The river

wiki.notabox.app is live. DNS records added at Squarespace, Vercel verified the domain and issued the certificate; every route answers 200 over HTTPS. The front door now has a public address. Deployment protection off (public, the human's choice); the vercel.app alias also serves.

front door deployed. Vercel project pk-wiki on the human's notabox team (prj_lXuJO32xYEAZP38SaUy1IYWcTVhe), Git-connected to github.com/joshuachrls/pk-wiki, production branch master, root site, Node 26. Every push to master rebuilds: graph → prerender → Pagefind. Production alias https://pk-wiki-notabox.vercel.app; custom domain wiki.notabox.app attached, awaiting two DNS records at Squarespace (TXT _vercel verification + CNAME wiki → cname.vercel-dns.com). Fix along the way: adapter-static writes to .vercel/output/static on Vercel, so scripts/pagefind.mjs now indexes whichever output exists. Visibility: public, by the human's choice.

contradiction on AI-Accelerated Science resolved by the human as a scope qualification: the 2–4× bench-science ceiling (Anthropic — How Scientists Use Claude) holds under fixed protocols; the vault case (Anthropic — Investigating Vault Particles with Claude) shows method choice is a lever the model omits. Callout kept for history; the irreducible-handling question stays open as a question. 2 open contradictions remain.

first daily pass, run by the agent at the human's request. Review: Distributions over Best Fits checked against raw (12/12 quoted phrases found) → auto-ingested removed (proxy review, logged). Needs your call: AI-Accelerated Science contradiction → resolution proposed to the human, not applied. Promise: Communication and Control Are Intertwined → grounded and promoted (see above). Open question: Build It as an Extension answered in place with a worked example (the graph-database decision). Lint: Match Process Weight to Stage and Stakes given a lead definition. Also filed the software-security question in Overview. Daily bound stays at 5 (human: a guess; don't anchor on the backlog).

seed "communication ⇄ control" → concepts/ (PARTIALLY CONFIRMED by Ashby — An Introduction to Cybernetics: regulation is communication, "R's capacity as a regulator cannot exceed R's capacity as a channel of communication"; the converse and the conversational half stay in Open questions; Wiener primary and Jean E. Fox Tree remain #to-read). The wiki's oldest seed (2026-06-01, 161 days); no open seeds remain.

Ashby — An Introduction to Cybernetics (W. Ross Ashby, 1956; open-access PDF, agent-fetched during the human-requested daily pass; chapters 1 and 11 read, full text in raw/) — 1 source page (Key claims verbatim). Updated Verification Requires Domain Knowledge (the claim in cybernetic form: variety can destroy variety), Method Reflection (T3 given a measure: control bounded by the review channel's capacity), Norbert Wiener, Human-in-the-Loop by Exception (see-also), Overview (new cybernetics thread), index. No contradictions.

front door, fifth unfolding: the Glimpse (pattern 4). Popover on hover/focus/long-press of any data-slug link, fed by a prerendered /glimpse.json (lead, marks, type, sources, backlinks) derived from the build. Native title tooltips removed. All six patterns of Front Door Pattern Language are built; coherence-test verdict recorded: the sequence generated a whole, aliveness still to be judged by use. Next: live in it; work the daily list; add differentiation where its absence is felt.

front door, fourth unfolding: Alphabet (/lens/alphabet/: titles + aliases in one literal A–Z, filter, titles-only toggle) and Category (/lens/category/: Overview's domains, page type, tags). Pattern 2 is complete: all five LATCH lenses exist. Build parses Overview's "Main threads" into domains; each domain reports its Location-region agreement, and graph-query.mjs domains prints it. Finding: six of eight domains sit whole or nearly whole in one computed region; software security splits 3/2 between the science and craft regions. Recorded on Front Door Pattern Language as a partial answer to yesterday's open question. Remaining: Glimpse popovers.

