nx_frontier_scan_review_candidate_t346.nx

nx_frontier_scan.nx -- THE AUTONOMOUS FRONTIER-NOTICER for a /compare domain. Operator 2026-08-20: "we want to be able to notice things like the metahuman gap without needing claude by having our own research on /compare sota and best of breed and workflowed and mcp and api etc". WHAT THE METAHUMAN GAP ACTUALLY WAS (mined from the transcripts, not assumed -- writer session 55724797-9f7b-4f68-81aa-fd2de7997175, 2026-08-18 18:15:25Z): the operator said the characters looked wrong, a seat ran FIVE WebSearch calls, and appended four citations to charsim.refs whose URLs IT NEVER OPENED (three carry mirror '-' pin '-'; the fourth points at a page fetched three days earlier from a DIFFERENT url). The gap therefore had TWO halves and only ONE of them needed the outside: (A) the INTERNAL deficit signal -- rivals coded Best on skin/hair rows where our symbol is _ABSENT_. Already mechanical, already in our own data, and NOBODY HAD EVER COMPUTED IT. (B) the OUTSIDE BAR -- the 2026 numbers. That genuinely needed the live web. This organ computes (A) from the domain's own files and pulls (B) through the sovereign fetcher. IT EMITS PROPOSALS AND NEVER ADMITS ONE. It does not write <domain>.matrix, <domain>.refs, or any watch cell. The measured law it obeys: no margin threshold makes auto-declaration safe (nx_memsugg, precision 808->890 permil at 24 percent coverage), and A COMPLETION SIGNAL THAT KEYS ON A NAME REWARDS WRITING THE NAME. The list is the answer; a seat adjudicates it. SEEDS ARE DATA, NEVER HAND-TYPED. They come from the domain's own board: <domain>.axes field 4 is literally "frontier-kw;frontier-kw" -- the estate ALREADY declares per-row frontier keywords and nothing had ever consumed them. <domain>.matrix field 1 (the row label) -- always, so a domain without .axes still has seeds. SOURCES ARE DATA TOO: knowledge/frontier_sources.conf. Both endpoints shipped with it were PROVEN reachable through nx_research_fetch before a line of this was written (arXiv API 200/12267 B, HN Algolia 200/17063 B) -- necessary because nx_websearch is OUR OWN CRAWL INDEX (0 results for "metahuman"), so it structurally cannot see the outside. VERBS seeds <domain> -- what this domain is about, from its own data. No network. deficit <domain> -- the internal signal, ranked, WITH THE REASON PER ROW. No network. scan <domain> [maxrows] [maxitems] -- fetch + diff + print proposals. Network. Writes nothing. propose <domain> [maxrows] [maxitems] -- scan, then file the worklist + the frontierprop- plane rows. RESOURCE ENVELOPE (a shipping criterion, not an afterthought): outbound requests are exactly rows_scanned x sources, both bounded by conf and echoed in every run line; buffers are sized FROM THE INPUT (sys_read_file, matrix+refs bytes, source count) so there is no ceiling to guess; the only writes are one worklist file and at most rows x sources x 2 plane upserts, keyed by url hash so a re-scan UPSERTS rather than grows.