This document describes the systematic enumeration of orthographic zomeable projections for all 47 convex uniform 4-polytopes, beyond the 8 polytopes (6 regulars + snub 24-cell + grand antiprism) already documented in this repo.
The 47 unique convex uniform 4-polytopes are generated as Wythoff constructions on the rank-4 Coxeter groups:
| Coxeter group | diagram | Wythoff bitmasks | unique forms |
|---|---|---|---|
| A₄ | 3-3-3 | 9 (palindromic) | 9 |
| B₄ | 4-3-3 | 15 | 15 |
| F₄ | 3-4-3 | 9 (palindromic) | 8 (24-cell ≡ B₄ rectified-16-cell) |
| H₄ | 5-3-3 | 15 | 15 |
| total | 48 | 47 |
Related corpus-completeness analysis: see
INHERITANCE_FREE_SWEEP.mdfor a 38-hour, 180-cell matrix sweep that searches every (group, bitmask, rng ≤ 4) cell without assuming kernel-inheritance from the parent regular. It confirms that the production sweep below misses zero shapes that would snap toℤ[φ]³.
Plus the 2 non-Wythoff convex uniforms (snub 24-cell, grand antiprism)
gives the well-known total of 47 + 2 ≡ 49 — but the snub 24-cell is
Coxeter-equivalent to B₄ (0,0,1,0) snubbed
and is already enumerated in output/uniform/snub_24cell/. Likewise the grand
antiprism is in output/uniform/grand_antiprism/.
Throughout this repository — including filenames like
tools/emit_novel.py, ongoing_work/novel_rng<N>.json, and the
[NOVEL] tag in tools/analyze_sweep.py — the word “novel” is
used in a narrow, deduplication sense:
codebase “novel” = a Wythoff projection’s 3D shape that is not congruent to any of the 6 regular polytopes’ own zomeable projections.
This is a redundancy filter, not a structural-novelty claim. The
regular polytopes’ projections are catalogued separately in
output/{5cell,8cell,16cell,24cell,120cell,600cell}/; the codebase’s
“novel” tag just keeps the Wythoff-sweep output from re-emitting any
Wythoff projection that happens to coincide with one of those.
Empirically at rng=2 the filter caught 0 cases (descendant
polytopes have strictly more vertices than their regular, so their
projections cannot coincide).
This is distinct from, and weaker than, the more intuitive meaning of “genuinely new”:
structurally novel = a 3D shape that is not obtainable by projecting some Wythoff polytope along a kernel direction that already worked for one of the 6 regulars.
Under the structurally-novel definition, the set is provably empty
at every kernel range, by Lemma A (the representation-theoretic
shortcut on which the whole sweep is built — see §Method
below): every zomeable kernel of any Wythoff variant P_G(b) is also
a zomeable kernel of the regular P_G((1,0,0,0)) of the same group.
So every Wythoff zomeable shape this sweep produces is a “natural
Wythoff extension” of a regular’s projection — applying the
truncation/cantellation/etc. operation in 4D and projecting under a
kernel direction the regular already provides.
The 69 .vZome files in the output/<polytope>/ folders are therefore best
understood as “every distinct natural Wythoff extension of the 6
regulars’ projections at rng=2, modulo 3D-shape congruence and
modulo coincidence with a regular’s own projection”. There are no
“surprise” shapes outside this scope.
We re-use the search engine from lib/search_engine.py (the same engine
used for the 8 canonical polytopes) but exploit the following
representation-theoretic shortcut:
For any Wythoff construction
P_G(b)on Coxeter group G, every kernel directionn ∈ ℤ[φ]⁴that yields a zomeable orthographic projection ofP_G(b)is also a zomeable kernel for the regular polytopeP_G(1,0,0,0)of the same group.
(This is empirically verified up to the rng we test: the regular’s hit set strictly contains the variants’ hit sets across A₄/B₄/F₄/H₄ at rng=2.) So the sweep proceeds in two steps:
rng. (rng=2: |a|, |b| ≤ 2 in a + bφ.)Both steps live in tools/run_wythoff_sweep.py. Wythoff polytopes are
generated by lib/wythoff.py (Cholesky-frame simple roots, then
Procrustes-aligned to lib/polytopes.py’s canonical embedding so kernel
directions agree across the regular and its Wythoff descendants).
| group | regular polytope | |V| | |E| | step-1 kernels |
|---|---|---|---|---|
| A₄ | 5-cell | 5 | 10 | 324 |
| B₄ | tesseract | 16 | 32 | 2640 |
| F₄ | 24-cell | 24 | 96 | 608 |
| H₄ | 120-cell | 600 | 1200 | 432 |
Step 1 for H₄ takes ~48 minutes single-threaded; results are cached at
ongoing_work/kernels_<group>_rng2.npy after first computation.
All shape counts below are at rng = 2 across the full corpus of 47 unique convex uniform 4-polytopes.
Each table reports two columns:
lib.search_engine.shape_fingerprint
hashes produced by step-2 of the search (the pre-dedup, pre-emit
count)..vZome files that actually survive in
output/<polytope>/ (the post-dedup, post-emit count).
This is the user-visible corpus. For the 6 regulars (5-cell,
tesseract, 16-cell, 24-cell, 600-cell, 120-cell) the corresponding
models live in their own dedicated folders (output/regular/5cell/,
output/regular/8cell/, output/regular/16cell/, output/regular/24cell/,
output/regular/600cell/, output/regular/120cell/) and are noted as → <folder> in
the files column rather than counted here.The B₄ tesseract figure of 32 shapes matches the documented
tesseract enumeration at rng=2 in
output/regular/8cell/CLASSIFICATION.md:
“rng = 1, 2, 3 (yielding 8, 32, 129 distinct shapes — the count grows
without bound)”. The growth comes from the infinite split-cuboid family
(kernels of support 2; see CLASSIFICATION.md §3); the other 14 B₄
forms have only 1–3 shapes each because their edge directions break the
support-2 kernel automatic-zomeability argument.
² The raw fp column is the original step-2 census, kept here as a reference because some figures (notably the omnitruncated 120-cell’s 65 distinct fingerprints, and the cantitruncated 600-cell’s 26) are of independent interest as sweep-engine load metrics. The files column is what is shipped on disk after three downstream filters are applied to that raw census:
tools/dedup_corpus_by_direction.py):
collapses orbit-equivalent kernels at φᵏ scale that
shape_fingerprint’s 3-decimal binning had over-counted.status=snap_failed in the
manifest and no .vZome file is emitted. This is why several
B₄/F₄ rows show files = 0 despite a non-zero raw-fp count
(e.g. truncated tesseract: 3 raw fp, all 3 snap-fail → 0 files).tools/dedup_corpus_by_shape.py): the rotation/reflection/scale-
invariant pairwise-distance multiset, computed directly on the
emitted .vZome ℤ[φ] coordinates, finds orbit-equivalent kernels
at cos = ±½ that Stage A missed. This is the dominant collapse
for high-V H₄ polytopes (omnitruncated 120-cell: 38 ok → 1 file).The omnitruncated 120-cell row is the most extreme example: 65 raw fp → 38 ok manifest entries (Stage A) → 1 file (Stage B).
| bitmask | name | V | E | raw fp² | files² |
|---|---|---|---|---|---|
| (1,0,0,0) | 5-cell | 5 | 10 | 4 | → output/regular/5cell/ |
| (0,1,0,0) | rectified 5-cell | 10 | 30 | 4 | 4 |
| (1,1,0,0) | truncated 5-cell | 20 | 40 | 4 | 4 |
| (1,0,1,0) | cantellated 5-cell | 30 | 90 | 4 | 4 |
| (0,1,1,0) | bitruncated 5-cell | 30 | 60 | 4 | 4 |
| (1,1,1,0) | cantitruncated 5-cell | 60 | 120 | 7 | 4 |
| (1,0,0,1) | runcinated 5-cell | 20 | 60 | 4 | 4 |
| (1,1,0,1) | runcitruncated 5-cell | 60 | 150 | 6 | 4 |
| (1,1,1,1) | omnitruncated 5-cell | 120 | 240 | 10 | 4 |
A₄ subtotal: 32 files.
| bitmask | name | V | E | raw fp² | files² |
|---|---|---|---|---|---|
| (1,0,0,0) | tesseract | 16 | 32 | 32 ¹ | → output/regular/8cell/ |
| (0,1,0,0) | rectified tesseract | 32 | 96 | 3 | 3 |
| (1,1,0,0) | truncated tesseract | 64 | 128 | 3 | 0 (snap-fail) |
| (0,0,1,0) | 24-cell | 24 | 96 | 2 | → output/regular/24cell/ |
| (1,0,1,0) | cantellated tesseract | 96 | 288 | 3 | 0 (snap-fail) |
| (0,1,1,0) | bitruncated tesseract | 96 | 192 | 3 | 3 |
| (1,1,1,0) | cantitruncated tesseract | 192 | 384 | 3 | 0 (snap-fail) |
| (0,0,0,1) | 16-cell | 8 | 24 | 3 | → output/regular/16cell/ |
| (1,0,0,1) | runcinated tesseract | 64 | 192 | 3 | 0 (snap-fail) |
| (0,1,0,1) | cantellated 16-cell ³ | 96 | 288 | 2 | → rectified_24-cell/ |
| (1,1,0,1) | runcitruncated tesseract | 192 | 480 | 3 | 0 (snap-fail) |
| (0,0,1,1) | truncated 16-cell | 48 | 120 | 3 | 3 |
| (1,0,1,1) | runcitruncated 16-cell | 192 | 480 | 3 | 0 (snap-fail) |
| (0,1,1,1) | cantitruncated 16-cell ³ | 192 | 384 | 2 | → truncated_24-cell/ |
| (1,1,1,1) | omnitruncated tesseract | 384 | 768 | 3 | 0 (snap-fail) |
B₄ subtotal: 9 files. Five 0 (snap-fail) rows have raw fp > 0
but every candidate shape failed the snap-to-ℤ[φ]³ step required for
vZome export — see the
snap-failure analysis
below for why this happens (e.g. the truncated tesseract has shapes
in ℤ[√2]³ but not ℤ[φ]³).
³ The cantellated 16-cell is the same uniform 4-polytope as the rectified 24-cell, and the cantitruncated 16-cell is the same as the truncated 24-cell (see the Equivalences section below). Their .vZome files live under the canonical F₄ folder.
