nishi code wiki / learnings

learnings

every lesson this ecosystem has banked, ranked and mined from its own record - not asserted.

43977

learnings banked

2879

rank 6

7566

rank 5

13085

rank 4

20447

rank 3

1936

source files

What this is. The ecosystem states lessons constantly - laws, findings, callouts, root causes - and until 2026-07-31 none of them were captured mechanically. The harvest that ran every six hours extracted only "type":"tool_use" frames, i.e. the names of tools called, so a pipeline reporting fail=0 forever was moving zero learnings. This page and the SessionStart pulse are both emitted from the one store that now mines them.

Ranking follows the operator's own convention: a run of N stars is rank N, a callout glyph is rank+2, and marker phrases (LAW:, ROOT CAUSE, REFUTED) carry configured weights. Rank is the MAX of those signals, never the mean. Every weight lives in learn_markers.conf, never in code.

Two sources. curated (31880) are hand-distilled memory files; transcript (12097) are lessons mined forward from live sessions that were stated but never banked - that second number is the capture-leak gauge.

rank 6 2879 banked

6One correction I owe you: the banked law says *"no intake handler exists estate-wide, every generated contact form is decorative."* That is stale — nx_site_intake.nx + nx_site_intake_serve.nx were built since it was written. But it's only half wrong: nx_site_intake_serve.elf (:8033) is not in the live service list and there's no /intake/ edge route. So a dead form still ships today, just for a01f8dcee · transcript

6B We show the omission is not cosmetic. The two weight-side floors bind in different regimes and cross at {E03} per-mille of total solids. Above that crossing the one-floor law is correct by luck. Below it the law over-permits, licensing {E04} per-mille of air beyond what the standard allows: {E06} per-mille where the true ceiling is {E05}. The error direction is the unsafe one, because it9d0062a9 · transcript

6The organ's own header states the law it now violates:e494a0ed · transcript

6// LAW: an instrument that reports OK during the worst outage class it could ever see is not an instrument.e494a0ed · transcript

6The kill landed (the expired session was a dropped response, exactly as the law says) — and the numbers moved:e494a0ed · transcript

6// are named, single-sourced, and read by every consumer of the law.9d0062a9 · transcript

6🔴08-05 TLS REACH + STATIC-KEY SEV-9 — SEV-9 1785970852 (38 sites hardcoded the X25519 priv key) ✅ALL 22 PROD SITES FIXED + compile-verified 08-06; 🔴scope corrected OUTBOUND-CLIENT-ONLY — served TLS never lost forward secrecy ⇒ ★SCOPE A VULN BY MEASUREMENT BEFORE LETTING ITS BLAST RADIUS STAND. ⚠⚠I CALLED A CRASH LOOP LIVE THAT DIED 07-19 (never-rotated log) ⇒ ★★★★★★A CUMULATIVE LOG MINED WITHreference-index-recent-refs-dormant · curated

6🟢★★★★★★08-06 SEV-9 PAIR DISPROVEN — nx_cc IS INNOCENT — 1785530065+1785529609 blamed a ==-compiles-as-> codegen bug; THE SOURCE SAID != — an ESCAPED nx_gate_bite mutant (its phase-1 operator IS ==→!=). One byte, 8/10 RED→10/10 GREEN, both directions. ✅producer fixed+promoted: write-ahead .bitepristine sidecar, recover-BEFORE-trust (else the corruption self-certifies), disarm only afterMEMORY · curated

6🏁08-05/06 NISHILANG SOTA RADAR + COMPILER SAFETY — /compare/lang radar LIVE; 7 compiler capabilities bite-witnessed; CLIMB v10 = 1785971135. ✅08-06 5W+H VOICE LIVE on undefined-name (rubric ~24/25 from ~14) ⇒ ★★★★★★SEARCH MEMORY FOR THE VOICE STANDARD THE ESTATE ALREADY BOUGHT BEFORE DESIGNING A SURFACE · A RICHER MESSAGE IS A LARGER PAYLOAD — RE-AUDIT EVERY BOUNDED WINDOW DOWNSTREAM (hostctlMEMORY · curated

6🟢🔴08-05 ICE CREAM SOTA RULER — ruler 17/17, paper FD-001 LIVE, nx_ice_scale 13/13 bite-proven. ★★★★★★WRITING A LIMIT DOWN IS NOT APPLYING IT (⇒ grep a const's own name; count=1 = DEAD) · A ROUNDED-DOWN DIVISOR RELAXES A LEGAL FLOOR · A DIFFERENCE IN A TABLE IS NOT ONE THE STUDY FOUND · SCALING UP IMPROVES CRYSTAL SIZE — THE ARTISAN PREMIUM IS BACKWARDS · LABOUR IS THE COST DRIVER (100×,MEMORY · curated

6🔴08-05 nx_txtscan BYTE-ORIENTED + 0x0E — 211 laptop .nx carry 0x0E runs, NAS CLEAN ⇒ PULL from NAS, NEVER push · ★★★★★★THE MCP TOOL NAME IS NOT THE BUILD TARGET NAME (nor is a SOURCE FILE a target: a lib with no main() fails rc=102).reference-index-recent-refs-dormant · curated

6🟢🔴★★★★★★08-06 CRAWLER: GZIP + SHAPE POLICY + A SECOND SYSTEM — ✅1.3 leg was INDEXING GZIP (6,079B 1f 8b vs 1.2's 30,385B plaintext, same url same second) ⇒ inflate at the chokepoint, all 3 legs now identical · ✅HOST ALLOWLIST RETIRED (operator: "a dumb way to have a search engine") ⇒ shape/param policy from urlpolicy- plane, gate 21→23/23; uncurated query urls now crawl (v2ph + loc.govMEMORY · curated

6★★★★★★A CRITICAL SECTION MADE SAFE AGAINST YOUR OWN CONTROL FLOW IS NOT SAFE AGAINST PROCESS DEATH. Fail-safe *by finishing* is not fail-safe *by construction*. Any tool that edits source to test it is a writer, and every writer needs crash recovery.project-nishi-escaped-mutant-compiler-exoneration-2026-08-06 · curated

6★★★★★★A DIFFERENTIAL PROBE HAND-COPIED FROM ITS SUBJECT IS ONLY EVIDENCE IF THE COPY IS PROVEN CHARACTER-IDENTICAL. rdy_probe — same module, same imports, same consts, same long literals — silently corrected the defect while being copied, so four rounds of it exonerated the innocent (import count, const collision, module-vs-main, literal length) and never indicted the guilty character. It is keptproject-nishi-escaped-mutant-compiler-exoneration-2026-08-06 · curated

6Banking the law so no future session re-chases this. Writing the topic file first.e494a0ed · transcript

6// ★★★★★★OPERATOR REPERTOIRE IS A COVERAGE AXIS IN ITS OWN RIGHT. Until 2026-08-04 this matched ONLY ==e494a0ed · transcript

6285 1784577350 7 osbench open OSBENCH ALTITUDE DEFECT found by running my own F853 harness and reading the output (2026-07-20). nx_osbench_micro emits create_write/read/resolve rows tagged MEASURED, but the two sides are NOT the same altitude: PABI_NISHIOS block-VFS is an in-memory mmap 256KB image (zero syscalls, no page cache, no fsync, no disk) while PABI_LINUX does real openat+write+closee494a0ed · transcript

63. ★★★★★★SUCCESS WAS WEATHER, NOT CAPABILITY — AND NOTHING RECORDED ITproject-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

62. ★★★★★★OPERATOR: THE HOST ALLOWLIST WAS "A DUMB WAY TO HAVE A SEARCH ENGINE"project-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

6is pacing, not anti-bot capability. ★★★★★★A 403 AFTER SUCCESSFUL 200s FROM THE SAME HOST IS A RATEproject-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

61. ★★★★★★THE 1.3 LEG WAS INDEXING COMPRESSED GARBAGEproject-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

64. ★★★★★★THE SECOND SYSTEM — AND IT REFUTED ME ON ITS FIRST REAL RUNproject-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

6★★★★★★A DECODER ONE CONSUMER IMPORTS IS NOT A CAPABILITY THE FLEET HAS.project-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

6★★★★★★REMOVING A BLANKET RULE BILLS YOU FOR EVERY CASE IT WAS SILENTLY COVERING.project-nishi-crawler-outcome-log-and-auditor-2026-08-06 · curated

6Shipping the voice also caught a real stack bug, which is the session's pattern paying off again: the richer message couldn't reach the API at all — hostctl echoes only the last 600 bytes of the builder's capture, smaller than one 5W+H diagnostic, so the head of the first error was being amputated upstream of everything I'd fixed. Law: a richer message is a larger payload — re-audit every bounded526e5f4a · transcript

6🔴★★★★★★THE VOICE EXPOSED A STACK TRUNCATION: hostctl hc_tail("/tmp/buildrun.out", 600) echoes only 600B of the builder's FULL capture — ONE 5W+H diagnostic is bigger, so the head could NEVER reach the API no matter what mgmtapi did. A RICHER MESSAGE IS A LARGER PAYLOAD — RE-AUDIT EVERY BOUNDED WINDOW DOWNSTREAM OF A VERBOSITY CHANGE. Fix: mgmtapi ma_read_tail reads /tmp/buildrun.out (64KB tail)project-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★THE OPERATOR WAS POINTING AT A STANDING CARDINAL ALL ALONG: nxc2-5wh-substrate-2026-04-29 ("why who what where when how... one line answers any issue") + feedback-elm-class-friendliness-cardinal (rubric WHAT/WHERE/WHY/HOW/no-jargon ≥20/25 BEFORE shipping; internal organ names forbidden; five-minute-new-user test = ship blocker). I shipped THREE diagnostic versions without checking theproject-nishilang-sota-radar-2026-08-05 · curated

6🟢🔴★★★★★★08-05 CWE-330 ENFORCER + FETCH12 BODY — ✅enforcer LIVE (nx_cwe_scan 155,033B, gate 9/9, ratchet plane cweratchet-); ✅fetch12 chunked-body fixed, graphis 200 + 43,976B live. ★★★★★★OPERATOR: DATA MUST NOT BE WELDED TO LOGIC — the plane joined counters BY ROW ORDER, so reordering data silently mis-attributed a security count ⇒ re-keyed by CWE id; ★PROVE A DECOUPLING BY MUTATION (fixturesMEMORY · curated

6🟢🔴★★★★★★08-05 ICE CREAM SOTA RULER vs BEN & JERRY'S — ruler 17/17 + hydrocolloid term 14/14, paper FD-001 LIVE /papers. ★★★★★★WRITING A LIMIT DOWN IS NOT APPLYING IT: IC_FDA_MIN_FOOD_SOLIDS_Q2 declared + grep=1 occurrence (its own decl) ⇒ 1 of 4 CFR floors unenforced for the organ's life ⇒ TEST: grep a const's own name and COUNT; 1 = DEAD. Two floors cross at 356‰ TS; below it the shipped lawMEMORY · curated

