A Divination Corpus Misdescribed as a Panel Corpus

1. What the pipeline actually returned, and why that matters first

The headline number is a null result. Of 156 approved detections across 43 panels, the corpus summary reports 155 left unclustered and one motif family surviving — yet three cluster descriptions were supplied (8, 14, 16), and cluster 8 alone claims nine members. The bookkeeping does not close. Worse, the panel readings for FoA_04-5580 and FoA_04-5583 both state flatly that all their detections are unclustered, while cluster 8's distribution claims two members from FoA_04-5580 and one from FoA_04-5583. The one positive family-level finding in this run is therefore in direct contradiction with two of the panel-level records. Nothing downstream of that should be treated as settled.

This is not a peripheral complaint. It means the collection's shared visual vocabulary has to be reconstructed almost entirely from the panel readings — that is, from human-and-model description of individual images — with the clustering contributing almost no independent corroboration. Every "family" named below is a typological claim, mine or the panel readers', not a machine-discovered one, and I will keep saying so.

Two further structural facts about the run:

HDBSCAN split a bilateral pair into two singleton clusters. Clusters 14 (vertical_granular_border_strip) and 16 (narrow_vertical_textured_band) have effectively the same visual definition — a tall granular column, framed on the long sides, capped by a pale X- or cross-shaped terminal — and both are logged to the same object, the Ado-Ekiti palace door (EBA-Div_00303). The panel reading for that door identifies detections #0 and its mirror #29 as chequered/plaited filler strips. Clusters 14 and 16 are, on the balance of evidence, the left-leaf and right-leaf instances of one carved band, which the algorithm failed to recognise as the same thing. That is a diagnostic failure: the door's single most legible organising principle is mirroring, and the pipeline could not see across the mirror.
Both of those clusters inherited the label warriors_procession_panel from their parent panel, and both cluster notes reject it. The labels are leaking from panel-level captions onto motif-level segments. Any label in this dataset that describes figures where the crop shows texture should be assumed to be inherited rather than observed.

2. What the collection actually is

Read across the panel titles, this is not "a collection of carved West African panels." Counting object classes as the readings identify them:

Divination trays (opon Ifa), circular and rectangular: EBA-Div_00302, FoA_04-5580, FoA_04-5583, FoA_04-5585 (arched format), FoA_04-5586, FoA_04-5587, FoA_04-5588, FoA_04-5589, FoA_04-5947, plus the drawn trays KBA_10298, KBA_10299, KBA_10300, KBA_10305, KBA_15356, KBA_17555, KBA_17556, and the excerpted tray faces on KBA_17553.
Door leaves, stiles and posts: EBA-Div_00303 (Ado-Ekiti palace door), EBA-Div_00311 panels 00–02, EBA-Div_00312 panels 00–01, FoA_04-5578 panels 00–02.
Shafts, finials, tappers, staffs: FoA_04-5597, FoA_04-5598 (four iroke Ifa on one plate), FoA_04-5599.
Lidded vessel: KBA_10274.
Archive mount sheets reproducing published plates: KBA_19153, KBA_19154, KBA_19155, KBA_19157, and the drawing sheet KBA_17557.

So roughly half the corpus is divination furniture and about a fifth is not an object at all but a document about an object. Only one item — the Ado-Ekiti door — is a large multi-register figural door of the kind the corpus title implies. Localities, as given: Modakeke (Ife) on eleven FoA items and, if the caption handwriting is being read correctly on the mount sheets, "Modeke" again on KBA_19153–19155; Ado-Ekiti on one; Ife on EBA-Div_00311 and 00312. This looks like one collecting concentration, not a regional survey.

3. The shared vocabulary

Structural families — framing, dividing, filling. These are the collection's true recurrences, and they cross object class freely:

Interlace / guilloche / plaitwork. The single most pervasive idiom. It appears as a full four-sided frame (EBA-Div_00311_panel_01, where the reader notes interlace on all four sides as unusually complete), as paired full-height lateral bands enclosing a figural channel (EBA-Div_00312 panels 00 and 01), as an A–B–A alternation with figural episodes on a post shaft (FoA_04-5578_panel_00, panel_02), as an entire undivided panel face (FoA_04-5578_panel_01), as a tray-rim band diametrically opposite the head (FoA_04-5583 #0; KBA_10305 #3), and as the whole decorative programme of a rectangular tray frame (FoA_04-5587 #0, #1, #2). Notably, FoA_04-5587's reader observes that the top and bottom runs use different interlace idioms — guilloche versus hook-chain — inside an otherwise symmetrical frame. That is a carver's choice, and it is the kind of observation clustering ought to have surfaced and did not.
Hatching, combing, crosshatch, mat/basketry diaper. Used as compartment divider on tray rims (EBA-Div_00302, FoA_04-5580, FoA_04-5583, FoA_04-5589's alternating combed blocks), as ground and register-divider on planks (EBA-Div_00311_panel_02's X-in-square, herringbone and ladder fields), as body and garment texture, and as the draughtsmen's convention for recessed ground on the KBA sheets. It is also, almost certainly, what clusters 14 and 16 are.
Chevron, sawtooth, zigzag, triangle chains. Register dividers and terminal bands: the lintel-like zigzag frieze over the raised-arm figure in FoA_04-5578_panel_00, the chevron divider in EBA-Div_00311_panel_00, the triangle chain on the lower rim of KBA_15356, the allover triangle rows on the vessel body of KBA_10274.
Concentric bosses, discs, rosettes, spirals. The roundel column on the board intruding at the left of the FoA_04-5578_panel_01 plate; the coiled spiral terminating EBA-Div_00311_panel_00; the paired spirals flanking the lower face on KBA_17556; the apex medallion of KBA_10274; the toothed ring of EBA-Div_00312_panel_00.

Subject-bearing families. Fewer, more consistent, and — this is the point — deployed almost always inside the structural families rather than alongside them:

The axial frontal face. On trays it is at the top of the rim and the readings uniformly gloss it as oju opon, conventionally Esu (FoA_04-5580 #0, FoA_04-5583 #1, FoA_04-5586, FoA_04-5587, FoA_04-5588 #0, FoA_04-5589 #0, FoA_04-5947, plus the drawn trays KBA_10298, KBA_10299, KBA_10300 #0, KBA_10305, KBA_15356, KBA_17555, KBA_17556). It recurs off the trays too, as the single relief face on the shafts FoA_04-5599 panels 00 and 01, as the head finial FoA_04-5597_panel_02, and as the crowned figure on the mount-sheet vignettes KBA_19153–19157.
The frontal standing figure, splayed or arms-akimbo, oversized head, striated lower body. Triplicated on an axis in EBA-Div_00311_panel_01 and EBA-Div_00312_panel_01; single in EBA-Div_00311_panel_00; paired and mirrored in FoA_04-5578_panel_02; arms raised in FoA_04-5578_panel_00; massed in files on EBA-Div_00303.
The small kneeling / crouching / prostrate figure. This is what cluster 8 (crouching_relief_figure) is about, and it is the corpus's only surviving family. Its distribution is worth stating plainly: it is tray-rim material. Six members from the EBA-Div_00302 tray, two from FoA_04-5580, one from FoA_04-5583 — all three divination trays, two of them Modakeke.
Border fauna. Mudfish, saurians, birds, tortoise, snake — read on FoA_04-5586, FoA_04-5589, FoA_04-5947, and in the drawings KBA_10300 (long hatched saurian bodies as the most repetitive type), KBA_10305, KBA_17555.
Regalia and implements. Bowls, fans, whisks, staffs, the ovoid clusters read as possible ikin on FoA_04-5588, the flared arch on EBA-Div_00311_panel_02, the horn on KBA_17557.
The equestrian. Once, securely: on the lid of KBA_10274. Possibly in register 1 of EBA-Div_00303.

4. The compositional grammar

Two grammars, sharing four devices.