¹ The 32 tesseract shapes at rng=2 are the finite truncation of an
infinite cuboid family. See
output/regular/8cell/CLASSIFICATION.md
for the master theorem and the family parametrization. This is a
property of the tesseract specifically (its edge directions align with
the 4D coordinate basis); the 14 other B₄ forms do not enjoy this
support-2 automatic-zomeability and have only 1–3 shapes each.
| bitmask | name | V | E | raw fp² | files² |
|---|---|---|---|---|---|
| (1,0,0,0) | 24-cell ³ | 24 | 96 | (≡ B₄ (0,0,1,0)) | → output/regular/24cell/ |
| (0,1,0,0) | rectified 24-cell ³ | 96 | 288 | 5 | 3 |
| (1,1,0,0) | truncated 24-cell ³ | 192 | 384 | 6 | 3 |
| (1,0,1,0) | cantellated 24-cell | 288 | 864 | 2 | 0 (snap-fail) |
| (0,1,1,0) | bitruncated 24-cell | 288 | 576 | 1 | 0 (snap-fail) |
| (1,1,1,0) | cantitruncated 24-cell | 576 | 1152 | 2 | 0 (snap-fail) |
| (1,0,0,1) | runcinated 24-cell | 144 | 576 | 1 | 0 (snap-fail) |
| (1,1,0,1) | runcitruncated 24-cell | 576 | 1440 | 2 | 0 (snap-fail) |
| (1,1,1,1) | omnitruncated 24-cell | 1152 | 2304 | 1 | 0 (snap-fail) |
F₄ subtotal: 6 files (3 rectified + 3 truncated 24-cell). The rectified- and truncated-24-cell counts include shapes originally discovered under the equivalent B₄ Wythoff names (see footnote ³ on the B₄ table and the Equivalences section below). Six F₄ forms snap-fail because their ball-coordinate fields contain √2; see the snap-failure analysis below.
All 15 H₄ polytopes complete. The omnitruncated 120-cell (V=14400)
required ~137 min CPU on a 24-core box for the step-2 group-by-shape
pass; bounded RAM was achieved by the multiset shape-fingerprint path
in lib/search_engine.shape_fingerprint (see implementation notes).
| bitmask | name | V | E | raw fp² | files² |
|---|---|---|---|---|---|
| (0,0,0,1) | 600-cell | 120 | 720 | 1 | → output/regular/600cell/ |
| (1,0,0,0) | 120-cell | 600 | 1200 | 1 | → output/regular/120cell/ |
| (0,0,1,0) | rectified 600-cell | 720 | 3600 | 2 | 1 |
| (0,1,0,0) | rectified 120-cell | 1200 | 3600 | 1 | 1 |
| (0,0,1,1) | truncated 600-cell | 1440 | 4320 | 2 | 1 |
| (1,1,0,0) | truncated 120-cell | 2400 | 4800 | 3 | 1 |
| (1,0,0,1) | runcinated 120-cell | 2400 | 7200 | 4 | 1 |
| (1,0,1,0) | cantellated 120-cell | 3600 | 10800 | 5 | 1 |
| (0,1,1,0) | bitruncated 120-cell | 3600 | 7200 | 4 | 1 |
| (0,1,0,1) | cantellated 600-cell | 3600 | 10800 | 5 | 1 |
| (1,0,1,1) | runcitruncated 600-cell | 7200 | 18000 | 11 | 1 |
| (1,1,0,1) | runcitruncated 120-cell | 7200 | 18060 | 6 | 1 |
| (1,1,1,0) | cantitruncated 120-cell | 7200 | 14520 | 11 | 1 |
| (0,1,1,1) | cantitruncated 600-cell | 7200 | 14508 | 26 | 1 |
| (1,1,1,1) | omnitruncated 120-cell | 14400 | 28800 | 65 | 1 |
H₄ subtotal: 13 files. Every H₄ Wythoff descendant collapses to a single 3D shape after Stages A + B.
(The cantitruncated, runcitruncated, and omnitruncated 7200/14400-vertex
polytopes report slightly more than 7200/14400 vertices and the
expected edge counts in the engine due to a known tol_decimals=8
rounding issue in lib.wythoff.orbit; the correct kernels and shape
counts are still found because the search engine deduplicates
internally.)
H₄ raw-fp diversity grows sharply with V: the cantitruncated 600-cell contributes 26 distinct fingerprints alone, and the omnitruncated 120-cell 65 — every one of its 432 step-2 hits lands in a distinct fingerprint equivalence class. After Stages A + B all 65 collapse to the same 3D shape (cell-first truncated icosidodecahedron).
| group | files | running total |
|---|---|---|
| A₄ | 32 | 32 |
| B₄ | 22 | 54 |
| F₄ | 2 | 56 |
| H₄ | 13 | 69 |
The 69 total matches the count of .vZome files in the
output/<polytope>/ folders and the count of status=ok entries in
output/wythoff_sweep_manifest.json. Of these 69:
tools/run_wythoff_sweep.py).4φ + 2 ≈ 8.47 that lie outside the rng = 2
candidate set. See
B₄ rng = 4 finding.tools/run_wythoff_sweep.py rejected them because the
tesseract’s edge set fails to snap along those directions while
the bitruncated tesseract’s edge set succeeds. See
Step-1 blind-spot audit.Three of the 16 nominal B₄ Wythoff bitmasks describe the same
uniform 4-polytope as a smaller-bitmask F₄ Wythoff form (because the
24-cell ≡ rectified 16-cell is built into the embedding of F₄ inside
B₄). We file every shape under the F₄ canonical name and use only one
folder per polytope; the B₄ rows in the table above point to the F₄
folder via → <folder>/ and have no folder of their own.
| F₄ canonical name | F₄ bitmask | ≡ B₄ name | B₄ bitmask | folder |
|---|---|---|---|---|
| 24-cell | (1,0,0,0) | rectified 16-cell | (0,0,1,0) | output/regular/24cell/ |
| rectified 24-cell | (0,1,0,0) | cantellated 16-cell | (0,1,0,1) | output/uniform/rectified_24cell/ |
| truncated 24-cell | (1,1,0,0) | cantitruncated 16-cell | (0,1,1,1) | output/uniform/truncated_24cell/ |
The qfbox.info reference lists these three coincidences explicitly: “Rectified 16-cell: identical to the 24-cell”, “Cantellated 16-cell: identical to the rectified 24-cell”, “Cantitruncated 16-cell: identical to the truncated 24-cell; a uniform polychoron bounded by 24 truncated octahedra and 24 cubes” (qfbox.info/4d/uniform).
In the manifest, every entry that was originally found under the B₄
bitmask carries an alias_of field recording the original
(group, bitmask, source_polytope) triple, so the search-engine
provenance is preserved while the surface organization stays clean.
qfbox.info/4d/uniform lists 47
non-prismatic convex uniform 4-polytopes, plus the snub 24-cell and
grand antiprism (49 non-prismatic uniforms in total). After
deduplication via the Equivalences table,
this corpus contributes 34 of those 47 as .vZome files. The
13 missing are all in B₄ / F₄ and all snap-fail at rng=2 because
their projections live in ℤ[√2]³ rather than ℤ[φ]³:
| family | missing polytopes |
|---|---|
| tesseract / 16-cell (7) | truncated tesseract, cantellated tesseract, runcinated tesseract, runcitruncated tesseract, cantitruncated tesseract, omnitruncated tesseract, runcitruncated 16-cell |
| 24-cell (6) | bitruncated 24-cell, cantellated 24-cell, runcinated 24-cell, cantitruncated 24-cell, runcitruncated 24-cell, omnitruncated 24-cell |
Increasing rng will not rescue these — the field mismatch is a
structural feature of B₄/F₄ vertex coordinates, not a search-precision
artefact (see the
snap-failure analysis
below). The two non-Wythoffian uniforms (snub 24-cell, grand
antiprism) are out of scope for this Wythoff-only sweep but have
hand-curated projections at output/uniform/snub_24cell/ and
output/uniform/grand_antiprism/.
Step 1 of Method seeds Step 2 with the kernels found by
projecting only the regular member of each Coxeter group (bitmask
(1,0,0,0) or, for B₄, additionally (0,0,0,1)), under the assumption that
every Wythoff descendant is zomeable along exactly the same kernels.
A kernel-completeness audit was run after the rng = 2 sweep finished:
for each Wythoff descendant we ran an independent rng = 2 search
without the regular’s kernel restriction (tools/run_wythoff_sweep
--rng 2 --bitmask 0XXX, full 195 312 candidate directions), and
compared the resulting raw fingerprint count against the descendant’s
pipeline count in the tables above.
| group | descendants tested | descendants with gap | gap range | new shapes that snap to ℤ[φ]³ |
|---|---|---|---|---|
| A₄ | 8/8 | 0 | — | 0 |
| B₄ | 6/6 | 6 | +1 .. +4 | 6 (at rng = 4) ⁵ |
| F₄ | 2/2 | 0 | — | 0 |
| H₄ | 5/13 (600-cell, 120-cell, rectified 600-cell, rectified 120-cell, cantellated 600-cell) ⁴ | 0 | — | 0 |
⁴ Only five H₄ completeness probes finished within session-feasible wall time; the remaining eight H₄ cases (V ≥ 1200, E ≥ 3600 each) are pending — see the Coverage status subsection below.
⁵ At rng = 2 none of the B₄ gap-shapes snap to ℤ[φ]³. At rng = 4 the rectified tesseract and the truncated 16-cell each gain 3 new shapes that do snap and use only standard zometool struts — see B₄ rng = 4 finding below. The other four B₄ gap-bearing descendants reconcile to zero new shapes once master regulars and B₄ ↔ F₄ aliases are indexed.
The B₄ pattern is the only group with kernel-completeness gaps, and every descendant is affected. This is an artefact of B₄ having two non-isomorphic regulars (tesseract and 16-cell are dual but distinct polytopes with different edge orbits): kernels that align with 16-cell edges are not a subset of kernels that align with tesseract edges. A₄, F₄, and H₄ are all self-dual at the level of edge orbits (5-cell self-dual; 24-cell self-dual via the F₄ outer automorphism; 600-cell ↔ 120-cell are dual but the seeding bitmask is the cell-first 600-cell already), so descendants don’t gain new kernels relative to their single regular seed.
At rng = 2, none of the B₄ extra shapes snap to ℤ[φ]³. Every
gap shape’s coordinates contain √2 (the silver ratio), the same
field mismatch that produces the B₄/F₄ snap-failures already
documented in the
Wythoff-extension shape inventory.
So at rng = 2 the kernel-completeness gap is real but does not
translate into any additional .vZome files: the existing 60-file
corpus is complete at rng = 2 as a zometool-realisable set.
However, this is not the final word — at rng = 4 (i.e., when the
zomeable-direction generator is allowed to use larger ℤ[φ]-integer
coefficients) the rectified tesseract and truncated 16-cell each
gain 3 new zomeable shapes, see
B₄ rng = 4 finding.
The 16-cell case is a particularly clean validation: the regular-only
seed found 3 shapes; the full-rng=2 search found 6, exactly matching
master’s hand-curated 6 16-cell projections in output/regular/16cell/.