6★★★★★★A 200 ALONGSIDE AN EMPTY BODY IS A PARSE BUG, NOT A TRANSPORT ONE — and the tooth must assertproject-nishi-cwe330-enforcer-and-fetch12-body-2026-08-05 · curated

63. ★★★★★★THE SCANNER CAUGHT A STATIC KEY MY GREP MISSEDproject-nishi-cwe330-enforcer-and-fetch12-body-2026-08-05 · curated

61. ★★★★★★OPERATOR CORRECTION MID-TURN: DATA MUST NOT BE WELDED TO LOGICproject-nishi-cwe330-enforcer-and-fetch12-body-2026-08-05 · curated

6🔴★★★★★★GATE SIGPIPE IMMUNITY SHIPPED: BOTH verdictless-run pairs (1785972850/65, 1785975277/91) died because the gate streams to stdout — caller's transport drops ⇒ next stdout write ⇒ SIGPIPE kills the run MID-MEASUREMENT, so even the LOG never gets the verdict. sys_ignore_sigpipe() at main entry ⇒ A GATE WHOSE VERDICT DEPENDS ON ITS CALLER'S CONNECTION SURVIVING IS NOT A MEASUREMENT. (Withproject-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★METADATA ABOUT A DECLARATION IS STILL A LOOKUP; QUOTE THE DECLARATION. 1 arg(s), defined at line N makes the reader decode shorthand and chase a line. v2 prints the ACTUAL SOURCE LINE of the suggestion (new nx_diag_show_line = the caret's line-extractor minus the caret, 6-space indent):project-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★A ROUNDED-DOWN DIVISOR IS A RELAXED CONSTRAINT WHEN THE DIVISOR IS A LEGAL FLOOR.project-nishi-icecream-sota-ruler-2026-08-05 · curated

6★★★★★★SCALE RUNG SHIPPED — nx_ice_scale + gate 13/13, BITE-PROVENproject-nishi-icecream-sota-ruler-2026-08-05 · curated

6★★★★★★THE HEADLINE: WRITING A LIMIT DOWN IS NOT APPLYING ITproject-nishi-icecream-sota-ruler-2026-08-05 · curated

6★★★★★★A DIFFERENCE VISIBLE IN A TABLE IS NOT A DIFFERENCE THE STUDY FOUNDproject-nishi-icecream-sota-ruler-2026-08-05 · curated

6★★★★★★SCALING UP *IMPROVES* CRYSTAL SIZE — THE ARTISAN PREMIUM IS BACKWARDS HEREproject-nishi-icecream-sota-ruler-2026-08-05 · curated

6★★★★★★LABOUR IS THE COST DRIVER — NOT INGREDIENTS, NOT ENERGYproject-nishi-icecream-sota-ruler-2026-08-05 · curated

6🔴★★★★★★THE SAME CHECK THAT CONFIRMED THE WRITE FALSIFIED A PUBLISHED CLAIM. Moving to GPT madeproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6🔴★★★★★★THE STICK WRITE EXPOSED A DEFECT THE IMAGE COULD NEVER SHOW: AN IMAGE IS A FIXED-SIZEproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6(NishiLang → NATIVE Windows PE). ★★★★★★"WORKS ON EVERY OS" MEANS ONE ORGAN EMITTED PER HOST, NOTproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6run failed with a bare -39. The law is banked and I still skipped the assert. Every scriptedproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6"ON" into *which one, and from where*. ★★★★★★A STATUS LIGHT SAYS SOMETHING HAPPENED; A NAME SAYSproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6🟢🔴★★★★★★08-05 SEARCH: 8 DEFECTS, ALL LIVE — root cause COVERAGE, NOT RANKING; R4-R8 shipped, ruler 366 FLAT · ★READ THE BYTES BACK · 🔴R9 TLS 0x1302 undecryptable ⇒ ★RETIRE ON EVIDENCE THE HOST IS DEAD, NEVER ON YOUR OWN INABILITY; hash_len48 remains 1785966831. Open 1785959927/943 · 1785965632. Detail→file+OVERFLOW.MEMORY · curated

6🏁08-05 NISHILANG SOTA RADAR + COMPILER SAFETY — /compare/lang radar LIVE; 7 compiler capabilities shipped bite-witnessed; v9 rungs 1-2 done, CLIMB v10 = 1785971135 · ★THE MESSAGE IS PART OF THE FEATURE · ★★★★★★A RADAR THAT CANNOT MINE IS A REPORT, NOT AN INSTRUMENT (nx_capwitness = next). ⚠equiv gate: pass ../<staged>; RED-all+cclogb=0 = challenger NEVER RAN (now LOUD preflight). ✅late2:MEMORY · curated

6★★★★★★08-05 MOVE WORK TO THE SIDE WITHOUT BUDGET PRESSURE — a feature losing a competition for a FIXED budget needs to MOVE, not be tuned (index-side pyramid ⇒ query-side sweep: 17/17 classes). PUT THE IDENTITY CASE IN THE SWEEP + TAKE THE MIN ⇒ NON-REGRESSION IS A PROOF, NOT A TEST · ★★★★★WIDENING A SEARCH REQUIRES RAISING THE BAR — BEST-OF-N IS N CHANCES TO BE FOOLED (false identity match atMEMORY · curated

