Appendix — Report → Wiki-Page Cross-Reference Map

This appendix is a provenance index over neuronx_cc 2.24.5133.0+58f8de22 wiki artifacts, not a binary-RE page. It is generated by scanning two corpora that live in this repository: the raw evidence reports under neuronx-cc/raw/ (the D-*.txt deep-dive notes) and the published wiki pages under neuronx-cc/wiki/src/. Every edge below is a real rg citation found in a committed page — no mapping is hand-asserted. Because it is repo-derived, this page has no version-pinned addresses of its own; its accuracy is exactly the accuracy of the citations the authors wrote, and it goes stale the moment a page adds, drops, or renames a D-ID. Regenerate it (see §5) after any wiki edit that touches an evidence anchor.

Abstract

Each normative wiki page anchors its claims to one or more deep-dive reports — the D-<family><n>.txt notes under raw/ that record the binary-analysis evidence a page is built on (a disassembly trace, an enum dump, a struct layout, a verifier walk). A page cites a report by writing its D-ID inline in a BACKING: line or an evidence callout. This appendix inverts and totals those citations so a maintainer can answer two questions in O(1):

• Forward (§2): given a report D-Xnn, which page(s) consume it? — so editing or correcting a report immediately shows every page that must be re-checked.
• Reverse (§3): given a page, which reports back it? — so a reviewer can pull the exact evidence behind a page without re-grepping.

The index is also a completeness audit. It flags two failure modes that a provenance map exists to catch: a report cited by zero pages (an unconsumed evidence note — work captured but not yet written up), and a D-ID cited by a page that does not exist in raw/ (a dangling citation — a claim whose evidence is missing or mis-typed). The gap census is §4.

The map is the substance of this page; the prose is scaffolding. Its value is completeness and accuracy of the edges, not narrative. Every (report → page) row was produced by rg -l '\bD-ID\b' against wiki/src/, and every page slug named is a file that exists on disk.

0. How to read an edge

A single edge is a citation. When bir/dtype-tables.md contains the string D-D04, that is the page declaring "the dtype-table claims on this page are backed by the evidence in raw/D-D04-dtype-tables.txt." This appendix records that as the forward edge D-D04 → bir/dtype-tables.md and the reverse edge bir/dtype-tables.md ← D-D04.

Three properties of the citation graph a maintainer should know:

• Many-to-many. A report can back several pages (a struct-layout report like D-E12 accesspattern struct feeds four pages), and a page can rest on several reports (nki/bircodegen-collective.md cites eleven). The map below lists every page for each report (forward) and every report for the highest-fan-in pages (reverse).
• Word-boundary matching. D-IDs are matched with \bD-[A-Z]+[0-9]+\b. This is deliberate: a naïve rg 'D-SB2' would false-positive on the hardware term GPSIMD-SB2SB in walrus/legality-dispatch.md. The \b…\b anchoring rejects that substring, which is why the dangling count is exactly 0 rather than a spurious 1. Any future regeneration must keep the boundary anchors.
• Stem vs file. Two reports were split across a base note and a continuation (D-P02 + D-P02b, plus two D-Z06 variants). They share one D-ID stem, so the corpus is 506 files / 504 stems. Forward edges key on the stem.

QUIRK — a D-ID appearing in a page is not proof the page is correct, only that it claims that report as evidence. This index audits linkage, not truth. Truth is the job of each page's own adversarial-verify section and of §0.3. A row here means "this page says report X backs it"; whether the page read X faithfully is out of scope.

1. Family map

Reports are grouped by the leading-letter family, which tracks the analysis theme the note belongs to. Knowing the family lets a maintainer guess a report's subject from its D-ID alone. The table gives each family's theme, total stem count, and how many of those stems are currently consumed by at least one page.

2. Forward map — report → consuming page(s)

The forward index, grouped by family. Each row is report → the page(s) that cite it. Only consumed reports appear here; the 269 unconsumed reports are listed in §4.1. Page slugs are relative to wiki/src/.