This confirms (i) the missing 3 shapes exist, (ii) the search engine
recovers them correctly when given the wider kernel set, and (iii) the
production pipeline could in principle be extended to the
multi-regular-seed scheme, but only at the cost of adding three
duplicate-of-existing-corpus B₄ records.
The reproducibility script is preserved as
ongoing_work/probes/probe_kernel_completeness.py.
The default sweep uses rng = 2, generating 195 312 candidate kernel
directions per Step-1 search and ~3 × that per Step-2 zomeable-check.
Targeted probes at rng = 3 (2 882 400 directions, ~15 × the rng = 2
candidates) and rng = 4 (21 523 360 directions, ~110 × rng = 2)
were run on representative cases.
Among the regulars, every rng = 3 and rng = 4 count matches master’s hand-curated count exactly after the full snap+dedup pipeline:
| polytope | V | rng = 2 raw | rng = 3 raw | rng = 4 raw | rng = 4 snap | rng = 4 unique | master |
|---|---|---|---|---|---|---|---|
| 5-cell | 5 | 4 | 4 | 4 | 4 | 4 | 4 |
| 16-cell | 8 | 6 ⁵ | 6 | 6 | 6 | 6 | 6 |
| 24-cell | 24 | 3 | 3 | 3 | 3 | 3 | 3 |
| tesseract | 16 | 32 ⁶ | 129 | 344 | 11 | 8 | 9 ⁶ |
⁵ At rng = 2 the full-search count is 6, matching master. The pipeline restricts to regular-only kernels and finds 3 (see kernel- completeness audit above).
⁶ The tesseract is an outlier with a documented infinite-family
behaviour: its rng = 1, 2, 3 raw fingerprint counts are 8, 32, 129
(matching the
output/regular/8cell/CLASSIFICATION.md
master theorem exactly), and at rng = 4 raw fingerprints jump to
(a, b) ∈ ℤ[φ]² with a² + b² ∈ ℤ[φ] (silver-
ratio coordinates rapidly fall out of ℚ(φ)). After Stage-B dedup
8 distinct vZome shapes remain. Master’s hand-curated 9 reflect a
stable subset chosen for graphic distinctness; the rng = 4 search
finds 8 of them in this single run, validating the search engine’s
behaviour on the most numerically demanding regular. No 5-cell /
16-cell / 24-cell shape was missed at any rng tested.Among the descendants the picture is more nuanced: V ≥ 60
polytopes pick up additional Stage-A fingerprint groups at rng = 3
(omnitruncated 5-cell goes 10 → 14, cantellated 16-cell ≡ rectified
24-cell goes 5 → 7, etc.). A targeted snap-test on four of these
cases — cantitruncated 5-cell, runcitruncated 5-cell, omnitruncated
5-cell, and cantellated 16-cell — fed each rng = 3 Stage-A
fingerprint through lib.emit_generic.project_and_emit and then
through the same Stage-B 3D-shape-congruence dedup
(tools/dedup_corpus_by_shape.shape_signature) used by the
production pipeline:
| polytope | rng = 2 corpus files | rng = 3 raw fp | rng = 3 snap pass | rng = 3 unique after Stage B | new vs corpus |
|---|---|---|---|---|---|
| cantitruncated 5-cell | 4 | 10 | 10 | 4 | 0 |
| runcitruncated 5-cell | 4 | 9 | 9 | 4 | 0 |
| omnitruncated 5-cell | 4 | 14 | 14 | 4 | 0 |
| cantellated 16-cell ⁵⁄ | 3 | 7 | 7 | 3 | 0 |
⁵⁄ Aliased to F₄ rectified 24-cell — the same uniform polytope, see the Equivalences table.
Every additional rng = 3 Stage-A fingerprint snapped to ℤ[φ]³ but
collapsed onto an existing rng = 2 shape under Stage-B dedup. The
extras are orbit-equivalent kernel directions (the polytope’s own
symmetry group) whose 3D projections are rigid-motion equivalent;
their fingerprints differ only because of the floating-point-precision
binning in shape_fingerprint. No new .vZome files emerged. The
full rng = 3 snap-test data is preserved at
ongoing_work/rng3_snaptest_summary.json
and
ongoing_work/rng3_dedup_summary.json,
the rng = 4 data at
ongoing_work/rng4_snaptest_summary.json
(probe scripts in ongoing_work/probes/).
The current corpus is therefore stable under rng = 3 (and the regulars under rng = 4) for every polytope where the snap+dedup test was feasible (5-cell V = 5, 16-cell V = 8, 24-cell V = 24, tesseract V = 16, cantitruncated/runcitruncated/omnitruncated 5-cell V = 60–120, cantellated 16-cell V = 96 — representative across A₄, B₄, F₄ at low to moderate V). The H₄ descendants (V ≥ 720, ~3600 edges each) make rng ≥ 3 search infeasible per case at ~hours of compute each on the present implementation; the 600-cell and 120-cell families’ single-shape claim at rng = 2 is consistent with their high symmetry but has not been independently re-verified at rng ≥ 3.
The 6 regular convex 4-polytopes plus the 2 non-Wythoffian
exemplars all show full master-shape recovery through the production
pipeline. This is the strongest end-to-end calibration of the
search + snap + dedup chain available, since these 8 polytopes have
been studied by zometool/4D enthusiasts for many years and their
projection inventories in output/<polytope>/ are independently
curated reference data.
| polytope | V | E | master shapes | search depth that confirmed full recovery | unique snapped | recovered | extra |
|---|---|---|---|---|---|---|---|
| 5-cell (A₄) | 5 | 10 | 4 | rng = 4 (Milestone 4) | 4 | 4 / 4 | 0 |
| tesseract (B₄) | 16 | 32 | 2 + ∞ inf-family | rng = 4 (Milestone 4) | 8 | family ✓ | 0 |
| 16-cell (B₄) | 8 | 24 | 6 | rng = 4 (Milestone 8) | 6 | 6 / 6 | 0 |
| 24-cell (F₄) | 24 | 96 | 3 | rng = 4 (Milestone 8) | 3 | 3 / 3 | 0 |
| 600-cell (H₄) | 120 | 720 | 1 | rng = 2 audit (gap = 0) | 1 | 1 / 1 | 0 |
| 120-cell (H₄) | 600 | 1200 | 1 | rng = 2 audit (gap = 0, this milestone) | 1 | 1 / 1 | 0 |
| snub 24-cell | 96 | 432 | 2 | rng = 2 (Milestone 5) | 2 | 2 / 2 | 0 |
| grand antiprism | 100 | 500 | 2 | rng = 2 (Milestone 5) | 2 | 2 / 2 | 0 |
The infinite cuboid family in the tesseract row is parameterised
(output/regular/8cell/*inf_family*.vZome); the search engine generates
the family correctly (8 unique snapped shapes at rng = 4 cover all
generic-cuboid projections), but the inf-family .vZome files
themselves use fractional coordinates that the
tools/dedup_corpus_by_shape.parse_vzome
re-loader does not currently parse, so they don’t participate in
the corpus index for direct hash matching — the family check is
“the search produced ≥ master count of generic cuboids”, not a
per-file hash match. All other rows are exact 1-to-1 hash matches.
The probe scripts and per-fingerprint records are preserved at
ongoing_work/probes/ and the
corresponding *.json summaries (e.g.,
ongoing_work/h4_120cell_rng2_audit.json
for the 120-cell, ongoing_work/non_wythoff_sanity_summary.json
for snub-24 / grand antiprism).
The snub 24-cell and grand antiprism are non-Wythoffian uniform
4-polytopes (no Coxeter–Dynkin bitmask generates them — the snub-24
has D₄ symmetry of index 6 in F₄, and the grand antiprism has the
ionic-diminished H₄ subgroup of order 400). This means
build_polytope(group, bitmask) cannot construct them, so they sit
outside the Wythoff sweep. But their vertex/edge data is available
via lib/polytopes.py (snub_24cell,
grand_antiprism), and they share the rest of the pipeline (the
rng = 2 search engine, the snap, and the Stage A / Stage B dedup) with
the Wythoff sweep. Master maintains a hand-curated set of 4 zomeable
projections for them (output/uniform/snub_24cell/*.vZome and
output/uniform/grand_antiprism/*.vZome, two per polytope) which have been
studied by zometool/4D enthusiasts for years; recovering this set
end-to-end through the same pipeline is the calibration captured
above. The exact recovery confirms that for these two non-
Wythoffian exemplars the shape census is not undercounted at
rng = 2 — every master shape is found, no master shape is missed,
and no extra zomeable shape exists in the rng = 2 search range.
This is consistent with the rng = 3 / rng = 4 stability proven
above for the regulars and the small Wythoff descendants.
The rng = 3 / rng = 4 snap+dedup probes above were focused on A₄ descendants and the regulars. Re-running the rng = 4 snap+dedup test on B₄ Wythoff descendants whose kernel-completeness audit showed a non-zero gap surfaces a new result. The audit had reported six gap-bearing B₄ descendants; rng = 4 snap+dedup probes resolve the gaps as follows:
| polytope | V | E | audit gap | rng = 4 raw fp | snapped | Stage-B unique | in corpus | new | resolution |
|---|---|---|---|---|---|---|---|---|---|
| 16-cell | 8 | 24 | 3 | 6 | 6 | 6 | 6 | 0 | audit miss-counted (master regulars not indexed) |
| rectified tesseract | 32 | 96 | 4 | 7 | 7 | 6 | 3 | 3 | real new shapes |
| 24-cell (B₄) | 24 | 96 | 1 | 3 | 3 | 3 | 3 | 0 | audit miss-counted (master regulars not indexed) |
| truncated 16-cell | 48 | 120 | 4 | 7 | 7 | 6 | 3 | 3 | real new shapes |
| cantellated 16-cell | 96 | 288 | 3 | 8 | 8 | 3 | 3 | 0 | alias of rectified 24-cell (shapes already in F₄ folder) |
| cantitruncated 16-cell | 192 | 384 | 4 | 12 | 12 | 3 | 3 | 0 | alias of truncated 24-cell (shapes already in F₄ folder) |
Two polytopes grow. The rectified tesseract and the truncated
16-cell each gain 3 zomeable shapes that the rng = 2 sweep does not
find. In both cases 7 raw fingerprints reduce to 6 unique shape
signatures (one fingerprint pair collides on the same canonical
shape), and the 3 existing rng = 2 shapes are recovered identically.
All new shapes use only the 4 standard zometool strut colours and
share the same 4-edge-length signature shape
((1.0, k), (1.47337, k), (1.61803, k), (1.7013, k))
where k = 24 for the rectified tesseract (V = 32) and k = 30 for
the truncated 16-cell (V = 48). Their snapped 4-D kernel directions
all involve coordinates of size 4φ + 2 ≈ 8.47 or 9φ + 2 ≈ 9.47
(ℤ[φ]-integers that lie outside the rng = 2 candidate set), which
explains why the rng = 2 sweep cannot reach them: the relevant kernel
directions simply don’t exist among rng = 2 candidates.
Four polytopes are clean — their audit gaps were artifacts.
Three of them (16-cell B₄, 24-cell B₄, cantellated 16-cell) were
flagged because the
probe_kernel_completeness.py
counter only inspected output/<polytope>/ and ignored
both the master regular folders (output/regular/16cell/, output/regular/24cell/)
and the cross-group alias folders (cantellated 16-cell ≡ rectified
24-cell, cantitruncated 16-cell ≡ truncated 24-cell, see
B₄ ↔ F₄ aliases). Once those
locations are added to the corpus index, every snapped rng = 4 fp
matches an existing shape; the gap collapses to zero. The
cantitruncated 16-cell rng = 4 probe (V = 192, E = 384) ran in
~24 h wall on this machine and confirmed the alias prediction
exactly: 12 raw fp → 3 Stage-B unique shapes, all already present
in the truncated 24-cell folder.
The 6 new shapes have been promoted to the main corpus at
output/uniform/rectified_8cell/rectified_8cell_oblique_00.vZome,
oblique_01.vZome,
oblique_02.vZome
and the analogous trio at
output/uniform/truncated_16cell/.
Each entry in output/wythoff_sweep_manifest.json
carries a "rng": 4 field marking it as a higher-rng discovery.
The promotion script
ongoing_work/probes/promote_b4_rng4.py
is reproducible: re-running it would (a) re-verify Stage-B
distinctness from corpus and within the candidate set, (b) drop
fp-index duplicates, (c) classify each kernel as oblique, and
(d) overwrite-refuse on idempotency.
The original per-fingerprint kernel-direction records and the raw
emitted .vZomes are also preserved at
ongoing_work/rectified_tesseract_rng4/,
ongoing_work/truncated_16cell_rng4/,
and the matching *.json files. The reconciliation evidence for
the four clean cases lives at
ongoing_work/cantellated_16cell_rng4.json,
ongoing_work/cantitruncated_16cell_rng4.json,
and
ongoing_work/b4_small_rng4_verify.json.
The probe scripts are in
ongoing_work/probes/.
This contradicts the earlier “corpus is stable at rng = 4” shorthand: the regulars (5-cell, 16-cell, 24-cell) are stable at rng = 4, the tesseract follows its known infinite cuboid family, A₄ Wythoff descendants are stable at rng = 3, but the rectified tesseract and truncated 16-cell each gain 3 zomeable shapes at rng = 4. Of the six audit-flagged B₄ descendants, the 3-new- shapes-per-polytope pattern is confirmed on 2 polytopes and ruled out on the other 4 (now including the cantitruncated 16-cell). All B₄ Wythoff descendants are therefore audited at rng = 4, with a final +6 zomeable shapes attributable to the B₄-without-master-indexing audit gap.
This subsection inventories which Wythoff descendants have been
audited at which rng, so it is explicit which corners of the
search space are proven complete vs. probable but unverified
vs. completely untouched. The verdict for each cell of the
table is one of:
.vZome files).lib.search_engine.search
is invoked directly on the descendant’s edge set, bypassing the
parent-regular Step-1 filter in tools/run_wythoff_sweep.py. See
Step-1 blind-spot audit.ongoing_work/ as new.| group | polytope (bitmask) | V | E | rng = 2 sweep | rng = 2 audit | rng = 2 d-d | rng = 4 probe | new shapes | notes |
|---|---|---|---|---|---|---|---|---|---|
| A₄ | 5-cell (1,0,0,0) | 5 | 10 | done | done | n/a (regular) | done | 0 | regular |
| A₄ | rectified 5-cell (0,1,0,0) | 10 | 30 | done | done | done | rng = 3 done | 0 | A₄ self-dual; 4 d-d fp all collapse via Stage-B to corpus |
| A₄ | truncated 5-cell (1,1,0,0) | 20 | 40 | done | done | done | done | 0 | |
| A₄ | bitruncated 5-cell (0,1,1,0) | 30 | 60 | done | done | done | rng = 3 done | 0 | |
| A₄ | cantellated 5-cell (1,0,1,0) | 30 | 60 | done | done | done | rng = 3 done | 0 | |
| A₄ | cantitruncated 5-cell (1,1,1,0) | 60 | 120 | done | done | done | done | 0 | 3 d-d new fp all collapse via Stage-B to corpus |
| A₄ | runcinated 5-cell (1,0,0,1) | 20 | 60 | done | done | done | done | 0 | |
| A₄ | runcitruncated 5-cell (1,1,0,1) | 60 | 150 | done | done | done | done | 0 | 2 d-d new fp all collapse via Stage-B to corpus |
| A₄ | omnitruncated 5-cell (1,1,1,1) | 120 | 240 | done | done | done | rng = 3 done | 0 | 6 d-d new fp all collapse via Stage-B to corpus |
| B₄ | 16-cell (0,0,0,1) | 8 | 24 | done | done | n/a (regular) | done | 0 | regular |
| B₄ | tesseract (1,0,0,0) | 16 | 32 | done | done | n/a (regular) | done | infinite | follows known cuboid family |
| B₄ | rectified tesseract (0,1,0,0) | 32 | 96 | done | done | done | done | +3 | new V=32 shapes (Milestone 6, rng=4) |
| B₄ | 24-cell-as-B₄ (0,0,1,0) | 24 | 96 | done | done | done | done | 0 | 24-cell sits here too; 3 d-d new fp = 3 master 24-cell projections |
| B₄ | truncated 16-cell (0,0,1,1) | 48 | 120 | done | done | done | done | +3 | new V=48 shapes (Milestone 7, rng=4) |
| B₄ | cantellated 16-cell (0,1,0,1) | 96 | 288 | done | done | done | done | 0 | F₄ alias of rectified 24-cell |
| B₄ | cantitruncated 16-cell (0,1,1,1) | 192 | 384 | done | done | done | done | 0 | F₄ alias of truncated 24-cell; 12 raw fp → 3 unique, all in corpus |
| B₄ | bitruncated tesseract (0,1,1,0) | 96 | 192 | done | done | done | not yet | +3 | new V=96 shapes (Milestone 17, rng=2 d-d): oblique_00, oblique_01, oblique_02 (originally misclassified face_first_hexagon, renamed in M17b); full 4-colour palette |
| B₄ | other tesseract / 16-cell records (truncated tesseract, runcinated tesseract, omnitruncated tesseract, runcitruncated 16-cell, runcitruncated tesseract, omnitruncated tesseract, …) | 64 .. 384 | 128 .. 768 | done | not yet | not yet | not yet | (likely 0) | not on the audit case list, since each is either an F₄ alias or a tesseract record covered by the cuboid-family argument |
| F₄ | 24-cell (1,0,0,0) | 24 | 96 | done | done | n/a (regular) | done | 0 | regular |
| F₄ | rectified 24-cell (0,1,0,0) | 96 | 288 | done | done | done | done¹ | 0 | ¹ tested via B₄ alias cantellated 16-cell; 4 d-d new fp all collapse via Stage-B to corpus |
| F₄ | truncated 24-cell (1,1,0,0) | 144 | 288 | done | done | done | done | 0 | 5 d-d new fp all collapse via Stage-B to corpus |
| F₄ | cantellated 24-cell (1,0,1,0) | 288 | 864 | done | done | done | partly | 0 | 1 d-d new fp snap-failed (ℤ[φ,√2]-flavoured kernel [0,0,1/φ²,−1/φ²]) |
| F₄ | runcinated 24-cell (1,0,0,1) | 192 | 576 | done | done | done | partly | 0 | 1 d-d new fp snap-failed (kernel [0,0,0,2φ] along F₄ axis; ℤ[φ,√2] field) |
| F₄ | runcitruncated 24-cell (1,1,0,1) | 576 | 1440 | done | done | done | partly | 0 | 1 d-d new fp snap-failed (kernel [0,0,0,2φ] along F₄ axis; ℤ[φ,√2] field) |
| F₄ | other F₄ Wythoff records | – | – | done | done | done | partly | 0 | most are B₄ aliases; remaining F₄ rng=4 probes covered transitively |
| H₄ | 600-cell (0,0,0,1) | 120 | 720 | done | done | not yet | running | 0 (rng=2) | rng=4 currently running (~19 h wall) |
| H₄ | rectified 600-cell (0,0,1,0) | 720 | 3600 | done | done | not yet | infeasible | 0 (rng=2) | rectified-600-cell ≡ rectified-120-cell up to H₄ outer auto |
| H₄ | rectified 120-cell (0,1,0,0) | 1200 | 3600 | done | done | not yet | infeasible | 0 (rng=2) | |
| H₄ | cantellated 600-cell (0,1,0,1) | 3600 | 10800 | done | done | not yet | infeasible | 0 (rng=2) | |
| H₄ | 120-cell (1,0,0,0) | 600 | 1200 | done | done | n/a (regular) | infeasible | 0 (rng=2) | full audit run; 1 raw fp = 1 unique shape = 1 master shape (120cell_H4_to_H3.vZome); rng=4 infeasible |
| H₄ | omnitruncated 120-cell (1,1,1,1) | 14400 | 28800 | done | done¹ | infeasible | infeasible | 0 (rng=2, parent-filtered) | ¹ rng=2 Step-2 sweep at parent-filter (120-cell) found 65 raw fingerprints → all 65 snap to 1 Stage-B-unique shape = the existing corpus cell_first_truncated_icosidodecahedron_00512c47fe.vZome master. Salvaged from a stuck 88-h shell by ongoing_work/probes/omnitr120_rng2_focused.py; results in ongoing_work/h4_omnitr120_rng2.json. Caveat: descendant-direct (blind-spot) audit not run — would require ~10⁵ kernel × 14400-vertex passes at rng=2, prohibitively slow; the 0-gap finding here only covers kernels that survive the 120-cell parent-snap filter. |
| H₄ | other 7 H₄ Wythoff records (truncated 600-cell, truncated 120-cell, cantellated 120-cell, bitruncated 120-cell, cantitruncated 600-cell, cantitruncated 120-cell, runcinated 120-cell, runcitruncated 600-cell / 120-cell) | 1440 .. 7200 | 4320 .. 14520 | done | untested | untested | infeasible | unknown | rng = 2 audit estimated >24 h CPU per case at the present lib/search_engine speed |
Summary of unfinished territory:
lib/search_engine.search (precomputing per-
direction strut hits is expected to give a 5-10× speedup) and/or
a duality argument that halves the audit table. All 5 audited
H₄ cases show 0 gap, and H₄’s self-dual structure (analogous to
A₄’s) suggests the remaining 8 will also have 0 gap, but this is
not empirically verified.wythoff.build_polytope, so extending them to
rng = 4 requires a separate explicit V/E loader. Not done.field = sqrt2, etc.) — outside
scope. The B₄ silver-ratio shapes that fail to snap in ℤ[φ]³
could in principle live in a vZome field = sqrt2 model, but
the project explicitly declined this direction.The “infeasible” labels in the H₄ rows are not fundamental limitations — they are limitations of the present search-engine implementation. A vectorised per-direction strut-hit cache or a GPU port would unlock several H₄ rows at rng = 2 and possibly the smaller H₄ rows at rng = 4.
tools/run_wythoff_sweep.py is a two-step pipeline. Step 1
(find_group_kernels, line ~108) enumerates every rng candidate
direction and snap-filters them against the parent regular
polytope only — i.e. against the (1, 0, 0, 0) bitmask form of each
group. Step 2 reuses that filtered kernel set for every Wythoff
descendant of the same group.
This architecture has a blind spot: if a kernel direction makes
the descendant’s edge set zomeable (passes _try_align) but the
parent regular’s edge set fails the same test, the kernel gets
dropped in Step 1 and Step 2 never sees it. Such kernels can exist
because the descendant has a different (richer) edge multiset than
the parent — a descendant edge set may admit a zomeable rotation that
the parent’s denser-or-sparser edge set rejects. This blind spot
was first identified empirically when the
B₄ rng = 4 finding
section traced its 6 new shapes to kernels that snap-fail on the
tesseract at every rng but classify on rectified tesseract / truncated
16-cell.
The Milestone 17 audit (ongoing_work/probes/blind_spot_audit.py)
re-ran the rng = 2 search directly against each descendant’s edge
set, bypassing Step 1 entirely. It covered every A₄/B₄/F₄
Wythoff descendant not already addressed by the rng = 4 audit
(25 polytopes total), reproduced every existing manifest fp_hash
(0 misses across all 25), and flagged 31 raw fingerprints whose hashes
were absent from the manifest. A follow-up triage
(ongoing_work/probes/blind_spot_triage.py) snapped each candidate
to ℤ[φ]³ and matched the resulting Stage-B signature against the
full corpus (output/ + master regulars):
| group | new fp | snap OK | snap FAIL | matched corpus (same poly) | matched corpus (other poly) | intra-audit dup | GENUINELY NEW |
|---|---|---|---|---|---|---|---|
| A₄ | 11 | 11 | 0 | 11 | 0 | 0 | 0 |
| B₄ | 8 | 8 | 0 | 0 | 3 | 2 | 3 |
| F₄ | 12 | 9 | 3 | 9 | 0 | 0 | 0 |
| total | 31 | 28 | 3 | 20 | 3 | 2 | 3 |
Interpretation by group:
cantitruncated_5-cell (3 fp), runcitruncated_5-cell (2 fp),
and omnitruncated_5-cell (6 fp) each have rng = 2 kernels where
the 5-cell parent fails to snap but the descendant succeeds. All
11 of these rng = 2 d-d kernels project to 3D shapes that are
Stage-B-equivalent to existing entries in the same polytope’s
folder under the production sweep. Self-duality of A₄ does not
prevent the blind spot — it only means descendants of A₄ are
particularly sparsely edged, which gives them more _try_align
freedom than the 5-cell parent has — but in the rng = 2 case every
recovered shape happens to coincide with one already in the corpus
via a different (parent-snapping) kernel orbit.24cell_triality.vZome, 24cell_long_root_rhombic_dodecahedron.vZome,
24cell_short_root_cuboctahedron.vZome in output/regular/24cell/); these
are aliases of pre-existing master shapes, not new shapes. Five
fp belong to B₄ (0,1,1,0) ≡ bitruncated tesseract; intra-audit
dedup collapses them to 3 distinct Stage-B shapes, all V = 96
with the full 4-colour zometool palette (B:48, Y:48, G:48, R:48),
and none match anything in the existing corpus. These 3 shapes
were promoted to
output/uniform/bitruncated_8cell/{oblique_00,oblique_01,
oblique_02}_*.vZome (the third was originally promoted as
face_first_hexagon_df821cc628.vZome, but the heuristic
classify_kernel in lib/polytope_features.py over-matched the
kernel direction [φ³, −φ, φ, φ] against the bitr_tesseract’s
hexagonal face normals; the projection’s two visible hexagons are
off-centre, so it is correctly oblique — renamed to
oblique_01_df821cc628.vZome in M17b/c, the M17c renumbering also
shuffled the suffixes so oblique_00/01/02 align row-for-row with
the antiprism inheritance table below). All 3 have rng=2 flag
and a Milestone 17 note in manifest.json.cantellated 24-cell with kernel [0, 0, 1/φ², −1/φ²] and two in
runcinated / runcitruncated 24-cell with the same [0, 0, 0, 2φ]
along-axis kernel. (Earlier drafts wrote 2φ²; the actual
numeric is 3.236… = 2φ, not 5.236… = 2φ².) These 3 are
recorded as snap_fail in the triage JSON; they pass
lib.search_engine.search (so _try_align accepts them) but
lib.emit_generic._snap_coords cannot find a scale that brings
them into ℤ[φ]³. An empirical field check
(ongoing_work/probes/f4_snap_fail_field_check.py, M18) confirmed
that each projected vertex coordinate lies in ℤ[φ, √2] — most
in the pure ℤ[√2] sub-field — with edge lengths squared in the
perfectly clean rational set {0, 1, 3/2, 2}. The same kernels
applied to the corresponding B₄ snap-fails (runcinated /
cantellated tesseract, runcitruncated 16-cell) give the same
ℤ[φ, √2] field, so the obstruction is intrinsic to the kernel
direction, not to the host group. The shapes are geometrically
perfectly regular projections; they simply cannot be encoded in
vZome’s φ-only field (analogous to the 13 ℤ[√2]³ snap-failures
already documented in Coverage vs. qfbox.info).The 3 promoted bitruncated tesseract shapes were initially flagged
as exceptions to the oblique-kernel inheritance pattern
that earlier milestones described, because their kernels are not
inherited from any tesseract master projection (the parent
tesseract’s edge set fails to snap along these directions). A
follow-up inheritance probe (ongoing_work/probes/test_antiprism_hypothesis.py)
resolved this completely: all 3 M17 bitr_tesseract shapes are
projections of the bitr_tesseract along the 3 documented 16-cell
antiprism master kernels (B6R12Y6, R12B6Y6, Y6R12B6), and each
also coincides with the 24-cell triality master projection:
| 16-cell antiprism master kernel (axis-aligned B4 basis) | 16-cell V=8 image | 24-cell V=24 image | bitr_tesseract V=96 image (M17 file) |
|---|---|---|---|
[2φ³, 2, −2φ³, 2φ³] (B6R12Y6_master_A) |
16cell_antiprism_B6R12Y6.vZome |
24cell_triality.vZome |
oblique_00_26f0b7c6e6.vZome |
[1/φ², 3φ−4, −1/φ², 1/φ²] (R12B6Y6_master_B; 3φ−4 ≈ 0.854) |
16cell_antiprism_R12B6Y6.vZome |
24cell_triality.vZome |
oblique_01_df821cc628.vZome |
[1+2φ³, √5, √5, −√5] (Y6R12B6_master_C) |
16cell_antiprism_Y6R12B6.vZome |
24cell_triality.vZome |
oblique_02_80a961d9da.vZome |
So the corpus has 0 documented exceptions to the inheritance
pattern — all 28 oblique_*.vZome files in output/,
including the 3 M17 additions, project from a parent regular master
kernel.
The reason the M17 audit reported a 3rd “descendant-only” kernel
[φ³, −φ, φ, φ] (the kernel originally assigned to
oblique_01_df821cc628.vZome, originally promoted as
face_first_hexagon_df821cc628.vZome and renamed in M17b /
renumbered in M17c) is
subtler: that kernel and
the R12B6Y6 antiprism master kernel [1/φ², 3φ−4, −1/φ², 1/φ²]
project the bitr_tesseract to the same 3D V=96 shape (Stage-B
sig df821cc628…) even though they are not B4-Weyl-equivalent
4-vectors (different magnitude multisets, so no permutation +
sign-flip can relate them). The bitr_tesseract has more symmetry
than its B4 Weyl group: its Coxeter symbol 0–1–1–0 is palindromic
under the B4 outer Z/2 automorphism (the duality flip that swaps
the bitmask (1,0,0,0) ≡ tesseract orbit with (0,0,0,1) ≡
16-cell orbit), so the bitr_tesseract is invariant under B4 × Z/2.
The Z/2 factor maps [φ³, −φ, φ, φ] to a 4-vector parallel to the
R12B6Y6 antiprism kernel, giving the same 3D image — but the
16-cell (only B4-invariant) sees only one of the two pre-images,
which is why the [φ³, −φ, φ, φ] kernel snap-fails on the 16-cell
while its Z/2-image snap-OK’s.
For the other two M17 kernels (oblique_00 [1/φ², 1/φ², −1/φ², −φ]
and oblique_01 [1/φ², −φ, φ, φ]), both the M17 kernel AND the
corresponding antiprism master kernel make the 16-cell zomeable
(and they project to the same V=8 antiprism shape), so the
inheritance is direct via plain B4-Weyl orbit equivalence.
Bonus connection: the oblique_00 4-tuple
[1/φ², 1/φ², −1/φ², −φ] is, under our convention disclaimer,
simultaneously the A4 5-ball master kernel (vertex-first 5-cell
basis), the F4 triality master kernel (D₄ long-root 24-cell basis),
the B4 16-cell antiprism Y6R12B6 direction (B4 axis-aligned basis),
and the B4 24-cell triality direction — the same 4 numbers
reading as four different geometric meanings across the four
groups’ conventions.