6🔴🔴★★★★★★08-05 TLS REACH + STATIC-KEY SEV-9 — SEV-9 1785970852: 38 sites hardcoded the X25519 priv key = NO FORWARD SECRECY ⇒ ✅CRAWLER SHIPPED LIVE sha a8b3a686 1,459,688B (grew, gate 21/21, run-proven) + cert_autorenew + gallery_auth · 🔴SCOPE CORRECTED: OUTBOUND-CLIENT-ONLY — all serving daemons csprng per-connection (nx_sites_daemon_v2:1370) ⇒ SERVED TLS NEVER LOST FORWARD SECRECY ★SCOPE AMEMORY · curated

6🟢🔴08-05 WRITE BRAIN SOTA SWAP — llama-server:8091 brain + embedded 4B fast lane (Gemma-4-26B-A4B by measurement) · ★★★★★★STANDING: every nx_cc change runs nx_cc_equiv_gate BEFORE promotion (caught its 1st change; /tmp IS noexec) · ⚠⚠nxmcp.ps1 in-process only · ⚠open 1785935429; registry COMPILED-IN. Detail→file+OVERFLOW.MEMORY · curated

6★★★★★★08-05 N SEGMENTS, NOT ONE INDEX (⚠file CORRECTED same-session after operator pushback — I answered a global-scalar-quality strawman when he described the SUBREDDIT model; CONCEDE ONCE AND MOVE, repeating a caveat IS the moralising). Segments = exemplar seeds + own depth/cadence/quality/demand ⇒ query scans 1/N = the delivery-cost lever; scarce segments sit at T0 8B forever. DUPLICATE COUNTMEMORY · curated

6🟢🔴★★★★★★08-05 CODEC COVERAGE→IMGSEARCH SOTA — 5 rungs GREEN incl VP8 keyframe; corpus 1000‰ (120/120) + TGA/PCX/ICO/PNM 12/12; cross-format content-match proven · ★A RED FROM AN IMPOSSIBLE EXPECTATION IS A DEFECT IN THE RULER · ★DECODABLE ≠ SEARCHABLE. Debts 1785955112/1785955132. Detail→file+OVERFLOW.MEMORY · curated

6🔴★★★★★★08-05 nx_txtscan BYTE-ORIENTED + 0x0E CORRUPTION — line-oriented grep inherits the wrong UNIT ⇒ records are OFFSETS; 211 laptop .nx carry 0x0E runs, NAS CLEAN ⇒ PULL from NAS, NEVER push · ★★★★★★THE MCP TOOL NAME IS NOT THE BUILD TARGET NAME — check tool_allowlist.conf before believing a promote · slopmeter: real gap = VOCABULARY. Detail→file+OVERFLOW.MEMORY · curated

6Your fairness principle is now estate law: the comp inherits the donor's own construction rules — and the measured fact that licenses it was sitting in our own data all along. Seed-san is *not* one sculpted surface: she's 5 meshes, 21 primitives, 74 named parts — hair as separate strand tubes, a head that alone carries 16,778 triangles (half her body's budget!), clothing draped over a core. Our05646302 · transcript

6★★★★★★A MULTI-MESH DONOR HAS ONE SKIN PER MESH — taking skins[0] labelled all 34k triangles withproject-nishi-vrm-donor-matrix-2026-08-04 · curated

6★★★★★★VERIFY A DIAGNOSTIC BY READING ITS EMITTED BYTES, NEVER THE SOURCE THAT PRINTS IT — second catch of the day by this rule (the other: an off-by-one sys_write length silently truncating at line 1118).project-nishilang-sota-radar-2026-08-05 · curated

6Operator caught a real regression against our own bar: my first exhaustiveness message carried *no line and no enum name*, while the duplicate-definition diagnostic shipped hours earlier cites both lines. ★★★★★★THE MESSAGE IS PART OF THE FEATURE — A CHECK THAT CANNOT SAY WHERE IS A WORSE CHECK, AND YOUR OWN BEST DIAGNOSTIC IS THE BAR, NOT THE COMPETITOR'S. Fixed before promotion:project-nishilang-sota-radar-2026-08-05 · curated

6🔴★★★★★★THE GATE'S OWN TEMPS WERE THE seq236 CLASS: every temp was ONE FIXED PATH (/tmp/eqa.s, worst _build/_eq_asm_out.elf = the RENAME SOURCE — run A could rename run B's assembled binary into its own slot and compare CROSS-WIRED artifacts). OBSERVED LIVE: a dropped-response MCP call + its transport retry = TWO CONCURRENT RUNS 15s apart (epochs 1785972850/65, both died verdictless when theirproject-nishilang-sota-radar-2026-08-05 · curated

6⚠FIRST ATTEMPT POINTED AT THE WRONG LINE: I anchored on tok_at(P.toks, P.pos) at *error* time, but the parser has advanced past the call by then, so the caret landed under the NEXT statement. Fixed by capturing the call's own token at parse_primary_call entry (the caller consumed the identifier, so it's one slot back). ★★★★★★A CARET THAT POINTS AT THE WRONG LINE IS WORSE THAN NO CARET — it sendsproject-nishilang-sota-radar-2026-08-05 · curated