Grammar A: annular, oriented, void-centred. Every tray reading arrives at the same structure independently — a deliberately blank working field, all carving displaced to a continuous rim, the rim compartmented by hatched spacers, and one axial accent (the face) fixing a top. EBA-Div_00302 gives the cleanest demonstration, because the reading identifies the failure of symmetry as the informative fact: the crowned head at the bottom of the photograph has no counterpart at the opposite point, so the vertical axis is polarised rather than merely symmetrical. KBA_10305 makes the same point quantitatively — the maximum adjacency distance on the panel (0.77) is face-to-plait, i.e. the two most emphatic elements sit diametrically opposite.

Grammar B: axial, framed, register-stacked. Doors, stiles and posts: a frame of interlace or plait, a central channel, and content stacked vertically with scale doing the hierarchic work. EBA-Div_00312_panel_01 is described as "frame-plus-file," with ornament taking ~51% of detected area and three equal figures in the channel. FoA_04-5578_panel_00 is A–B–A: interlace, figure, interlace. EBA-Div_00311_panel_00 alternates geometric field, large figure, chevron divider, small-figure row, large abstract emblem — abstraction bracketing narrative.

Shared devices across both:

Enclosure over sequence. Almost no panel here narrates laterally. Even the Ado-Ekiti door, the one object with genuine processional rows, subordinates them to a mirrored grid.
Bilateral symmetry about a single vertical axis, broken once. Ornament is symmetrical; the axial accent is singular. FoA_04-5589 states this exactly: everything else is paired, banded or repeated, the face alone is frontal, singular and axial.
Scale as rank. Explicit on EBA-Div_00303 (the X-marked figure at roughly triple surrounding scale, flanked by miniature attendants) and on FoA_04-5588 (one oversized face against dozens of small figures).
The reserved void. The corpus's most distinctive formal fact, and the one the pipeline stumbled into: on FoA_04-5585, FoA_04-5586, KBA_10299 and FoA_04-5947, the detector's only or principal box lands on the empty field, not on the ornament. The narrow, high-contrast, often curved decorated band defeats a rectangular detector, while the large uniform void attracts it. That is an artefact — but it is an artefact that correctly registers where the visual mass isn't.

5. Workshops, regions, repertoire

What the evidence supports:

A locality concentration, not a region-wide sample. Eleven FoA items captioned Modakeke (Ife), plus the Ife-filed EBA-Div_00311/00312, plus the mount sheets' "Modeke." The Ado-Ekiti door is a single outlier of a different object class from a different town. Comparative claims about "Yoruba doors" from this corpus rest on one door.
A shared repertoire, at least at the level of ornament, spanning object classes within that concentration. The same interlace, the same combed spacers, the same axial face appear on trays, posts, shafts and finials. FoA_04-5597_panel_02's reader makes the sharp observation that the twisted collar band on a head finial is "the same rope-and-twist border motif" that frames door panels and tray rims — migrated from a border to a body ornament. That kind of migration is the strongest available argument for a single carving vocabulary rather than object-specific ornament sets.
One candidate workshop signature. FoA_04-5583's mirrored ring-eyed, curved-beaked creature with hatched upper field and inward spiral (#4/#6) is flagged as "a distinctive local formula." That is exactly the sort of narrow, idiosyncratic form whose recurrence would establish a hand or workshop — and exactly the sort the clustering failed to test, because it left all eight of that tray's detections unclustered.

What the evidence does not support: any inference about workshops from cluster 8. Its stated cohesion (0.940) is lower than its cosine distance to clusters 17, 20 and 5 (0.976, 0.972, 0.956), which the cluster note reads correctly as arbitrary slicing of one continuous population; and the cluster note itself suspects CLIP is keying on photographic grain and raking light rather than carved form. A family whose coherence may be photographic cannot license a claim about carvers.

6. As a communicative system

The panel readings converge, independently, on a proposition I think the collection does support: these objects communicate orientation, enclosure and rank, not episodes.

Almost every reading reaches for the same negative finding — no interaction between figures, no attribute passed along, no scene. EBA-Div_00311_panel_01: "repetition and enclosure rather than telling an episode." EBA-Div_00312_panel_01: "a repeated assertion rather than a story." FoA_04-5583: "an instrument rather than a narrative panel." FoA_04-5599: "no procession, no mounted figure, no attendant group." Even the door, the most narrative object present, is read as "retinue and increase" — an assertion of countable bodies, restated symmetrically so that either leaf makes the same claim.

