Binaries pinned on this page — libnrt.so → libnrt.so.2.31.24.0 (aws-neuronx-runtime-lib 2.31.24.0-0b044f4ce, BuildID 8bb57aba0fb2e0035f1d88e9fc4fb3e7387c102e, ELF64 DYN x86-64, not stripped, full DWARF v4); libnccom.so.2.31.24 (aws-neuronx-collectives 2.31.24.0-1a31ba186, BuildID 9c00176c081788c9435d27d11bb40e92495463f0, not stripped, .debug_info); neuron.ko (aws-neuronx-dkms 2.27.4.0, GPL-2.0 C source authoritative, binary stripped); libncfw.so (SONAME libncfw.so.2.31.1.0.cf13a49f, BuildID a98f8e1ca2294582835310c3a1092e0a5e500db5, no DWARF); libnrtucode_extisa.so (BuildID 7bb03bc42ce1530924a1797ec9d5e518a7ae5e44, stripped, no DWARF); plus the static archive libnds.a (6 TUs, 810,674 B, the same code linked into libnrt.so).
Part 0 — Reference Apparatus / REFERENCE · Evidence grade: this page grades other pages; its own claims are the per-page status tags carried in each cited page's own header blockquote, plus the DWARF coverage facts from the source-map and heavy-frame census re-run against these binaries · back to index
This is the book's honest self-assessment. Every other page asserts a confidence in its own header — "Reimplementation-grade", "byte-anchored", "boundary edge" — but a reader landing on one page cannot see how the whole stack lines up: which subsystems are decoded to the byte and which are an opaque blob with a name stuck on it. This page is that ledger. It assigns one of three evidence grades to every shipped page and to every subsystem cluster, names the principal gap that keeps each cell below the top grade, and routes the genuinely-opaque items to the Phase-3 Deep-Dive Backlog. It is a catalogue: its quality is the completeness and the honesty of the map, not new derivation.
The three grades are deliberately coarse. Byte-level decoded (reimplementation-grade) means a competent engineer could rebuild the artifact — wire format, ISA encoding, struct layout, algorithm — from the page alone, because the bytes were read and the decision logic recovered. Surface-mapped means the role, call surface, and inputs/outputs are pinned but at least one interior is not decoded to the byte: a config TU with no recovered static caller, a serializer with no DWARF, an upload path summarized rather than re-derived. Not-traced means the page is honest about an opaque region — a sequencer-internal microcode op, a per-arch leaf body left at a name, an interior math kernel that was sized but not walked. The grade is the floor of the page: a page that is byte-level on its main artifact but leaves one helper opaque is graded by what a reimplementer still cannot rebuild. Two of the five binaries — libncfw.so and libnrtucode_extisa.so — carry no DWARF and (for extisa) no symbols at all, which sets a hard ceiling on every page that depends on them; that ceiling is named explicitly in §5.
NOTE — "surface-mapped" is not a failure grade. For vendored OSS (protobuf, Abseil, simdjson, libarchive, zlib, the Rust crates) surface-mapping is the correct terminal state — a reimplementer links the pinned version rather than reversing the arena allocator. Those rows are graded by whether the boundary (which symbols, which version, what they back) is decoded, not the library interior. The version pins live in the vendored SBOM; this page does not re-grade them.
One row per shipped page, grouped by Part; the SUMMARY path is exact. Grade is the page floor per the legend. Status is a one-phrase state; Principal gap names the single thing that keeps the cell below byte-level (— if none). Conf grades this assessment of the page, not the page's own internal claims.
The artifacts a reimplementer can rebuild from the page alone, each with its owning page. These are the load-tested cores of the book — wire formats read field-by-field, ISA catalogues sized against their own operand tables, struct layouts confirmed against multiple sources.
These pages pin the role and call surface but stop short of a byte-level interior — correctly, because the interior either lacks a recovered caller, lacks DWARF, or is upstream OSS a reimplementer should link rather than reverse.
Some HAL config TUs — 17 of 24 al_hal_udma_config functions have no recovered static caller in libnrt; their role is known, their invocation context is not → runtime/hal-udma-iofic.md.
template/STL-inlined; folded into enc.cc by address, never name-cited
deep-dive-backlog (switch-platform events)
GOTCHA — "not-traced" never means "absent from the book." Each opaque region has a page that names its boundary, its inputs, and its callers; what is missing is the byte-level interior. A reimplementer treats the boundary as a contract and the interior as a build-against-the-binary task — not as a hole the book pretends is filled.
Two of the five binaries set a hard ceiling that no amount of analysis lifts to byte-level on internal names:
libncfw.so (BuildID a98f8e1ca…) carries no DWARF and self-versions in its SONAME (libncfw.so.2.31.1.0.cf13a49f) rather than in .git_hash/.nrt_brazil_version sections. Its public boundary (the libncfw_get_image_func get-image entry, the embedded blob set, the v2 reset prologue) is byte-anchored; its internal serializer and handler names are not recoverable, so those pages are graded surface-to-MED by construction.
libnrtucode_extisa.so (BuildID 7bb03bc4…) is stripped — no DWARF and no symbol table. It is reached only as a dlopen target whose 30-entry dlsym binding table names its API surface; everything below that surface is decoded by raw disassembly against the carved blobs, not by symbol.
NOTE — the byte-identity QUIRK that links these two carriers (the SUNDA v2 sequencer boot scaffold ships byte-identically in both, per firmware/upload-path.md) is itself a byte-level fact recovered despite the no-DWARF ceiling — proof that the ceiling caps internal naming, not byte-level structural claims. The grade floor of a no-DWARF page is therefore surface on names but can be byte-level on bytes; the table above grades the names a reimplementer would need, which is the binding constraint.
Subsystem ↔ Binary ↔ Source-TU Matrix — the orthogonal view: which TU in which binary implements each subsystem (this page grades coverage; that page maps ownership)