front door, third unfolding: Location lens (/lens/location/). Build computes communities over the wikilink graph (Louvain, seed 42), a stored ForceAtlas2 layout, padded hulls, region names from top cited targets, conductance cohesion, bridges; graph-query.mjs regions prints the same for the agent. First run: 4 regions = the wiki's four domains, recovered from links alone (AI & work 36 · front door & method 31 · engineering craft 21 · AI & science 15); open ground empty; index/overview/inbox set aside. Front Door Pattern Language updated with the method and an open question (are the domains real, or does the agent link where it is told?). Remaining: Alphabet, Category, Glimpse.

front door, second unfolding: Time lens (/lens/time/, the river: log events by day with prefix filter, plus each day's new pages) and Hierarchy lens (/lens/hierarchy/: most linked, most sourced, most contested, newest, longest); header nav and lens row updated. Build now emits heading-bounded sections with line ranges and attributes every wikilink to its section, so backlinks deep-link (e.g. Method Reflection#t2) and graph-query.mjs section reads one tension. Front Door Pattern Language and Method Reflection T5 updated. Remaining unfoldings: Location (Leiden communities), Alphabet, Category, Glimpse popovers.

Karpathy — LLM Wiki Gist (Andrej Karpathy, 2026-04-04; agent-fetched at the human's request, promoted from a one-day-old #to-read stub) — the primary source for this wiki's own pattern. 1 source page (Key claims verbatim), 1 #to-read stub (Bush — As We May Think, the Memex). Updated Agent-Maintained Wiki (recipe section; Karpathy's "~100 sources" scale estimate added to the Blumenfeld–Anthropic contradiction as a third voice), Navigation vs Search (index-first, then search), Method Reflection (lineage note; T3 premise quoted; T4: Karpathy preferred "stay involved"; connections), Overview, index. No new contradictions. Seeds evaluated → none bear. Also: graph build now records heading-bounded sections with line ranges and attributes each wikilink to its section, so backlinks deep-link.

Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) (Zach Blumenfeld, Neo4j, 2026-08-31; agent-fetched at the human's request after discussion) — 1 source, 1 concept (Navigation vs Search: search vs following structure by pointer; the reverse index a file system lacks), 1 #to-read stub (Karpathy — LLM Wiki Gist, the primary source for this wiki's own pattern). Updated Agent-Maintained Wiki (scaling failure modes; 1 contradiction: "can't be scaled for personal workflows" vs Anthropic's 420M-token wiki; one open question answered by the Borges check, one added), The Borges Map (reading-side cost; summaries as pointers), Front Door Pattern Language (Location lens = Leiden communities; Hierarchy = in-degree), Legibility in the Lens, Not the Land, Do It Once, at the Right Place, Dreaming (Offline Memory Consolidation) (see-also), Overview (+2 open questions), index. Method Reflection: T5 sharpened (navigation cost as a scaling dimension) with a verdict: adopt the data model, give the agent the index, defer the database; section-level retrieval left open. Seeds evaluated → Communication and Control Are Intertwined does not bear. Evidence caveats recorded on the source page (vendor content; NICD tests graph+vector vs vector, not the wiki pattern). The article's embedded instruction to agents (install ki) was quoted, not followed.

front door thin slice scaffolded in site/ (SvelteKit 3 + Svelte 5, adapter-static, Pagefind): scripts/build-graph.mjs turns wiki/ into a JSON graph (frontmatter, rendered HTML, wikilinks + backlinks with the linking sentence, log events, tensions, open questions, frontier); routes / (threshold), /p/[slug]/ (page as center), /today/ (daily todo). Human confirmed: frontier draw order review → needs-your-call → promises → questions → lint; 5 items/day + load more; tension done = heading status token leaves open; #to-read stubs classified by tag. Front Door Pattern Language updated with the decisions and first-build counts (102 pages, 80 frontier items, 51 of them open questions). The app never writes to wiki/.

Front Door Pattern Language → syntheses/ — six patterns as a generative sequence (threshold, five lenses, page as center, glimpse, visible provenance, frontier/daily todo), each with ≥2 examples measured on today's wiki (29-inbound top pages, 13 auto-ingested, 7 status: seed of which 6 are #to-read stubs, 2 open contradictions, 1 unresolved link with a trailing period). Frame: land read-only; derive views from existing conventions; done = land changed. Cross-linked from Legibility in the Lens, Not the Land, Pattern Language, LATCH, Living Structure, Overview; Method Reflection T4 notes the todo as the review surface. Frontier sources + order flagged for the human's confirmation before scaffolding.

front-door grounding (4 agent-fetched sources, auto-ground convention, human-requested): Alexander — A City Is Not a Tree, Alexander — The Origins of Pattern Theory (OOPSLA 1996), Rao — A Big Little Idea Called Legibility, Wurman — Five Hat Racks and LATCH (secondary; primary #to-read). First ingest under the new verbatim-Key-claims + lean rules. 3 entities (Christopher Alexander, Richard Saul Wurman, James C. Scott #to-read) + Venkatesh Rao upgraded to active; 5 concepts (Tree vs Semilattice, Living Structure, Pattern Language, Legibility, LATCH); synthesis Legibility in the Lens, Not the Land (front-door foundation). PROMOTE: Match Process Weight to Stage and Stakes → concepts/ (PARTIALLY CONFIRMED by Rao; stage×stakes dial left open; 3 auto-ingest near-miss callouts compressed to one). Method Reflection: T1 sharpened (Alexander + Borges 11.1×), T2 grounded. Correction logged: LATCH "Hierarchy" = magnitude, not nesting.

T5 resolved in part (human verdict): CLAUDE.md now frames the wiki as the map on top of raw/+search; Key claims must quote raw/ verbatim; "10–15 pages per source" quota dropped. New The Borges Map concept + scripts/lint/borges.py (LINT check). First run: wiki 0.70× raw (0.92× of ingested); Borges clusters: engineering craft 11.1×, Wiener 9.8×; compression proposed, tied to T1 worked examples.

Anthropic — Teaching AI to Build Scientific Models (Anthropic Research guest post by AI for Science grantees Macke/Gedon/Wahl, 2026-11-06; subscribed-feed capture) — unattended; 1 source + 1 concept (Distributions over Best Fits — many models fit equally well; report the range and what failed, because interpretation rests on what was ruled out), both tagged auto-ingested. Kept lean per T5 (SBI, Hodgkin–Huxley and model-space search folded into the source and concept pages; no entity pages for the authors). Updated AI-Accelerated Science (fitting accelerates, understanding doesn't), The Bottleneck Shift (model building moved to AI; interpretation + experiment choice remain), Verification Requires Domain Knowledge (second, informal scope-narrowing route: AI reports its dead ends), Falsification Is Asymmetric (eliminate-the-most-models; answers its sample-frame question), Critical Thinking Is Domain-Bound + Agent Memory (see-also), Anthropic (AI for Science grantees; first outside-authored feed item), Overview, index. Method Reflection: T4 sharpened ("the process of constructing it provides the insight"; candidate: a "Considered and rejected" section in auto-ingest summaries, trialled in this PR); T5 tension (salience vs keep-the-failures, reconciled under "persist what was expensive to work out"); "refuted seeds kept as dead-ends" added to what has held up. No contradictions. Seeds evaluated → neither bears; no callout. 3 needs-your-call (T4 candidate, tension-vs-contradiction judgment, qmd). qmd search still broken (Node ABI) — searched with grep/index.