6Mining: still open. matrix-class domains measure exactly one thing — a Nishi cell is "present" only if its declared implementing symbol exists on disk — so they catch a *deleted* organ but can never *discover* a new capability. sota-class domains (lang, llm, nishios, …) measure nothing: every cell is data and the liar-kill checks only *shape* (cols≥8, rows≥20, quantitative≥3, No≥Best), neverproject-nishilang-sota-radar-2026-08-05 · curated

6// per the law this session just gave a tooth). Detection is structural and must be code; the STANDARD's taxonomy8eaefd4f · transcript

6503 = dropped response on a write. Law: read back before retrying. Checking whether the edit landed.526e5f4a · transcript

6★★★★★★A TEST CONSTANT THAT ESCAPES INTO PRODUCTION IS INDISTINGUISHABLE FROM A BACKDOOR.project-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6★★★★★★A CUMULATIVE LOG MINED WITH NO TIME WINDOW REPORTS A FIXED PROBLEM AS A LIVE ONE — A COUNTproject-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6★★★★★★THE ADOPTION GAP, NOT A MISSING PRIMITIVE (4th time this lane) ·project-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6★★★★★★A RULER THAT REPORTS A CAUSE IT NEVER MEASURED WILL BE BELIEVED — print the byte, never the theory.project-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6★★★★★★A 200 ALONGSIDE A FAILURE VERDICT MEANS THE TRANSPORT SUCCEEDED AND THE PARSE DID NOT — theproject-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6★★★★★★A SIZE COMPARISON IS ONLY EVIDENCE BETWEEN TWO BUILDS OF THE SAME SOURCE SET — comparingproject-nishi-search-tls-reach-and-static-key-2026-08-05 · curated

6nx_meshview packed pixels R-high while the shared write_png consumes R-low: every render this lane ever produced had red and blue swapped — the bodies you've been seeing all night were flesh-toned in the data and blue on screen, and nobody could tell, because until the heatmap no render's colors *meant* anything. The law is worth the price of the bug: a color bug is invisible until a color05646302 · transcript

6The fix row is queued to the plane behind a background retry (the edge is in one of its self-healing flaps; the banked pattern says it recovers on its own, and the poller will land the row the moment it does). Memory carries the law either way.05646302 · transcript

6The lawyer example is the cleanest proof that this is a general mechanism rather than anything to do with faces: practice-area specialization is a facet, and the same segment machinery serves personal-injury sites, corgis, and a visual style with identical plumbing.1aa138c6 · transcript

6Mid-session the tools daemon 503'd and the edge dropped for ~2 minutes after I fanned out 8 parallel ingest jobs — it self-recovered, I finished over the raw curl /mcp lane, and banked the law: *the job daemon IS the control plane; fan out ≤4*.53b11736 · transcript

6🔴🔴★★★★★★08-05 TLS REACH + STATIC-KEY SEV-9 — SEV-9 1785970852: 38 SITES HARDCODE THE X25519 PRIVATE KEY (priv[i]=0xA0+i) ⇒ THE FLEET HAD NO FORWARD SECRECY (one ECDH key for every session ever, derivable from any binary; hits the crawler's shared fetch + ACME cert autorenew + an *authenticated* fetch). nx_csprng_fill was ALREADY IMPORTED IN THE SAME FILE ⇒ ★★★★★★THE ADOPTION GAP, NOT A MISSINGreference-index-recent-refs-dormant · curated

6★★★★★★graphis = TLS-1.2-ONLY Apache (suite 0xc02f), NEVER an 0x1302 host — found only after I made the diagnostic PRINT the suite instead of asserting it ⇒ A RULER THAT REPORTS A CAUSE IT NEVER MEASURED WILL BE BELIEVED. ✅TLS-1.2 fallback SHIPPED into the shared fetch lib (authenticated impl, graphis 200 + 1.3 control unchanged). 🔴Crawler's remaining blocker = BODY EXTRACTION, not TLSreference-index-recent-refs-dormant · curated

6🟢★★★★★★08-05 NISHI-OS N1-FULL — the shim READS /KERNEL.NXE off the ESP (embed deleted; BOOTX64 22016→1536B; census=EXACT 6 callrax+1 refusal+1 jmp; kernel growth no longer touches the shim). Real-TianoCore FAT witness (tally RISES across EBS 640→2080px; PCI 2×VERIFIED+2×UNBOUND) + literal NXE! neg on real serial. ✅08-05e FS-2 COMPLETE — image = GPT 2 partitions (ESP + NISHI DATA, Microsoftreference-index-recent-refs-dormant · curated

6★★★★★★THE R9 "BUILD TRUNCATION" SEV-8 WAS RETRACTED (1785968635+1785968500 EATEN): it was THE MCP TOOL NAME IS NOT THE BUILD TARGET NAME again (nx_https_get→cli2.elf, but target nx_https_get compiles the LIBRARY wrapper) ⇒ ★★★★★★A SIZE COMPARISON IS ONLY EVIDENCE BETWEEN TWO BUILDS OF THE SAME SOURCE SET.reference-index-recent-refs-dormant · curated

6🏁08-05 NISHILANG SOTA RADAR + COMPILER SAFETY — /compare/lang = sota-class 11-lang/35-axis radar LIVE (regen-owned; demand = 1651 oracle bugs) + hub↔spoke sync, frontier door, buildknow lane, arity cron. 7 COMPILER CAPABILITIES SHIPPED IN ONE SESSION, every one bite-witnessed refuse+accept and gated (equiv 10/10+selfhost → promote_toolchain canary): call arity · both-direction ptr/int arg typesreference-index-recent-refs-dormant · curated

