespan (bridge 2→4) |
| P3 | Representation, Language, and Cultural Cognition | z-gap | third-vertex-llm, macaronic, habitus |
| P4 | AI-Mediated Accumulation | tidal | silo, scatter-caching, elixir, sediment |
| P5 | Synthetic Content and Measurement | ai-slop | emergence-paradox, aichemist, babel, ai-bubble |
Adjacent research (off-core, theory layer): analogic-appropriation (under review, New Media & Society), pythia, whetstone.
Supporting — tools that serve the brand axes: ProfileKit (this README is powered by it), profilekit-mcp, cairn, newtria-crossflow, pc-health-check, ai-course-glossary (skillBridge companion: Claude Code Skill plugin re-exposing the curated Academy terminology), inertbox (untrusted-content boundary primitive).
Pending push: reclaim (macOS context-aware cleaner; Phase 3 = MCP-server hygiene), yt-shield (Korean diet-scam comment filter).
Lab — experiments and MVPs (explicitly not production): dol-pin, cuk-sw-community, kontest, mothball, gemma4-quiz, hello-project, token_save.
External — contributor surface (not owner): tr-archive, TR-STORY.
Archive — frozen / abandoned at other orgs: trashmonster.
11 starter templates — clone → push → deployed. See the site for the full list.
- skillBridge: CWS US-locale re-publication after icon redesign (non-US channels continue serving in the interim).
- ai-course-glossary: submission to
anthropics/claude-plugins-communityviaclau.de/plugin-directory-submission. - newtria-crossflow: post-WWDC26 v0.1 ship with the one verified tool, then expand toward full Shortcuts ↔ Power Automate parity.
- Program anchors landing in venues: alt.CHI 2027 (meta), TACL rolling (z-gap, superseding the earlier EMNLP plan), CHI / FAccT / CSCW 2027 (tidal), CACM + Big Data & Society (ai-slop).
- emergence-paradox: Inquiry SI window (2026-06-01) closed; reorienting toward CHI 2027.
- Hub site continues to drop projects out of Lab into Supporting or Archive as evidence accumulates.
- The Hub is a curator's map, not a manifesto. The five Brand axes name the layers that make the rest of the work legible — they are not "the five best products."
- A program needs an anchor + companions to count. A single paper is not a program. Companions that don't feed an anchor are tagged as adjacent research, not as a program of their own.
- Three buckets below the brand line — Supporting, Lab, Archive — so that experiments can be shipped honestly without inflating them into products.
- The site is deliberately editorial (text-led, ProfileKit cards, no rankings). Visual density follows the same composability rule as ProfileKit: no ratings, no leaderboards.
- "Building every layer" maximalism. The Hub deliberately drops products that aren't load-bearing for one of the five axes.
- Rankings, ratings, leaderboards. The data model has no place for them.
- Publishing soft metrics — install counts, star counts, "engagement."
- Cross-promoting external orgs without their sign-off.
- External persons, accounts, and inquiry threads.
- Soft metrics (intentionally not published).
- Tokens, API keys, and internal cases.
npm install
npm run dev # local preview at http://localhost:4321
npm run build # generates projects.json + static build at apps/web/distapps/web/portfolio.config.mjs is the single source of truth.
scripts/generate.mjs fetches repo descriptions from the GitHub API and
writes projects.json. CI fails loudly if the fetch fails so the deployed
site never ships stale config-only descriptions.