D-A* family

D-AA* family

D-AB* family

D-AC* family

D-AD* family

D-AE* family

D-AF* family

D-AG* family

D-B* family

D-C* family

D-D* family

D-E* family

D-F* family

D-G* family

D-H* family

D-I* family

D-J* family

D-K* family

D-L* family

D-M* family

D-N* family

D-O* family

D-P* family

D-Q* family

D-R* family

D-S* family

D-T* family

D-U* family

D-V* family

D-W* family

D-X* family

D-Y* family

D-Z* family

3. Reverse map — page → backing reports

The inverse of §2: for the highest-fan-in pages (those resting on the most distinct reports), the reports that back them. A page absent here either cites a single report — read it off the forward map — or cites none. The full reverse map for every page is reconstructible from §2 by transposing the edges; this section spotlights the pages a reviewer most wants to audit, because they integrate the most evidence.

4. Gaps — unconsumed reports and dangling citations

A provenance index is only as useful as the gaps it surfaces. Two are actionable.

4.1 Unconsumed reports (269)

These reports exist in raw/ but are cited by zero wiki pages. They represent evidence that was captured during analysis but is not yet written up in a normative page — the to-do backlog of the wiki, not an error. A maintainer planning new pages should mine this list first. Listed by family (the consumed ones are in §2):

D-A* (4): D-A02, D-A04, D-A07, D-A08

D-AA* (6): D-AA01, D-AA02, D-AA03, D-AA06, D-AA08, D-AA09

D-AB* (4): D-AB02, D-AB03, D-AB04, D-AB09

D-AC* (1): D-AC04

D-AD* (3): D-AD02, D-AD04, D-AD05

D-AE* (3): D-AE01, D-AE02, D-AE04

D-AF* (2): D-AF03, D-AF04

D-AG* (7): D-AG01, D-AG02, D-AG03, D-AG04, D-AG05, D-AG06, D-AG10

D-B* (16): D-B05, D-B06, D-B07, D-B11, D-B12, D-B16, D-B18, D-B19, D-B20, D-B21, D-B22, D-B25, D-B26, D-B27, D-B30, D-B32

D-C* (12): D-C03, D-C05, D-C06, D-C08, D-C13, D-C16, D-C17, D-C18, D-C19, D-C20, D-C23, D-C24

D-D* (1): D-D06

D-E* (10): D-E01, D-E04, D-E05, D-E08, D-E10, D-E11, D-E15, D-E17, D-E18, D-E20

D-F* (12): D-F01, D-F05, D-F06, D-F07, D-F09, D-F10, D-F11, D-F12, D-F14, D-F15, D-F16, D-F17

D-G* (5): D-G06, D-G07, D-G09, D-G12, D-G13

D-H* (26): D-H02, D-H04, D-H05, D-H07, D-H08, D-H09, D-H11, D-H12, D-H14, D-H15, D-H18, D-H19, D-H20, D-H22, D-H25, D-H27, D-H28, D-H29, D-H30, D-H31, D-H32, D-H33, D-H34, D-H36, D-H37, D-H39

D-I* (19): D-I02, D-I03, D-I04, D-I05, D-I06, D-I07, D-I08, D-I13, D-I14, D-I15, D-I16, D-I18, D-I19, D-I20, D-I22, D-I24, D-I25, D-I26, D-I27

D-J* (24): D-J01, D-J02, D-J07, D-J08, D-J10, D-J11, D-J12, D-J15, D-J16, D-J17, D-J18, D-J19, D-J21, D-J22, D-J23, D-J24, D-J25, D-J26, D-J27, D-J30, D-J31, D-J33, D-J35, D-J36