6code-metrology code research_metrology Ruler-craft: descriptive geometry to modern metrology, applied to precision realism 2026-08-03 PUBLISHED code-digitaltwin the freemason ruler answered: Monge descriptive geometry (2 orthogonal views determine a solid) exposes that our 1-view instrument CANNOT SEE depth -- nose/chin projection was inference not measurement; ladder M1 orthogonal pair, M253b11736 · transcript

6ROW: 1785968500 7 tls/0x1302-integration open R9 0x1302 FULLY WIRED IN SOURCE + COMPONENT-GATED, BUT THE INTEGRATED CLIENT BINARY IS DEAD -- STAGED ONLY, NOTHING DEPLOYED (2026-08-05). STATE: source complete across nx_tls13_transcript.nx (dual-hash), nx_tls13_record.nx (aes256 dispatch), client_session recv_sh/recv_hs/derive_app/emit_finished (hash_len 48 + suite-selected transcript),962ef9d8 · transcript

6I'll run this as a full Nishi lane: boot a seat, pull relevant banked laws, research the July-2026 SOTA for automotive diagnosis (using the 2012 Sonata and 2013 Sienna as the two hard test vehicles), then publish the findings and the exceed-plan to the estate. Starting with the laws query and loading the tool schemas I'll need.53b11736 · transcript

6ROW: 1785968635 8 build/silent-truncation open R9 INTEGRATION FAILURE ROOT-NARROWED -- IT IS A BUILD DEFECT, NOT TLS LOGIC (2026-08-05, refines 1785968500). TWO MEASUREMENTS SETTLE IT. (1) The staged binary prints ZERO BYTES even when run with NO ARGUMENTS, so it never reaches main's first write -- the crypto path never executes and none of the 0x1302 wiring was ever under test. (2) THE DECIDING962ef9d8 · transcript

6ROW idx=3194 1785968635 8 build/silent-truncation open R9 INTEGRATION FAILURE ROOT-NARROWED -- IT IS A BUILD DEFECT, NOT TLS LOGIC (2026-08-05, refines 1785968500). TWO MEASUREMENTS SETTLE IT. (1) The staged binary prints ZERO BYTES even when run with NO ARGUMENTS, so it never reaches main's first write -- the crypto path never executes and none of the 0x1302 wiring was ever under test. (2) THE962ef9d8 · transcript

6This estate has a banked law: *a byte decrease after an addition is a revert*. I applied it all day — I used it correctly to confirm the crawler and gate builds. But I only ever compared each build to my *previous staged attempt*, never to the live binary. Against the prior staged artifact my numbers looked fine; against the thing actually running, it was 167KB short the whole time.d1fe301a · transcript

6★★★★★★A GAP YOU ASSERT WITHOUT GREPPING IS A GAP YOU INVENTED — CHECKING WHAT EXISTS IS CHEAPERproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★A CLIENTHELLO IS A PROMISE — OFFERING A SUITE YOU CANNOT COMPLETE IS A LIE THE SERVER WILLproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★A BUILD THAT REPORTS SUCCESS AND SHRINKS IS A FAILED BUILD WEARING A SUCCESS SHAPE.project-nishi-search-canonical-lane-2026-08-05 · curated

6any existing row ⇒ the host is gone for good. ★★★★★★A TRANSPORT-LAYER CAPABILITY GAP BECOMESproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★COMPONENT GATES ARE NECESSARY, NOT SUFFICIENT — aead1302 4/4 AND transcript384 4/4 bothproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★FOUR AUTHORS EACH DEFERRED THE SAME WIRING, EACH CITING PARTS THAT WERE ALREADY BUILT —project-nishi-search-canonical-lane-2026-08-05 · curated

6🔴★★★★★★MY FIRST CONTROL VARIED THE HOST INSTEAD OF THE BINARY AND WAS THEREFORE USELESS (Iproject-nishi-search-canonical-lane-2026-08-05 · curated

6And it caught a bug that would have poisoned everything downstream. My first version took skins[0] and produced 16 tidy, contiguous, correctly-named parts — every one of them *hair*. Seed-san has one skin per mesh, so a skin-local joint index means a different bone in every skin. The fix resolves mesh → node → skin and labels by node index, which is globally unique. The law is worth keeping: a05646302 · transcript

6That's now a law: a caret that points at the wrong line is worse than no caret — it sends the reader to innocent code, which is precisely the confusion the feature exists to remove. And the meta-law that caught it, for the second time today: verify a diagnostic by reading its emitted bytes, never the source that prints it. The first catch was the off-by-one sys_write length that silently ate a3fb80e53 · transcript

6Everything live and honest. Closing the loop with the debt row and the laws.1aa138c6 · transcript

62. The transcript is shared by client, server, and mTLS — which means sites.elf, the mgmt API, and the tools daemon. Those are the services carrying every tool call I'm making. Breaking the transcript would lock me out with no way to revert. That's this estate's own banked law: *do not put the only road back inside the thing you are changing.* It's the same reasoning that kept nx_mgmt_apid1fe301a · transcript

6🟢🔴★★★★★★08-05 CODEC COVERAGE→IMGSEARCH SOTA — 5 rungs GREEN one session (JPEG-RST · GIF colour · PNG Adam7 · BMP+TIFF · C3 LOSSY VP8 KEYFRAME, gate 8/8 exit 0); nx_image_gray = PROJECTION of the ONE chokepoint; engine 985‰ 16/17 zero-regression; corpus 650‰→1000‰ (120/120, 0 undecodable; webp 42/42). Tables GENERATED from RFC6386 (3268 consts, 0 keystrokes); oracle = PIL/libwebp tile means ±2.MEMORY · curated