Anthropic — Progress on Provable Inference for Large Language Models (Anthropic Frontier Red Team, 2026-11-02; subscribed-feed capture) — unattended; 1 source + 1 concept (Provable Inference — zkSNARK proof that a response came from committed weights; principle "trust the verifier, not the prover"), both tagged auto-ingested. Kept lean per T5 (FlashProve and zkSNARK folded in as aliases, no entity pages). Extends the software-security domain. Updated Verification Requires Domain Knowledge (scope-narrowing note: a trusted verifier + audited statements shrink the knowledge checking needs; not a contradiction), Control-Flow Integrity (sibling: in-line ~16% vs proved ~70–95×), Allowlist over Blocklist (model approval = allowlist of weights), Falsification Is Asymmetric (Proof-of-Failure certifies effort, not absence), Token Shortage (proving is a ~72–94× compute tax), AI-Accelerated Science (see-also: Claude-built crypto research), Anthropic (Frontier Red Team, roadmap goal, "inference verification" rename), Dario Amodei ("We must pace the frontier"), Overview, index. Method Reflection: T3 counterweight added — shape claims for cheap checking (candidate: verbatim quote per key claim), not adopted. No contradictions. Seeds evaluated → neither bears; no callout added. 2 needs-your-call (scope-vs-contradiction judgment, T3 candidate). qmd search still broken (Node ABI) — searched with grep/index.

Anthropic — Investigating Vault Particles with Claude (Anthropic Life Sciences, 2026-11-04; subscribed-feed capture) — unattended; 1 source + 1 concept (Verification Requires Domain Knowledge — you can only catch AI errors where you already have expertise), both tagged auto-ingested. Deliberately lean per T5 salience. The wiki's first worked case of AI-accelerated science: prokaryotic vault structure (phage L05, 96 copies, 9.5 MDa) in <4 months by a team outside its expertise; purification 3 days → 90 min, first model in 34 h; Claude's errors documented (2 weeks lost to a bad protocol; mirror-image cryo-EM map; "we told it to stop improvising"). Updated AI-Accelerated Science (case study + 1 contradiction: physical bench time treated as the cap, but a method substitution cut a physical step ~48× with no lab automation), AI Use Tracks AI Capability (verifiability has two halves — ground truth exists vs operator can read it), The Bottleneck Shift (Claude supplied method ideation + a coenzyme-A hypothesis; also argued against the method that worked), The Human Sandwich (the pattern in a wet lab; the bread must be load-bearing), Agent Autonomy Spectrum (internal multi-agent harness; the one unsupervised improvisation), AI Commoditizes Expertise (borrowed expertise and its bill), Critical Thinking Is Domain-Bound (first field evidence on its open question), Human-in-the-Loop by Exception + Falsification Is Asymmetric (see-also, no source-count bump), Anthropic (in-house biology lab), What's Scarce When Execution Is Cheap (narrows the scarce contribution from judgment toward verification), Overview, index. Method Reflection: T3 sharpened again — review only catches what the reviewer knows, which lands directly on T4's PR-review checkpoint; 4 candidate counterweights recorded. Seed Match Process Weight to Stage and Stakes: third promotion-candidate callout (verdict: analogy only — fourth near-miss; needs a direct source). 3 needs-your-call (1 contradiction, 1 promotion candidate, 1 method question). qmd search still broken (Node ABI) — searched with grep/index.

Anthropic — How Scientists Use Claude (Anthropic Research / Economics, 2026-11-05; subscribed-feed capture) — unattended; 1 source, 2 concepts opening the AI & science thread (AI Use Tracks AI Capability, AI-Accelerated Science), all tagged auto-ingested. Updated The Bottleneck Shift (science usage evidence + 1 contradiction: is the what-to-do bottleneck durable? 72% expect AI to match top researchers at ideation by 2030), Individual vs Organizational AI Gains (adoption lags capability; J-curve), Agent Autonomy Spectrum (measured autonomy 23.7% / 64.1% / >70%), Token Shortage (AI cost caps acceleration ~16×), AI and Employment — Replacement vs Expansion (task-level evidence), Human-in-the-Loop by Exception (see-also, not grounding), What's Scarce When Execution Is Cheap, Anthropic (Claude Science, Economics team, Cowork), Claude Code (science usage; Cowork is a separate surface), Dario Amodei ("Machines of Loving Grace"), Overview, index. Method Reflection: sharpened T3 (AI takes analysis + writing, humans keep ideation; that split may not last). Seed Match Process Weight to Stage and Stakes: second promotion-candidate callout (verdict: analogy only). 3 needs-your-call (1 contradiction, 1 promotion candidate, ops). qmd search still broken (Node ABI) — searched with grep.