Reproducer:
python ongoing_work/probes/blind_spot_audit.py # 25 polytopes, ~9h on the present engine
python ongoing_work/probes/blind_spot_triage.py # snap+classify the 31 fp
python ongoing_work/probes/promote_blind_spot_m17.py
python ongoing_work/probes/m17_inheritance_full.py # final inheritance check vs all B4 parents
Outputs:
ongoing_work/blind_spot_audit_rng2.json — per-polytope new-fp
records.ongoing_work/blind_spot_triage.json — per-fp snap, Stage-B,
classification, staged path.ongoing_work/blind_spot_candidates/<group>_<bitmask>_<slug>/fp_<hash>.vZome
— staged .vZome files for each successfully snapped candidate.While auditing the snap-fail entries that survived from M17 into the
manifest (output/wythoff_sweep_manifest.json), three kernel
directions appear repeatedly across both B₄ and F₄ snap-fail
records:
| Kernel (Cholesky frame) | Decimal | Polytopes that hit it (B₄ + F₄) |
|---|---|---|
[√5, −√5, −√5, √5] |
[2.236, −2.236, −2.236, 2.236] |
7 B₄ (runcinated/cantellated/runcitruncated tesseract, runcitruncated 16-cell, …) |
[0, 0, 0, 2φ] |
[0, 0, 0, 3.236] |
7 B₄ + 4 F₄ (bitruncated/cantellated/cantitruncated 24-cell, etc.) |
[0, 0, 1/φ², −1/φ²] |
[0, 0, 0.382, −0.382] |
7 B₄ + 2 F₄ |
(Earlier drafts of this document referred to one of these as
[0, 0, 0, 2φ²] ≈ [0, 0, 0, 5.236]. That was a transcription
error: the actual numeric in the manifest, 3.236…, is 2φ, not
2φ².)
ongoing_work/probes/f4_snap_fail_field_check.py projects each
host polytope along its snap-fail kernel and reports the algebraic
field of the resulting 3D vertex coordinates. After rescaling so
the smallest non-zero edge length squared is 1, every test case
has the same picture:
Vertex coordinates lie in ℤ[φ, √2], with the majority in the
pure ℤ[√2] sub-field (i.e. (p + q √2)/d for small integers
p, q, d, with no φ component). Examples (runcinated 24-cell,
scale = 1):
±√2/2 = (0 ± 1·√2)/2
±(1 + √2/2) = (±2 ± √2)/2
±(1 + √2)
The same kernels applied to the corresponding B₄ snap-fails
(runcinated tesseract, cantellated tesseract,
runcitruncated 16-cell) give projections in the same
ℤ[φ, √2] field, so the obstruction is intrinsic to the kernel
direction (the 4D rotation maps a basis vector into a √2-flavoured
combination), not to the host group.
Conclusion. These projections are not numerically degenerate: they are clean, regular shapes with the simplest possible rational edge structure. They simply live in ℤ[φ, √2]³ rather than ℤ[φ]³, and vZome’s golden-zome field can only encode ℤ[φ]³. They join the 13 ℤ[√2]³ snap-failures already documented in Coverage vs. qfbox.info; the same algebraic obstruction blocks all of them.
Reproducer: python ongoing_work/probes/f4_snap_fail_field_check.py.
Cached output: ongoing_work/f4_snap_fail_field_results.txt.
runcinated 24-cell shares fp with cantellated tesseractWhen triaging the snap-fail kernels, runcinated 24-cell looked
“missing” from output/wythoff_sweep_manifest.json even though
it is a Wythoff descendant present in ongoing_work/shapes_rng2.jsonl
with fp 4507a6810f54bab2. The cause is in
tools/emit_novel.py:80–85 (pick_representative): when the same
fp_hash occurs in two polytopes, the occurrence with the smallest
(n_balls, group, bitmask, shape_idx) tuple wins. For
4507a6810f54bab2 (n_balls = 84), B₄ wins on the alphabetic group
tiebreak, so the manifest records the shape under
source_polytope = "cantellated tesseract" and the fact that
runcinated 24-cell produces the same shape is silently dropped.
Cross-polytope same-fp records found in novel_rng2.json (7 total):
| fp_hash | n_balls | shared between |
|---|---|---|
4507a6810f54bab2 |
84 | B₄ cantellated tesseract ↔ F₄ runcinated 24-cell (snap-fail, ℤ[φ,√2]) |
aabdf1c8acc26695 |
312 | B₄ omnitruncated tesseract ↔ F₄ runcitruncated 24-cell (snap-fail, ℤ[φ,√2]) |
5452b207028c5f19 |
60 | B₄ cantellated 16-cell ↔ F₄ rectified 24-cell (B₄ ≡ F₄ same polytope) |
cca712ec6eab236a |
52 | B₄ cantellated 16-cell ↔ F₄ rectified 24-cell (B₄ ≡ F₄ same polytope) |
598a9495c4a6ee08 |
120 | B₄ cantitruncated 16-cell ↔ F₄ truncated 24-cell (B₄ ≡ F₄ same polytope) |
5f5834c16179095e |
96 | B₄ cantitruncated 16-cell ↔ F₄ truncated 24-cell (B₄ ≡ F₄ same polytope) |
f79e915e332666e1 |
156 | B₄ runcitruncated tesseract ↔ F₄ cantellated 24-cell (snap-fail, ℤ[φ,√2]) |
5 of the 7 are the well-known B₄ ≡ F₄ same-polytope identifications
(rectified-24-cell ≡ cantellated-16-cell, truncated-24-cell ≡
cantitruncated-16-cell). The remaining 2 are genuine cross-polytope
shape collisions, all in the snap-fail set above — geometrically
distinct 4-polytopes whose along-[0,0,0,2φ]-style projections
collapse to the same 3D shape because the kernel direction kills
the polytope-distinguishing 4D structure. No data is lost; the
manifest entry is simply attributed to the alphabetically-first
group. (Low priority cleanup: emit a per-occurrence aliases list
into the manifest so cross-polytope fp_hashes are searchable from
either polytope.)
tools/analyze_sweep.py labels each sweep
shape against the regular reference set built from the (1,0,0,0)/
(0,0,0,1) records of every group (the 6 standard regular convex
4-polytopes). At rng=2 none of the 47 descendant Wythoff polytopes
project to a 3D shape that coincides with any of the 6 regulars’ own
projections — descendants always have many more vertices than their
regular, so their projection inventories are strictly disjoint at this
kernel range. In codebase terms, every sweep shape is “novel”; in the
stronger structural sense every sweep
shape is still a natural Wythoff extension of a regular’s projection
(Lemma A).
| group | natural-extension shape entries | distinct fp_hashes |
|---|---|---|
| A₄ | 43 | 43 |
| B₄ | 34 | 34 |
| F₄ | 20 | 20 |
| H₄ | 145 | 145 |
| TOTAL | 242 sweep-shape entries | 235 distinct fp_hashes |
The seven near-duplicates are cases where two different Wythoff polytopes project to the same 3D shape under different kernels (the hash is identical because shape fingerprints are rotation- and uniform-scale-invariant).
tools/emit_novel.py iterates the
natural-extension inventory (legacy filename: novel_rng<N>.json) and
calls lib.emit_generic.project_and_emit to write a .vZome file per
shape, then runs a 3D shape-congruence dedup pass to collapse any
congruent shapes the search engine reported as distinct. At rng=2 the
result is 60 of 235 distinct natural-extension shapes
successfully snapped to ZZ[φ]³ and confirmed pairwise non-congruent
(the zometool-realisable subset):
| group | sweep entries | snapped to vZome | distinct after shape-dedup |
|---|---|---|---|
| A₄ | 43 | 43 | 32 |
| B₄ | 34 | 13 | 13 |
| F₄ | 20 | 7 | 2 |
| H₄ | 145 | 145 | 13 |
A follow-up rng = 4 audit on the six B₄ Wythoff descendants whose
kernel-completeness probe reported a non-zero gap added 6 more
distinct shapes to the corpus (3 V = 32 from rectified tesseract,
3 V = 48 from truncated 16-cell), all under oblique classification
because their snapped kernels lie outside the rng = 2 candidate set
(typical kernel coordinates: 4φ + 2 ≈ 8.47, 9φ + 2 ≈ 9.47 — see
B₄ rng = 4 finding
below). These are flagged with "rng": 4 in
output/wythoff_sweep_manifest.json.
A second follow-up — the rng = 2 descendant-direct (Step-1 blind-
spot) audit on every A₄/B₄/F₄ Wythoff descendant — added 3 more
distinct shapes to the corpus, all from B₄ bitruncated tesseract
(output/uniform/bitruncated_8cell/{oblique_00,oblique_01,
oblique_02}_*.vZome). Each is V = 96 with the full 4-colour
zometool palette (B:48, Y:48, G:48, R:48). The blind spot is
architectural: tools/run_wythoff_sweep.py Step 1 snap-filters
against the parent regular polytope only, so any kernel where the
descendant’s edge set classifies but the parent’s does not gets
dropped before Step 2 ever runs. See
Step-1 blind-spot audit
below.
The shape-dedup totals above therefore become 69 distinct shapes after dedup when rng = 2 production + rng = 4 B₄ audit + rng = 2 descendant-direct audit findings are all counted.
The 27 snap failures are all in B₄/F₄ (cubic / 24-cellic geometry).
The truncated tesseract under kernel (0, 0, 0, 2 + 2φ) for example
projects to a 3D shape whose ball coordinates are (±1, ±1, ±(1+√2))
— the silver ratio, in ℤ[√2]³, not ℤ[φ]³. Every edge still
lies on a default zometool axis (the (√2, 0, 0) edges line up with
the Blue/X axis), so the projection passes the search-engine
zomeability test, but vZome cannot represent ball positions outside
Q(φ) = Q(√5) since √2 ∉ Q(√5). This is a structural feature of
B₄/F₄ projections, not a search-precision artefact. All the
icosahedral H₄ shapes snap because the icosahedral basis is naturally
aligned with the golden-ratio integer lattice.
vZome files are in output/,
organised into one subfolder per polytope common name (e.g.
output/uniform/cantellated_120cell/,
output/uniform/omnitruncated_120cell/). Each file is named
after the kernel direction’s classification against that polytope’s
features:
| basename pattern | meaning |
|---|---|
vertex_first_<hash>.vZome |
kernel parallel to a vertex direction |
cell_first_<celltype>[_<idx>]_<hash> |
kernel through a 3-cell centroid (e.g. cell_first_truncated_octahedron_03_…) |
face_first_<polygon>[_<idx>]_<hash> |
kernel through a 2-face centroid (e.g. face_first_hexagon_…) |
edge_first[_<idx>]_<hash>.vZome |
kernel through an edge midpoint |
oblique[_<idx>]_<hash>.vZome |
none of the above |
<idx> is appended only when multiple shapes within the same source
polytope classify identically (assigned alphabetically by <hash> for
stability). <hash> is the first 10 hex digits of the shape’s
fp_hash. The classifier lives in lib/polytope_features.py; it
extracts cell/face features via two nested scipy.spatial.ConvexHull
calls and returns the highest-priority match (vertex_first >
cell_first > face_first > edge_first > oblique).
A JSON manifest at output/wythoff_sweep_manifest.json cross-
references every file with its fp_hash, source polytope, kernel
direction, label, label-subtype, alias hashes (see “Direction-dedup”
below), and per-axis strut counts.
All emitted files use only the 4 standard zometool strut colours
(R = red, B = blue, Y = yellow, G = green) — non-standard directions
are rejected upstream in lib.search_engine._classify_dir. Coords
are post-snap normalised by a uniform φᵏ factor (lib/emit_generic
._normalize_scale) so the smallest pairwise vertex distance lands
near 30/φ² zome units (~11.5) across the whole corpus, giving a
consistent strut/ball ratio regardless of which s happened to snap.
oblique_*.vZome is a parent’s master projectionThe 25 oblique_*.vZome files in output/ come from
only 7 distinct kernel directions across the 4 groups, and each
of those 7 directions is exactly the kernel of one of the
hand-curated parent-regular master projection files in
output/{5cell,16cell,24cell}/.
