nx_memguard.nx

⛔⛔ DISPROVEN 2026-08-06 -- BUILT, MEASURED, AND RETRACTED. NOT DEPLOYED, NOT HOOK-WIRED. The ELF was deliberately removed from _offc. Source kept as a recorded negative result, because the idea is attractive enough that someone will otherwise build it again. WHAT IT WAS MEANT TO DO: derive a query from a new memory's own name+description and warn if the corpus already holds that doctrine -- prevention at write time, where post-hoc detection has a proven ceiling. WHY IT DOES NOT WORK (clean replay + control, both failed): * replay: fed the ORIGINAL July store-law frontmatter, it did NOT surface the June file that says the same thing. It returned project-nofloat-models-vision and netobs-observability instead. * control: a birdsong-synthesis note fired 4/6 against a FETCHER memory. A guard that fires on everything is indistinguishable from no guard, except it also teaches the reader to ignore it. ⚠MY FIRST "VALIDATION" WAS CONTAMINATED and nearly shipped: I pointed it at the July file AFTER stubbing it, and the stub's description literally contains the canonical filename. It scored 17/20 and looked like a triumph. ★ A FIXTURE THAT NAMES ITS OWN ANSWER VALIDATES NOTHING. ★★★★★★ THE ROOT CAUSE, MEASURED, AND IT INDICTS MORE THAN THIS ORGAN: IDF ASSUMES A DIVERSE CORPUS. IN A SINGLE-PROJECT CORPUS THE TOPIC WORDS *ARE* THE COMMON WORDS, SO RARITY WEIGHTING SELECTS FOR THE INCIDENTAL. Document frequency over the 2,178-file corpus (common cutoff = 108): sovereign 1178 · data 1108 · files 803 · store 638 · tsv 438 · txt 388 · ssot 257 · seg 228 Every word that IDENTIFIES the store law is corpus-common and gets dropped; the survivors were "planes" (105) and "queues" (39) -- incidental words that then matched unrelated files. Lowering the word-length floor 5 -> 3 to admit "tsv"/"txt" changed NOTHING, because those are common too. ⇒ Rarity weighting works for distinguishing specific technical artefacts (coulombic, dendrite) and fails for a project's CENTRAL concepts -- which is exactly where duplicated doctrine lives. The same blind spot limits nx_memhealth's duplicate ladder. nx_memfind is unaffected ONLY because a human picks its query terms; that human judgement is the part that cannot currently be automated. ---- original design notes below ---- nx_memguard.nx -- at WRITE time, ask "does this memory already exist?" and name the candidates. WHY THIS AND NOT MORE DETECTION ------------------------------ Measured 2026-08-06, the same doctrine written four times in 90 seconds; the store law written twice by two sessions three weeks apart; nge-browser-first written twice by ONE session. Every one