heznpc

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.

Currently implemented

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.

Planned

• 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-community via clau.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.

Design intent

• 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.

Non-goals

• "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.

Redacted

• External persons, accounts, and inquiry threads.
• Soft metrics (intentionally not published).
• Tokens, API keys, and internal cases.

---

Running the site

npm install
npm run dev       # local preview at http://localhost:4321
npm run build     # generates projects.json + static build at apps/web/dist

apps/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.