D-K* (13): D-K02, D-K03, D-K04, D-K05, D-K06, D-K07, D-K08, D-K09, D-K11, D-K13, D-K16, D-K17, D-K18

D-L* (8): D-L01, D-L03, D-L04, D-L05, D-L06, D-L07, D-L08, D-L09

D-M* (6): D-M01, D-M02, D-M03, D-M09, D-M13, D-M14

D-N* (10): D-N01, D-N03, D-N04, D-N06, D-N07, D-N08, D-N09, D-N10, D-N11, D-N12

D-O* (6): D-O07, D-O13, D-O17, D-O19, D-O20, D-O21

D-P* (8): D-P03, D-P05, D-P07, D-P08, D-P09, D-P10, D-P16, D-P17

D-Q* (4): D-Q01, D-Q02, D-Q03, D-Q07

D-R* (7): D-R01, D-R02, D-R04, D-R05, D-R06, D-R07, D-R09

D-S* (8): D-S05, D-S07, D-S08, D-S09, D-S10, D-S11, D-S12, D-S13

D-T* (6): D-T01, D-T02, D-T03, D-T06, D-T07, D-T08

D-U* (5): D-U03, D-U04, D-U05, D-U06, D-U07

D-V* (12): D-V02, D-V04, D-V07, D-V08, D-V09, D-V10, D-V11, D-V12, D-V13, D-V14, D-V15, D-V16

D-W* (5): D-W01, D-W03, D-W06, D-W07, D-W08

D-X* (8): D-X01, D-X02, D-X03, D-X04, D-X07, D-X08, D-X09, D-X10

D-Y* (5): D-Y02, D-Y03, D-Y04, D-Y07, D-Y08

D-Z* (3): D-Z01, D-Z04, D-Z05

4.2 Dangling citations (0)

No page cites a D-ID that fails to resolve to a raw/ file. Every evidence anchor in the wiki points at a report that exists.

CORRECTION — the one apparent dangle is a regex artifact, not a real one. A loose scan (rg 'D-SB2' without word boundaries) reports D-SB2 as cited-but-missing. It is not a citation: it is the substring SB2 inside the hardware term GPSIMD-SB2SB on walrus/legality-dispatch.md (a SBUF-to-SBUF GPSIMD op, bir::InstGPSIMDSB2SB). Anchoring the match with \bD-[A-Z]+[0-9]+\b rejects it and the dangling count is 0. This is the canonical reason the regeneration recipe in §5 mandates word boundaries — drop them and this index manufactures a phantom gap.

5. How this page is regenerated

This appendix is a pure function of the repo state at generation time. To rebuild it after a wiki edit, re-run the same scans the edges above were produced from. The recipe is the page's reproducibility contract.

## step 1 — corpus of distinct D-IDs (stems) in raw/              -> 504
ls neuronx-cc/raw/ | rg '^D-' | rg -o '^D-[A-Z]+[0-9]+' | sort -u

## step 2 — distinct D-IDs cited anywhere in the wiki (WORD-BOUNDED) -> 236
rg --no-filename --only-matching -g '*.md' '\bD-[A-Z]+[0-9]+\b' \
   neuronx-cc/wiki/src/ | sort -u

## step 3 — forward edge for one report: which page(s) cite D-Xnn
rg -l -g '*.md' '\bD-Xnn\b' neuronx-cc/wiki/src/

## step 4 — consumed / unconsumed / dangling split (comm against the two sets)
comm -12 raw_dids cited_dids   # consumed   -> 235
comm -23 raw_dids cited_dids   # unconsumed -> 269
comm -13 raw_dids cited_dids   # dangling   -> 0   (with \b...\b anchors)

Two invariants any regeneration must preserve, or it will produce wrong findings:

• Always anchor with \b...\b. Without word boundaries, hardware tokens like GPSIMD-SB2SB masquerade as citations and inflate the dangling count (§4.2).
• Key forward edges on the stem, not the filename. D-P02/D-P02b and the two D-Z06 notes share a stem; counting files (506) instead of stems (504) double-counts them.

6. Adversarial verification

Five forward edges and the three census totals, re-checked by hand against the repo. Every check passed.

See also

• §0.3 Methodology & Confidence Model — what a CONFIRMED / INFERRED tag means and how reports earn one.
• §14.4 Symbol / Offset Index — the sibling provenance index, mapping binary symbols and addresses to the pages that document them.
• §14.1 Master Opcode Reference, §14.2 Master Enum & Dtype Table, §14.3 Error-Message Catalog — the consolidated lookup appendices whose own backing reports appear in §2.