Verified empirically by ongoing_work/probes/_verify_obliques_inherited.py:
for each unique oblique kernel n in the manifest, project the
parent regular along n via lib.emit_generic.project_and_emit and
compute the Stage-B fingerprint (rotation+uniform-scale invariant,
so basis differences between lib/emit_vzome.py and lib/emit_generic.py
do not interfere); the result matches a master file’s Stage-B
fingerprint in every case.
Convention disclaimer. A kernel vector is a 4D direction expressed in a particular coordinate system, so the specific 4-tuples below are meaningful only relative to this repository’s canonical embeddings:
- A4: vertex-first 5-cell with
v₅ = (0, 0, 0, 4√5)and the opposite tetrahedral cell atx₄ = −√5(lib/polytopes.py:fivecell()).- B4: axis-aligned tesseract
(±1)⁴(lib/polytopes.py:tesseract()).- F4: D₄ long-root 24-cell, all permutations of
(±1, ±1, 0, 0)(lib/polytopes.py:cell24_long_root()).- H4: standard 600-cell / 120-cell embedding from
lib/polytopes.py:cell600()/cell120().The sweep’s per-group bases are produced by Cholesky-frame simple roots Procrustes-aligned to these targets (
lib/wythoff.py:104-165), so descendant kernels always live in the same basis as their parent regular. Choosing different (but still SO(4)-equivalent) embeddings for any of these would rotate every kernel below by the corresponding O(4) matrix. Statements that compare kernels across groups (notably the A4↔F4 coincidence in the F4 subsection) are therefore convention- dependent rather than coordinate-free geometric assertions.
kernel n (4D, sweep basis) |
parent master file | descendant oblique_*.vZome files |
|---|---|---|
[1/φ², 1/φ², −1/φ², −φ] = [0.382, 0.382, −0.382, −1.618] |
output/regular/5cell/5cell_5ball_R6Y1B3.vZome (V=5) |
bitruncated_5-cell/oblique_01_9e2d21191f, cantellated_5-cell/oblique_00_1c346b0e6a, cantitruncated_5-cell/oblique_00_6189be49ed, omnitruncated_5-cell/oblique_9f354d53ae, rectified_5-cell/oblique_00_4fe32006bf, runcinated_5-cell/oblique_00_1c278a1d7d, runcitruncated_5-cell/oblique_db7b74a33d, truncated_5-cell/oblique_01_9fee63fe24 (8 files) |
[φ³−1, 0, φ³−1, 0] = [3.618, 0, 3.618, 0] |
output/regular/5cell/5cell_4ball_Y6B3.vZome (V=4, edge-collapses) |
cantellated_5-cell/oblique_01_e7d00a6967, cantitruncated_5-cell/oblique_01_befb910d19, rectified_5-cell/oblique_01_55abe81889, truncated_5-cell/oblique_02_cd65b76a7d (4 files) |
[0, 0, 1/φ², −1/φ²] = [0, 0, 0.382, −0.382] |
output/regular/5cell/5cell_5ball_Y4B2R4.vZome (V=5) |
bitruncated_5-cell/oblique_00_190a7b1c71, cantitruncated_5-cell/oblique_02_c1635833bd, rectified_5-cell/oblique_02_c2ff97e546, runcinated_5-cell/oblique_01_4547c11713, truncated_5-cell/oblique_00_7df137dba4 (5 files) |
The 4th master file 5cell_vertex_first_tet_plus_center.vZome is
not an oblique direction (its kernel (1−2φ, 1−2φ, 1−2φ, 1) = (−√5, −√5, −√5, 1)
points to a 5-cell vertex), so it has no descendant oblique_*.vZome
counterpart; instead it shows up under the descendants’ vertex_first_* /
cell_first_* labels.
These 3 kernels are reachable only at rng = 4 in the parent-filtered Step-1 search (the 6 rectified-tesseract / truncated-16-cell descendant files were promoted in milestones 13–14 from the rng=4 audit). Crucially, projecting the tesseract along any of these 3 kernels snap-fails — they are 16-cell- aligned only. The bitruncated tesseract additionally inherits all 3 directions at rng = 2 (Milestone 17b), giving 3 more descendant files (one per row).
kernel n (4D, sweep basis) |
parent master file | descendant oblique_*.vZome files |
|---|---|---|
[8.472, 2, −8.472, 8.472] |
output/regular/16cell/16cell_antiprism_B6R12Y6.vZome |
rectified_tesseract/oblique_00_e2b79a96f7, truncated_16-cell/oblique_00_9d18eb2806, bitruncated_tesseract/oblique_00_26f0b7c6e6 (3 files; M17 bitr-tesseract inherits via B4 Weyl orbit at rng=2 rep [1/φ², −φ, φ, φ]) |
[0.382, 0.854, −0.382, 0.382] (0.854 = 3φ−4 ∈ ℤ[φ]) |
output/regular/16cell/16cell_antiprism_R12B6Y6.vZome |
rectified_tesseract/oblique_01_b35b865a54, truncated_16-cell/oblique_01_ccdfd208c9, bitruncated_tesseract/oblique_01_df821cc628 (3 files; the M17 bitr-tesseract inherits via the bitr-tesseract’s B4 × Z/2 outer aut, see M17 audit) |
[9.472, 2.236, 2.236, −2.236] = [1+2φ³, √5, √5, −√5] |
output/regular/16cell/16cell_antiprism_Y6R12B6.vZome |
rectified_tesseract/oblique_02_2be6954c03, truncated_16-cell/oblique_02_2c50f047a8, bitruncated_tesseract/oblique_02_80a961d9da (3 files; M17 bitr-tesseract inherits via B4 Weyl orbit at rng=2 rep [1/φ², 1/φ², −1/φ², −φ]) |
The three antiprism master projections form the well-known D₄
triality orbit of the 16-cell (see output/regular/16cell/RESULTS.md).
Empirically all three triality directions propagate identically to
both rectified_tesseract and truncated_16-cell — these two
descendants share the same oblique-projection structure even though
their full descendant-vertex sets differ.
The M14/M17 milestones promoted the antiprism trio to 3 of the 12
non-regular B4 Wythoff polytopes (rectified_tesseract, truncated_16-cell,
bitruncated_tesseract). An exhaustive M17d audit
(ongoing_work/probes/m17d_antiprism_full_b4_audit.py) projected
every non-regular B4 Wythoff descendant along the 3 antiprism master
kernels, with the following clean dichotomy:
| bitmask | polytope | snap result |
|---|---|---|
| (0,1,0,0) | rectified tesseract | 3 distinct shapes (M14 trio above) |
| (0,0,1,1) | truncated 16-cell | 3 distinct shapes (M14 trio above) |
| (0,1,1,0) | bitruncated tesseract | 3 distinct shapes (M17 trio above) |
| (0,1,0,1) | rect 24-cell ≡ cantellated 16-cell | 1 shape (rectified_24-cell/oblique_56093e7cd7); F4 triality collapses 3 → 1 |
| (0,1,1,1) | trunc 24-cell ≡ cantitr 16-cell | 1 shape (truncated_24-cell/oblique_6b7eadcbcf); F4 triality collapses 3 → 1 |
| (1,1,0,0) | truncated tesseract | snap-fail (all 3 kernels) |
| (1,0,1,0) | cantellated tesseract | snap-fail (all 3 kernels) |
| (1,0,0,1) | runcinated tesseract | snap-fail (all 3 kernels) |
| (1,1,1,0) | cantitruncated tesseract | snap-fail (all 3 kernels) |
| (1,1,0,1) | runcitruncated tesseract | snap-fail (all 3 kernels) |
| (1,0,1,1) | runcitruncated 16-cell | snap-fail (all 3 kernels) |
| (1,1,1,1) | omnitruncated tesseract | snap-fail (all 3 kernels) |
The dichotomy is by bitmask[0]:
bitmask[0] = 0 (no tesseract orbit in the vertex set): the
3 antiprism kernels snap. This holds for all 5 such polytopes.
Of these, the 3 with pure B4 (or B4 × Z/2) symmetry produce 3
distinct V = 32 / 48 / 96 shapes; the 2 F4-symmetric polytopes
((0,1,0,1) and (0,1,1,1), with full F4 = B4 ⋊ S₃ symmetry)
collapse the 3 kernels to a single shape under triality.
bitmask[0] = 1 (tesseract orbit present): snap-fails on all
3 antiprism kernels. This is the same snap-failure as the bare
tesseract — the tesseract sub-orbit’s image is non-zomeable along
any antiprism direction (the antiprism kernels are 16-cell-aligned
only), so any descendant containing the tesseract orbit inherits
the snap-failure. This holds for all 7 such polytopes.
Conclusion: the corpus is complete with respect to the antiprism
trio. No further B4 polytope hides a missed antiprism-inherited
shape; every snap-able combination is already promoted. The 2
F4-symmetric singletons (oblique_56093e7cd7, oblique_6b7eadcbcf)
are documented under the F4 triality master in the next subsection;
they are simultaneously B4-antiprism-inherited and F4-triality-
inherited (the F4 triality master kernel [1/φ², 1/φ², −1/φ², −φ]
is exactly the B4 antiprism Y6R12B6 master kernel up to a B4 Weyl
rotation, again under the basis conventions noted at the top of
this section).
kernel n (4D, sweep basis) |
parent master file | descendant oblique_*.vZome files |
|---|---|---|
[1/φ², 1/φ², −1/φ², −φ] = [0.382, 0.382, −0.382, −1.618] |
output/regular/24cell/24cell_triality.vZome (V=24) |
rectified_24-cell/oblique_56093e7cd7, truncated_24-cell/oblique_6b7eadcbcf (2 files) |
Under the conventions noted at the top of this section, the same
4D 4-tuple [1/φ², 1/φ², −1/φ², −φ] is simultaneously the A4 5-ball
master kernel (in the vertex-first 5-cell basis) and the F4 triality
master kernel (in the D₄ long-root 24-cell basis). Concretely,
this 4-tuple — read in the A4 basis — projects the 5 vertices of the
5-cell to the 5ball_R6Y1B3 shape, and the very same 4-tuple — read
in the F4 basis — projects the 24 vertices of the 24-cell to the
24cell_triality shape.
This coincidence is basis-dependent. The two canonical
embeddings (vertex-first 5-cell, D₄ long-root 24-cell) were chosen
independently in lib/polytopes.py, with no cross-group anchoring
step that would force a particular kernel 4-tuple to mean “triality”
in both. Re-Procrustes-aligning either group to a different (but
SO(4)-equivalent) representative of its canonical embedding would
rotate one of the kernels by an O(4) matrix and the numerical
equality would disappear. So the observation is a real numerical
fact about the embeddings shipped in this repo, not a coordinate-
free statement that “the 5-cell sits inside the 24-cell along this
axis”. (Comparisons within a group — kernel ↔ parent ↔
descendants — are coordinate-free, since descendants and parent
share the group’s basis by construction.)