The system therefore has a small number of grammatical operations: mark an axis with a face; reserve a centre; enclose with continuous plait; rank by scale; fill the remainder with metrical texture. What the corpus does not give is a lexicon of identifiable subjects. The readings are candid about this: no fan, no mudfish legs, no opa, no secure cutlass or bird on EBA-Div_00311_panel_01; no diagnostic attribute at all on panel_02; the whiskered face on FoA_04-5578_panel_02 unresolvable between animal mask and cheek scarification; the ring on EBA-Div_00312_panel_00 undecidable between coiled serpent, armlet, drum-head and rayed disc.

Two specific temptations should be resisted. First, the interlace gloss. Most panel readings attach continuity, binding and endlessness to plaitwork, always flagged as inference. FoA_04-5578_panel_01 alone refuses: "such glosses are commonly offered for African interlace but are not supportable from the object alone." That refusal is the methodologically correct position, and its presence alongside a dozen panels that do offer the gloss shows how easily a conventional reading propagates through a corpus by repetition rather than by evidence. Second, Esu. The identification of the axial tray face as oju opon / Esu is a well-attested positional convention and I hold it, from knowledge, with reasonably high confidence as a convention. But it is asserted here on sixteen-plus objects on the basis of position and frontality alone, in no case with an inscription or a diagnostic attribute. It is a type-level inference repeated many times, not sixteen independent confirmations.

7. Weakest points, and what would settle them

The clustering contradicts itself. Cluster 8 claims members from FoA_04-5580 and FoA_04-5583; both panel readings say all their detections are unclustered. Settled by: reconciling the detection-level cluster assignment table against the panel records, and reconciling the 156 approved detections against the panel readings, which between them describe well over 150 detections on the first two panels alone.
Cluster 8 may be a photographic family, not a carving family. Its internal cohesion is lower than its distance to three unexamined neighbours. Settled by: inspecting clusters 5, 17 and 20 jointly with 8; and re-embedding after normalising tone and grain, or with a shape-biased encoder, to see whether the grouping survives the removal of raking-light cues.
Clusters 14 and 16 are probably one motif on two door leaves. Settled by: checking whether their members are #0 and #29 of EBA-Div_00303, and by testing both against a purpose-built set of rope-twist/plait crops versus standing-figure-row crops. Adjacency to the former confirms the border reading and kills the inherited procession label for good.
Label provenance is untracked. warriors_procession_panel on a texture crop shows captions descending from panel to motif. Settled by: recording, per label, whether it was generated from the crop or the parent image — a metadata field, not a modelling problem.
The detector cannot see narrow curved borders. On FoA_04-5585, FoA_04-5586, FoA_04-5587, FoA_04-5588, FoA_04-5947 and KBA_10299 the decisive elements — including every axial face on some of them — went undetected while voids and paper mounts were boxed. Settled by: polar unwrapping of circular trays into rectangular rim strips before segmentation, and rotated or polygonal proposals for rectangular frames.
Media strata are not separated. Photographs, pen-and-ink survey drawings, and mount sheets reproducing published plates (KBA_19153–19157, KBA_17557) are being embedded in one space. On KBA_17555 the single detection is the paper mount; on KBA_19154 and KBA_19157 it is the pasted illustration entire. Settled by: a medium classifier applied before clustering, and separate runs per stratum. Until then, no cross-panel family can be trusted not to be a family of reproduction techniques.
The one candidate workshop formula is untested. FoA_04-5583's ring-eyed beaked creature. Settled by: a targeted retrieval query using that crop against the whole FoA Modakeke run and the KBA tray drawings. A single confirmed recurrence would be worth more than any of the general ornament families.
The corpus does not support its own title. Settled by: re-tabulating by object class. If the analytical question concerns carved door panels, this collection contains one door (EBA-Div_00303) plus a handful of stiles and post faces, and the divination material should be analysed as its own group with its own grammar — annular, oriented, void-centred — rather than forced into the register model that the panel readings correctly reject on tray after tray.