6🟢🔴★★★★★★08-05 SEARCH: 8 DEFECTS, ALL LIVE — "small result pool" root-caused: COVERAGE, NOT RANKING (13 of 27 ruler queries wanted hosts holding 0-2 docs, all seeded + HTTP 200 + STARVED). R4/b/c+R5 crawler (gate 20/20; burn-down REVERSED) · R1f+R6 serve (dss 42/42; bare site: false-zero DEAD) · R7 CC artery (refusals→0; corpus 89K→97.8K in a day) · R8 trust-aware ?. Paper RT-011 LIVE, 10 debtsMEMORY · curated

6★★★★★★CHECK-BEFORE-SHIP CAUGHT MY OWN MISREPORT (2026-08-05 night — OPERATOR STANDING ORDER: "check it before you ship, that should be our native state")project-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★THE NET IS THE AUTHORITY: with the root fixed I flipped both-directions on → gate went RED 4/10 (jpeg_ascii, tls13_client_session_recv_sh, p256_keyshare, tls13_p256_loopback; challenger REFUSES to compile them, other 6 + selfhost GREEN) → reverted the flip, kept only the proven half. Residue narrowed 57/70 corpus targets → 4 crypto/codec rows = an unnamed cause 2; the measurement + methodproject-nishilang-sota-radar-2026-08-05 · curated

6My "param-type is RECORD-ONLY" claim (shipped to radar+debt+memory) was FALSE: the 2026-08-01 call-arg type check ENFORCES pointer-into-integer (bite-witnessed: argument type mismatch ... arg 1 is a POINTER but the parameter is an INTEGER, refuse+accept, witness nx_probe_ptrint_live.nx). My ptr-literal probe had tested the documented VK_CONST_INT exemption (integer literals = sentinel/null idiom,project-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★THE ENABLER WAS AN INSTRUMENT FIX, NOT A COMPILER ONE: nx_cc_equiv_gate was sending the compiler's stderr to /dev/null, so 4 prior sessions could see THAT rows failed but never WHY. Now captured per side (/tmp/eq_cc_a.log / _b.log, sizes in each row line) — the very first capture named the exact call (tls13_ext_emit_supported_versions_tls13, nx_tls13_hello.nx out + o). A GATE THAT REPORTS Aproject-nishilang-sota-radar-2026-08-05 · curated

6★★★★★★A GATE RUN OVER THE REAL CORPUS IS A BACKFILL CENSUS — equiv 10/10 + selfhost GREEN *was* the proof that the estate has zero non-exhaustive matches, so enforcement shipped with no amnesty phase. Backfill-before-enforcement satisfied by measurement, not assumption. Promoted 618,136B, canary GREEN.project-nishilang-sota-radar-2026-08-05 · curated

6(the distant cue's px term is ~10x the near one -- that is the law, printed)05646302 · transcript

6Still 18.8KB. Trimming my entry further — the laws live in the topic file, the index only needs the hooks:d1fe301a · transcript

6the morning's CC law: verify by the READER, never the writer's own tally. It cost me a buildproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★AN OK FROM A WRITER IS NOT PROOF THE CONTENT CHANGED — READ THE BYTES BACK. (Exactlyproject-nishi-search-canonical-lane-2026-08-05 · curated

6★★★★★★WHEN THE SOURCE IS PROVEN CORRECT AND THE TOOLCHAIN PROVEN HONEST, THE DEFECT IS IN THEproject-nishi-search-canonical-lane-2026-08-05 · curated

6🟢★★★★★★08-05 NISHI-OS N1-FULL — the shim READS /KERNEL.NXE off the ESP (embed deleted; BOOTX64 22016→1536B; census=EXACT 6 callrax+1 refusal+1 jmp; kernel growth no longer touches the shim). Real-TianoCore FAT witness (tally RISES across EBS 640→2080px; PCI 2×VERIFIED+2×UNBOUND) + literal NXE! neg on real serial. ✅08-05b SLITAZ-PARITY RULER LIVE /compare/smallos (42 axes/8 cols, MEASURED-HONEST,reference-index-recent-refs-dormant · curated

6🔴★★★★★★THE CALL CENSUS WAS A BYTE SCAN AND IT FALSE-FLAGGED ITS OWN lea: bad=1@349 decodedproject-nishi-os-nxe-native-boot-gui-2026-08-04 · curated

6★★★★★★A CORRECTION IS ITSELF A HYPOTHESIS AND MUST PASS THE EXPERIMENT IT DEMANDS (re-provenproject-nishi-codec-coverage-imgsearch-sota-2026-08-05 · curated

6Two failures, both diagnosable. The emulator's T8 predates name-aware Open (my harness needs the name), and the loader census flags a stray FF — the banked law says decode it, don't guess. Adding the byte-context print the law requires:1885a1e5 · transcript

6★★★★★★AN IMPORTED FACT IS NOT A MEASURED FACT (provenance gains IMPORTED; the round-trip gate scoresproject-nishi-vrm-donor-matrix-2026-08-04 · curated

6★★★★★★LOSSLESSNESS LIVES IN THE MEASUREMENTS; THE ONTOLOGY IS ONLY AN INDEX OVER THEM. An object isproject-nishi-vrm-donor-matrix-2026-08-04 · curated