No oblique_*.vZome entries exist for any H4 descendant. The two
hand-built H4 master files (120cell_H4_to_H3.vZome,
600cell_H4_to_H3.vZome) classify under vertex_first / similar
icosahedral-axis labels in their respective descendant trees, not
under oblique.
The wythoff sweep, when restricted to its parent-filtered Step-1 kernel set, does not discover any new oblique projection directions beyond what the parent-regular master files already exhibit. Every “obscure” oblique projection of every Wythoff descendant that the production pipeline has emitted is just the descendant’s vertex set viewed through one of the parent regular’s documented oblique axes. This is consistent with the way the production sweep is architected (Step 1 finds kernels for the parent regular only; Step 2 reuses the same kernel set for every descendant), but it is a property of the parent-filtered kernel set, not a property of the descendant polytopes.
The
Step-1 blind-spot audit (Milestone 17)
ran the rng = 2 search directly against each descendant’s edge set
and added 3 new V=96 shapes from B₄ bitruncated tesseract. A
follow-up inheritance check
(ongoing_work/probes/test_antiprism_hypothesis.py) showed that
all 3 of these new shapes are still inherited — they are the
bitr_tesseract images of the 3 documented 16-cell antiprism master
kernels (B6R12Y6, R12B6Y6, Y6R12B6), and equivalently the
bitr_tesseract images of the 24-cell triality master kernel under
B₄ Weyl orbit equivalence:
| 16-cell antiprism master kernel | 16-cell V=8 image | 24-cell V=24 image | bitr_tesseract V=96 (M17) image |
|---|---|---|---|
[2φ³, 2, −2φ³, 2φ³] |
16cell_antiprism_B6R12Y6.vZome |
24cell_triality.vZome |
oblique_00_26f0b7c6e6.vZome |
[1/φ², 3φ−4, −1/φ², 1/φ²] (3φ−4 ≈ 0.854) |
16cell_antiprism_R12B6Y6.vZome |
24cell_triality.vZome |
oblique_01_df821cc628.vZome (renamed from face_first_hexagon_df821cc628.vZome in M17b, renumbered from oblique_02_df821cc628 in M17c) |
[1+2φ³, √5, √5, −√5] |
16cell_antiprism_Y6R12B6.vZome |
24cell_triality.vZome |
oblique_02_80a961d9da.vZome |
So the corpus has 0 documented exceptions to the inheritance
pattern. All 28 oblique_*.vZome files in output/
are inherited from a parent regular master kernel. The 7 master
kernels listed in the oblique kernel inheritance table below
account for 25 of 28 files; the remaining 3 (the M17
bitr_tesseract additions) are inherited from the same 3 16-cell
antiprism master kernels but at the V=96 bitr_tesseract scale,
extending each antiprism row in the table to 3 inheriting files
instead of 2. Reproducer:
python ongoing_work/probes/_verify_obliques_inherited.py plus
python ongoing_work/probes/test_antiprism_hypothesis.py.
# Per-group sweep. H4 step 1 takes ~16 min vectorised (~48 min on the
# pre-vectorisation engine); other groups finish in seconds.
python tools/run_wythoff_sweep.py --group A4 --rng 2
python tools/run_wythoff_sweep.py --group B4 --rng 2
python tools/run_wythoff_sweep.py --group F4 --rng 2
python tools/run_wythoff_sweep.py --group H4 --rng 2 --sort-by-size
# For the 4 large H4 polytopes (V=7200..14400) it is faster to drive
# them in parallel via --bitmask, one process per polytope:
python tools/run_wythoff_sweep.py --group H4 --rng 2 --bitmask 1101 `
--log ongoing_work/sweep_log_rng2_H4_1101.txt `
--shapes-jsonl ongoing_work/shapes_rng2_H4_1101.jsonl
# (...repeat for 1110, 0111, 1111, then concatenate the per-job
# JSONLs into ongoing_work/shapes_rng2.jsonl.)
# Cross-reference sweep shapes against the regular polytopes.
# (Filters out any Wythoff projection that happens to be 3D-congruent
# to a regular's projection -- empirically zero hits at rng=2. The
# legacy filename `novel_rng<N>.json` reflects the codebase's narrow
# definition of "novel"; see Terminology section above.)
python tools/analyze_sweep.py --rng 2 `
--out-novel ongoing_work/novel_rng2.json
# Emit .vZome files for every Wythoff-extension shape that survived
# the regular-coincidence filter (output/).
python tools/emit_novel.py --rng 2
lib/wythoff.py builds the regular polytope of
each group in a Cholesky frame derived from the Coxeter matrix, then
Procrustes-aligns it to the canonical embedding in lib/polytopes.py.
Wythoff variants of the same group inherit that rotation, so step-1
kernel directions transfer directly to step 2 with no further
alignment. However the Procrustes rotation in general is not an
axis-permutation of the canonical lattice basis, which means
gen_dirs() samples a different subset of the zomeable kernels for
the Wythoff embedding than for lib/polytopes’s embedding of the same
regular polytope. For this reason tools/analyze_sweep.py builds
its reference set from the sweep’s own (1,0,0,0)/(0,0,0,1) records
rather than re-running lib/polytopes-side searches.lib.search_engine.search and group_by_shape unchanged. Both
functions are now numpy-vectorised internally (_check_cos_pairs for
edge alignment, cKDTree + Gram-matrix pairwise distances for shape
fingerprints) — speedups of 39–42× on the H₄ 600/120-cell unlocked
the V=14400 omnitruncated 120-cell case for tractable analysis.V*(V-1)/2; for V = 14400 that is ~3 GB per shape, which on a
~30-shape group accumulates past 40 GB. shape_fingerprint’s
large_n_balls path replaces the sorted tuple with an order-
independent multiset accumulated via np.bincount over chunked
Gram-matrix tiles — peak RSS stays under 1 GB regardless of V.
The two fingerprint formats use different hashes, but the multiset
path only fires for V > large_n_balls (default 5000) so no
small-polytope record overlaps the regime; within a single sweep
every hit for a given polytope goes through the same path
consistently, so internal dedup is unaffected.Two-stage shape-equivalence dedup of kernels and shapes.
shape_fingerprint’s basis comes from np.linalg.svd(I − n̂ n̂ᵀ),
which has a degenerate eigenvalue 1.0 with multiplicity 3 — its 3D
output basis depends on small floating-point differences in n̂.
Two failure modes follow:
Stage A — scalar-multiple kernels. Kernels that are positive
scalar multiples of one another (e.g. k and φ²·k) round to
slightly different unit vectors after the manifest’s 4-decimal float
storage, and their fingerprints differ in the last bin of the
quantised distance multiset. Direction-equivalence dedup catches
this:
tools/run_wythoff_sweep.find_group_kernels collapses scalar-
equivalent kernels to one canonical (smallest-|k|) representative
before step 2 begins. This applies even to old caches.tools/emit_novel._dedup_by_direction collapses any remaining
scalar-equivalent representatives that survived step 2, per
polytope, before emission. Aliased fp_hashes are recorded in
each manifest entry’s aliases field for forward traceability.Stage B — H4-orbit-equivalent kernels. The deeper failure: kernels
that are not scalar multiples of each other — but lie in the same
orbit of the source polytope’s symmetry group — produce 3D shapes
that are rigid-motion equivalent (rotation + reflection + scale).
For example, the bitruncated 120-cell has 120 truncated-icosahedron
cells in a single H₄ orbit; their cell-centroid directions occupy
~60 distinct lines through the origin. Each line is its own
direction class (Stage A doesn’t merge them), but every projection
along an orbit-equivalent kernel produces the same 3D shape up to
rotation. After projection these shapes have an identical sorted
pairwise-distance multiset to FP precision, but
shape_fingerprint’s 3-decimal binning shifts on values sitting on
a rounding boundary, yielding different fp_hashes for what is
geometrically a single shape. Stage B catches this by reading each
emitted .vZome file and grouping by an exact geometric signature:
the sorted pairwise-distance multiset (rotation/reflection/scale
invariant) plus the edge-length multiset. Implemented in
tools/dedup_corpus_by_shape.dedup_manifest_by_shape, called
automatically as a post-emission pass in emit_novel.py and
available as a standalone CLI:
python tools/dedup_corpus_by_shape.py
The corresponding repair on a pre-fix manifest is
tools/dedup_corpus_by_direction.py (Stage A only) and
tools/dedup_corpus_by_shape.py (Stage B; supersedes the former).
At rng=2 these two passes collapsed 235 → 208 → 164 → 60 entries
(raw fp_hashes → after Stage A on the search records → after Stage A
on the manifest → after Stage B on the emitted corpus); the H₄
omnitruncated 120-cell alone went from 38 → 1. Stage B is also
available as tools/audit_shape_duplicates.py for inspection
without deletion.
ongoing_work/kernels_<group>_rng<N>.npy. Loaded
caches are run through the direction-dedup step, so old caches do
not need to be regenerated.ongoing_work/sweep_log_*.txt; on rerun, completed (group, bitmask)
pairs are skipped.ongoing_work/shapes_rng<N>.jsonl. Pass --force to re-test
polytopes already in prior logs (used to backfill the JSONL when
the dump format was added retroactively).--bitmask BMBMBMBM restricts a sweep to a single Wythoff variant
so large polytopes can be driven in parallel (one process per V≥7200
uniform). The per-job dumps are independent JSONL files which can be
concatenated into the master ongoing_work/shapes_rng2.jsonl once
all jobs complete.--sort-by-size sorts step-2 by polytope V count ascending so
that, for H₄, the V=14400 omnitruncated 120-cell is processed last
and the other 14 forms complete first if the run is time-bounded.B': 24 sporadic plus 30 generic rectangular-cuboid
kernels with B': 32 signature) — see output/regular/8cell/CLASSIFICATION.md
§3 for the structural enumeration; the previous “32” tally double-
counted one cuboid direction.emit_generic._normalize_scale). After
snap-to-ZZ[φ]³ the smallest projected edge length depends on which
scale s happened to satisfy _snap_zphi, which differs between
polytopes and gives wildly inconsistent strut/ball ratios in the
rendered .vZome. After classification (so directions are unaffected),
every coord is multiplied by a uniform φᵏ factor — φ is a unit in
ZZ[φ], so coords stay integral — bringing the smallest pairwise
vertex distance to within a factor of √φ ≈ 1.27 of a ~30-zome-unit
target. This makes the corpus visually consistent.tools/emit_novel.py writes each
natural-extension shape (legacy code-name “novel”) into
output/<common-name-slug>/ (e.g.
omnitruncated_120-cell/, cantellated_5-cell/) so the directory
tree is browsable by Wythoff form rather than fp_hash. The shape
index and short fp_hash are still in the filename for cross-
reference with manifest.json.