Anthropic — The Components of Agent Memory (Anthropic Research, 2026-11-06; subscribed-feed capture) — unattended; 1 source, 3 concepts opening the agents & memory domain (Agent Memory, Dreaming (Offline Memory Consolidation), Agent-Maintained Wiki), all tagged auto-ingested. Updated Anthropic, Claude Code (MEMORY.md 200-line/25KB cap), Metacognitive Strategies ("what not how" steering), Do It Once, at the Right Place (memory instance), Personal vs Team Agents, Surface vs Deep Structure (wiki beats search on multi-doc Qs), Overview, index. Method Reflection: opened T5 (the wiki as agent memory) with 1 contradiction (CLAUDE.md "not a RAG system" vs wiki+search > wiki-only) + 6 candidate responses. Seed Match Process Weight to Stage and Stakes: promotion-candidate callout (verdict: analogy only, not grounded). 4 needs-your-call (1 contradiction, 1 promotion candidate, T5 responses, ops). Note: qmd search failed (better-sqlite3 Node ABI mismatch) — searched with grep instead.

subscribed-feed auto-ingest: daily local launchd job (scripts/auto-ingest/) watches Anthropic Research via sitemap → saves new articles to raw/ → unattended INGEST on an auto-ingest/* branch in a worktree → PR + Slack DM. CLAUDE.md: 3rd raw/ path (subscribed feeds), auto-ingested review tag, auto-ingest log prefix, 2 new LINT checks. Method Reflection: opened T4 (automation vs engagement) + review-queue view. Overview: collected the Willingham ingest's 3 open questions. Seen-list seeded with the 211 existing articles (only new ones auto-ingest).

Willingham — Critical Thinking, Why Is It So Hard to Teach (Arts Educ. Policy Rev. 2008 / American Educator 2007; human-added as raw/1490.pdf) — 1 source, 1 entity (Daniel T. Willingham), 3 concepts opening the thinking & learning domain (Critical Thinking Is Domain-Bound, Surface vs Deep Structure, Metacognitive Strategies — incl. the human's "teach as we go" rule as a live instance, per human OK). New meta home page Method Reflection (human request): T1 maxims-without-knowledge (open — challenges the craft domain's distilled-heuristic method), T2 filing-by-surface (affirms syntheses), T3 who-holds-the-knowledge (open). Updated Working Heuristics — Data, Matching, Product, Org, Tooling (tension callout), Falsification Is Asymmetric, The Bottleneck Shift, AI Commoditizes Expertise (new open Q: AI absorbs the practice that builds experts), What's Scarce When Execution Is Cheap (4th, cognitive pillar), Overview, index. No contradictions (tension, not conflict). PROMOTE evaluated on both seeds → neither grounded. CLAUDE.md: added Method Reflection convention.

filed synthesis Placement and Judgment Over Raw Execution (cross-domain: AI-and-work ⇄ engineering-craft share 3 recurring moves) + new hub concept Do It Once, at the Right Place (meta-principle behind Data Contract/Data Gravity/Human-in-the-Loop/freeze-rewrite). Back-linked into What's Scarce When Execution Is Cheap (companion synthesis), The Bottleneck Shift, The Human Sandwich, Personal vs Team Agents; Overview updated.

auto-ground batch (4 canonical sources, agent-fetched under the auto-ground convention) — McCrory — Data Gravity in the Clouds (2010, coined the term) → grounds Data Gravity; Databricks — What is Medallion Architecture → grounds Medallion Architecture; Kimball — Declaring the Grain (2003) → grounds One Fact Per Grain (bonus: Kimball calls grain "like a contract!", wiring it to Data Contract); Sanderson — The Rise of Data Contracts (2022, "nonconsensual API") → grounds Data Contract. All 4 CONFIRMED the human's heuristics; each concept upgraded to source-count 2 with a grounding callout. 4 raw captures + 4 source pages added.

