Home › Census

Function-name census Syntology

One name, many papers, one shared input. For each function or class name that appears in several papers' harvested code, Syntology ran every implementation it holds on one shared input fixed by the rank, kind and dtype of the implementation's array arguments, and recorded what each returned. Implementations whose recorded output digest is identical form a cluster; a name whose implementations fall into more than one cluster under the same bucket is split: the field means more than one computation by that name. Identical values on the shared input are agreement on those inputs, not equivalence, and no substitution claim is made. This is Syntology's own measurement; nothing on these pages is from the archive except paper titles, which are labelled. The per-paper sample record is on Code.

In numbers Syntology

Census date and battery sha are read from the provenance sidecars of the 17 result files; the battery sha is the sha256 of the harness that generated the shared input; the run driver's sha256 is 5a452d0e7c0d. Union of 17 files, 9,976 lines, 9,916 distinct samples after de-duplication on sha (0 duplicate lines dropped), 3,320 names and 7,537 papers examined. Of the samples, 4,946 ran on the shared input and were compared; 2,689 did not run on it; 2,272 took no array argument, so no shared input existed for them and they ran only on their own fixture arguments; 9 ran but returned an output the harness could not digest. Each of the three is counted per name on its page and none is in the totals above; a further 60 rows are the harness's own torch computations under named conventions and are not implementations. A name is reported when the implementations that were compared come from at least 5 distinct papers.

Bucket key. A bucket is the positional (rank, kind, dtype) of each array argument; the argument name is not part of the key because the harness draws the shared array from (rank, kind) and casts it to the dtype, whatever the name. Two implementations in one bucket therefore saw bitwise-identical inputs; implementations in different buckets took different arguments and are never compared. 670 buckets across the names below.

What is not on these pages

name not held

A name with no page on Code is not in Syntology's harvested code at all, and has no census line.

held, never run on the shared input

5,910 names that appear in ≥5 papers' harvested code were not put through the census; a further 781 names were put through it and no implementation ran on the shared input.

ran, but only on its own fixture arguments — no shared input

2,272 samples took no array argument, so the harness replaced nothing: their recorded "shared" call is the fixture's own arguments (on 2,225 of them the recorded shared-output digest is the own-fixture digest). They are not compared, do not count toward any name's eligibility, and are not in the tables. 675 names have only such samples (or only undigested ones) and get no page; 133 names would reach 5 papers only by counting them and get no page either. 153 such samples belong to names that do have a page and are counted on it.

ran on the shared input, output not digested

9 samples returned an output the harness could not digest (non-numeric type: NoneType 2, Tensor 5, ndarray 2) and are not compared.

run, but fewer than 5 papers — too thin to report

1,568 names were compared from 1 to 4 papers and get no page.

class-bearing names

The census files hold 1,808 class-bearing samples (a class instantiated with seeded weights in evaluation mode, then called); 586 were compared and appear on 24 of the pages below, tagged per member. For a class, the recorded output depends on the seeded initialisation as well as the forward computation, so a class-bearing name that is split may differ in construction order rather than in what it computes. A class-bearing name that is not in the census files is not yet in the census.

withdrawn fingerprints

The graph export read at build time carries no withdrawal flag on any sample, so none could be excluded on that basis: 0 excluded. 9,916 of the 9,916 census samples were found in the export by sha; 0 were not.

Names with ≥5 papers Syntology

296 names, ordered by papers. Implementations and papers count only what was compared. Buckets are argument buckets (rank, kind, dtype); implementations are only compared within one. Distinct outputs counts clusters across all of a name's buckets. Not run is the number of this name's implementations the harness could not run on the shared input; own args only took no array argument and ran only on their own fixture arguments; not digested ran but returned an output that was not digested. None of the three is compared. 9 page addresses needed a numeric suffix to keep case-different names apart.

Read from groundwork/55_differential_census/census/v2_2026-09-22 at build time. Each name has a JSON twin at /census/<name>.json.

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections