External bug-boards, RFCs and papers → Nishi Lang capability requirements & autonomous-engineering answer-keys. Real corpora as sovereign ORACLES (bench-only, never their code).
SOTA-LIVEfull-scale, no samplinginjection-scanned GREENtrust-fence PROVENsovereign authed fetch 5000/hrTwo real sources, fetched at full scale via a paginating sovereign fetch (stops on rel="next" gone, verified against the rate budget so it is never a truncated sample):
| Source | Total | Solved (answer-keys) | Open (gaps) |
|---|---|---|---|
rust-lang/rust I-unsound | 669 | 546 | 123 |
llvm/llvm-project miscompilation | 982 | 726 | 256 |
Every real bug is classified onto a Nishi Lang capability axis by a deterministic, data-driven, case-insensitive keyword map. Each axis is GAP against the measured 0 / 184 primitive registry — a filed requirement, ranked by real oracle-bug demand:
| # | Capability axis | Oracle bugs |
|---|---|---|
| 1 | compiler-miscompile-codegen | 638 |
| 2 | compiler-vectorization | 78 |
| 3 | compiler-poison-undef | 54 |
| 4 | memory-safety-pin | 24 |
| 5 | memory-safety-alignment | 17 |
| 6 | target-abi-soundness | 16 |
| 7 | type-soundness-lifetime / compiler-integer-overflow | 13 |
| 8 | move-safety-aliasing | 10 |
| … | + ownership-drop, trait, stack, atomics, repr, uninit, bounds, double-free … | 1–7 |
One OO base class nx_ingest_base carries the shared parser + primitives; every stage is a thin subclass — consolidation without losing functionality:
paginating fetch → ingest (Rust + LLVM connectors) → classify (data-driven hot-read conf) → verify (full-plane histogram, defeats the 64KB read cap) → rank (router → requirements registry) → integrity-scan.
The fetch is fully sovereign — our own TLS 1.3 client with authenticated GitHub access at 5,000 req/hr (runtime-proven: 200 on an auth-only endpoint; no curl, no GraphQL, no libc) — so all 1,272 fix-diff answer-keys can be captured full-scale, each passed through the trust fence before it can become training fuel.
The 1,272 solved bugs are (problem, solution) pairs in the SWE-bench shape — the training fuel for the ecosystem's autonomous-fix loop, which already resolves 71% (FAIL_TO_PASS + PASS_TO_PASS, judged by fresh compile+run, never by trusting the patch).
External data is an untrusted / poisoning surface, so it is guarded before, during, and after submission, deny-by-default, consolidated onto the ecosystem guard lifecycle:
| Stage | Control | Status |
|---|---|---|
| PRE-submit | Sovereign TLS 1.3 + trust store; trust-tier fence — a fix is a tier-1 answer-key ONLY if merged into the canonical repo (not a fork) with a real merge-commit = maintainer-vetted; injection scan of all 1,651 titles | PROVEN — real rust-lang/rust#50000 accepted; 3 crafted trojans (unmerged / attacker-fork / null-merge) each rejected for the correct reason; injection 0/1,651 |
| AT-submit | never-execute (patches are data, never run during ingest); isolated never-brick sandbox for patch-apply | PRE done; sandbox building |
| POST-submit | TESTS-JUDGE-NOT-TRUST (a patch failing FAIL_TO_PASS+PASS_TO_PASS is rejected); adversarial CI re-proves the trojan rejections each run | building — gated before any patch trains |
Sovereign Nishi ecosystem · LangIntel capability · all figures measured full-scale, never sampled. Corpus on the Nishi information plane: langintel-requirements-, langintel-rustunsound-, langintel-llvm-, langintel-axismap-.