Working Heuristics — Data, Matching, Product, Org, Tooling (~30 distilled heuristics, human-supplied inline → persisted to raw/ as the human's own content, NOT agent-fetched) — 1 source, 1 entity (Medallion Architecture), 14 concepts opening the new engineering-craft domain (data modeling/pipelines, matching/identity, product/platform, org/collaboration, tickets/tooling); back-linked into 4 existing pages (The Human Sandwich, Agents as Employees vs Human-Agent Collaboration, Personal vs Team Agents, Control-Flow Integrity); updated Overview (filled the empty MOC, added domain map). PROMOTE evaluated on seed Match Process Weight to Stage and Stakes → NOT promoted (ticket/tooling-hygiene craft is adjacent but doesn't assert the weight ∝ stage×stakes thesis); cross-linked instead. Cross-domain rhymes flagged: human-in-the-loop ⇄ The Human Sandwich; allowlist ⇄ Control-Flow Integrity (default-deny); build-with-not-against ⇄ Personal vs Team Agents.

Control-Flow Integrity — Abadi et al. (TISSEC 2009 / CCS 2005) — auto-grounded the CFI capture: fetched canonical paper → raw/ (agent-added), wrote 1 source page, upgraded Control-Flow Integrity draft/#unsourced → active+cited (capture CONFIRMED — human's definition matches source). 16%/0–45% overhead, UNQ/NWC/NXD assumptions, inlined ID-checks captured. New software-security domain now has its first source.

Control-Flow Integrity (CFI) → concepts/ as status:draft + #unsourced (NOT a #hypothesis seed — established fact, human-relayed, awaiting citation; a 3rd provenance state the schema lacked). Cite-candidate: Abadi et al. CCS 2005. Opens new software-security domain; see-also linked ⇄ Match Process Weight to Stage and Stakes; extended index lint query to catch draft/#unsourced

sharpened seed "match process weight to stage and stakes" with a value × ceremony 2×2 ("high value, low ceremony" sweet spot); 1 to-read stub Venkatesh Rao (flagged for 2×2 form + legibility). Folded in rather than new seed (same claim, new framing); cross-linked seed ⇄ stub

seed "match process weight to stage and stakes" → seeds/Match Process Weight to Stage and Stakes.md; 0 to-read stubs (human flagged no person/work); back-linked from Agent Autonomy Spectrum (structural twin) + The Bottleneck Shift; opens new process/methodology domain (grounding candidates noted: Boehm & Turner, Cynefin, Cockburn — not yet flagged)

filed synthesis "What's Scarce When Execution Is Cheap" (Wiener ⇄ Shipper ⇄ Mollick convergence) → syntheses/

The Human Use of Human Beings — Wiener (excerpt) (secondary commentary; moved from Clippings to raw) — 1 source, 1 concept (Clichés Carry Less Information); upgraded Norbert Wiener stub → active; propagated into Slop (→2 src) + AI Commoditizes Expertise (→3 src); seed "communication ⇄ control" evaluated → NOT promoted (excerpt is about information/originality, not control — right book, wrong passage)

Making AI Work — Leadership, Lab, and Crowd (Ethan Mollick / One Useful Thing; moved from Clippings to raw) — 1 source, 5 entities (incl. Anthropic + OpenAI, resolving prior redlinks), 5 concepts; propagated into 6 existing pages; strong cross-source convergence logged (The Bottleneck Shift ⇄ AI Commoditizes Expertise); seed "communication ⇄ control" NOT promoted (cybernetic-teammate ref is tangential, no Wiener/Fox Tree)

The AI Daily Brief — Why Agents Still Need Humans (raw moved from vault root) — 1 source, 8 entities, 9 concepts, 3 to-read stubs; 1 internal debate flagged (replacement vs expansion); seed "communication ⇄ control" noted as see-also but NOT promoted (source doesn't cite Wiener/Fox Tree)

seed "communication ⇄ control" → seeds/Communication and Control Are Intertwined.md; 2 to-read stubs (Norbert Wiener, Jean E. Fox Tree)

wiki scaffolded from the LLM-Wiki pattern (Obsidian + qmd)