6★★★★★★DO NOT CHOOSE A REFERENCE — OVER-DETERMINE AND LET THE RESIDUAL JUDGE. N cues vs 1 unknownproject-nishi-vrm-donor-matrix-2026-08-04 · curated

6<div class='meth'>How this is scored. This is a state-of-the-art comparison across the FULL competitor field: quantitative axes carry measured / published numbers (Nishi’s column is our own measurement, competitors are researcher-sourced), grade axes use Best / Yes / Part / No. Every axis carries a source note. No single vanity ‘coverage’ score — the1885a1e5 · transcript

6Four earlier sessions had hunted this and reported "isolated repros are all quiet." They were — because the bug is order-dependent, not shape-dependent. Two probes found it: one placed the known-refusing call *after* a cast on the same local (the refusal vanished — mutation reproduced), one cast a *different* local (still refused — so it's per-value, not check-wide). That's now a banked law: when3fb80e53 · transcript

6The scan loop overwrites its own cursor to signal "found" — destroying the position it exists to report. That's a banked law I just re-broke (4th occurrence in the estate). Fixing with a flag, in both trees to honor the merge-discipline law:05646302 · transcript

6My literal probe had tested the *documented exemption* and I read its green as absence-of-check. That's now banked as a law: a green from an exempted expectation is as false as a red from an impossible one — read the check's own exemption list before concluding from a probe.3fb80e53 · transcript

6R1f — the authority denominator fix. One line, but it's the ranking arc the R2d campaign named: stem-expanded variants were inflating the idf-mass obligation so no document could ever reach the 50% floor — the entire authority/trust branch was dead code on natural-language queries. The obligation now covers only original terms while credit still spans variants. The law: an obligation overd1fe301a · transcript

6★★★★★★nx_cc FINALLY HAS AN A/B NET: nx_cc_equiv_gate was RED for a stale baseline path + /tmp IS noexec (products never exec'd, immune to chmod) ⇒ now GREEN 10/10+selfhost. STANDING: every nx_cc change runs nx_cc_equiv_gate <candidate-elf> BEFORE promotion.MEMORY · curated

6🟢🔴★★★★★★08-05 NISHI-OS SPORE + nx_cc A/B NET — HW2 op-list interpreter LIVE (row carries a PROGRAM: READ/AND/EXPECT/SHR/END; row0=4 steps MASK-form, row1=10 steps SHR-form + BAR0-is-memory), witnessed on real OVMF + a real e1000 with the bite run on real hardware; img 946bb0db. Off-LAN lane → OFF-LAN API.MEMORY · curated

6★★★★★★IT CAUGHT THE FIRST CHANGE IT EVER GATED (my duplicate-let check ⇒ 6/10 rows + selfhost FAIL). ★★★★★'IT COMPILES' IS NOT 'IT COMPILES THE ESTATE' · ★★★★★TWO FAILURES THAT AGREE ARE NOT AN EQUIVALENCE (vacuity guard now live) · ★★★★★BACKFILL BEFORE ENFORCEMENT — same-block scoping BUILT+GREEN but inert: the estate declares one name TWICE IN ONE BLOCK (1785952187; parent 1785933085).MEMORY · curated

6What the renders showed: both families carry their design language cleanly through the new conversion path. The law page reads as a dignified firm page — serif headline, small-caps labels, squared navy inputs with the gold focus ring, the one-tap call button as the clear primary action, mailto beneath it, the working form below, and the attorney-advertising disclaimer sitting quietly at the rule.fd4df1ad · transcript

6🟢🔴★★★★★★08-05 CODEC COVERAGE→IMGSEARCH SOTA — 4 rungs GREEN one session (JPEG-RST real-500-RST witness · GIF colour+deinterlace · PNG Adam7 · BMP+TIFF readers); nx_image_gray = PROJECTION of the ONE chokepoint; engine 985‰ 16/17 zero-regression; corpus 650‰, residual = 100% lossy-VP8 ⇒ C3 IS the rung (debt 1785936277). ★★★A BIT-WRITER THAT ORs MAKES BUFFER REUSE A SILENT CORRUPTOR · CHECK WHICHMEMORY · curated

6🔴★★★★★★08-05 nx_txtscan BYTE-ORIENTED SEARCH + 0x0E CORRUPTION (19/19, 3 BITE-PROVEN) — ★★★★★★WE DID NOT ONLY AVOID THE UNIX TOOLS, WE INHERITED THEIR UNIT OF MEANING: both sovereign greps are LINE-oriented ⇒ a 112KB ONE-LINE corpus returns the WHOLE FILE as matches=1 ⇒ the record must be an OFFSET. ctl found 211 laptop .nx each w/ ONE 119-byte 0x0E run killing the emit loop ⇒ ★★★★★A COMPILERMEMORY · curated

6🟢🔴★★★★★★08-05 SEARCH R4 CANONICAL LANE — ruler 366 RED root-caused: COVERAGE STARVATION, NOT RANKING (who.int=0, britannica=0, python.org=1, wikipedia~64 of 89K — all seeded, all HTTP 200, all starved). ★★★★★★A FLAT WALK OVER A FRONTIER THAT GREW JUNK-FIRST IS A PERMANENT BLOCKADE OF ITS OWN HEAD · A RANKING CAMPAIGN AGAINST A COVERAGE GAP ARGUES ABOUT THE ORDER OF DOCS THAT DO NOT EXIST. LaneMEMORY · curated