hierarchical-bits
BH Applicability Scorecard — research-grounded sweep of 20 domains
Method: one web-grounded research agent per domain gathered real data volumes, the tools used today, and whether interpretations conflict; each scored 0-3 on five criteria. The composite below is computed from those scores with declared weights (transparent, reproducible in
scorecard.py). Scores are research-informed estimates, not market data.
Read it in two axes: the composite says how BH-shaped a domain is (substrate + many layers + selective read). The verdict says whether there is novel work to do — because being BH-shaped is not enough if the store-once + selective-read pattern is already mature SOTA.
| # | domain | composite | verdict | scale | why |
|---|---|---|---|---|---|
| 1 | Knowledge graphs (multi-ontology) | 100 | ANCHOR | 16B triples | named graphs + RDF-star already SOTA |
| 2 | Medical imaging (multi-reader) | 95 | ANCHOR | WSI 1-100 GB | DICOM SEG references source — mature |
| 3 | Earth obs / satellite | 95 | ANCHOR | 10+ PB | COG + STAC already do it |
| 4 | Legal eDiscovery | 95 | ANCHOR | TB / matter | Relativity store-once + coding fields |
| 5 | Video archives / MAM | 95 | ANCHOR | PB masters | MAM timecoded tracks — mature |
| 6 | Dataset versioning / data lakes | 95 | ANCHOR | PB | lakeFS/DVC zero-copy IS the win |
| 7 | Autonomous driving datasets | 95 | DELEGATE | Waymo 2 TB | sensor signal -> codecs; sidecars mature |
| 8 | Data labeling / annotation ★ | 93 | ANCHOR | COCO 20 GB | CVAT/Label Studio store-once — mature |
| 9 | Genomics (variant calling) | 93 | ANCHOR | ref 3 Gbp | CRAM + tabix selective — mature |
| 10 | Geospatial map tiles | 90 | ANCHOR | PB | COG/PMTiles/vector tiles — mature |
| 11 | CAD / BIM | 90 | BUILD | tens of GB | federation duplicates; rival overlays UNSERVED |
| 12 | Model checkpoints / MoE ★ | 85 | ANCHOR | base GB-TB | S-LoRA / Punica already SOTA |
| 13 | Agent memory ★ | 75 | ANCHOR | KB-MB text | substrate LIGHT; Mem0/Zep mature |
| 14 | Scientific sim / HPC ensembles | 75 | DELEGATE | CMIP6 30 PB | per-member dense -> Zarr/zfp |
| 15 | Time-series / IoT | 68 | DELEGATE | 1-50 TB/day | Gorilla codec owns the signal |
| 16 | RLHF / preference data | 62 | NO | MB-GB text | substrate = label order; payload IS the point |
| 17 | Distributed tracing ★ | 58 | ANCHOR | TB/day | many distinct traces; Tempo mature |
| 18 | 3D scenes / glTF (LOD) | 45 | DELEGATE | MB-TB | geometry -> Draco/Nanite; LOD = dup copies |
| 19 | Vector DBs / embeddings * | 30 | DELEGATE | TB | control: vectors ARE payload; PQ/HNSW own it |
| 20 | Image / audio codecs * | 25 | DELEGATE | exabytes | control: signal IS payload; codecs irreducible |
★ = prototype already built in this repo (bhmem / bhtrace / bhckpt / bhanno).
The honest finding
- Most domains score HIGH (90-100) — and most are
ANCHOR. The BH shape is everywhere, but store-the-substrate-once + selective-read is already mature production SOTA almost everywhere it matters: DICOM (medical), COG+STAC (satellite/GIS), lakeFS/DVC (data lakes), CRAM+tabix (genomics), MAM (video), Relativity (legal), named graphs (KGs), S-LoRA (checkpoints). BH earns credibility by analogy there, not a novel build. - Among the 20 domains surveyed, CAD/BIM (90) was the only one classified
BUILD. The research found no mainstream tool that stores ONE canonical building substrate once with many additive AND rival discipline/version overlays co-registered as first-class layers with selective branch/region reads — today federation duplicates and treats clashes as ad-hoc. That union is the gap. (A claim about this sample, not about every possible domain.) - The recurring under-served slice is the rival layer, not the shared one. Across annotation, MAM, eDiscovery, genomics and KGs, existing tools store the substrate once but treat conflicting interpretations as noise to adjudicate into one ground truth — not as first-class, queryable, co-registered rival layers. That is precisely what
bhannomodels. So the sweep suggests BH’s principal still-under-explored contribution is treating rival interpretations as first-class entities — narrower and sharper than ‘a universal format’. Stated as the differential observed so far, not a final reduction of what BH can be. - Our four built prototypes all land
ANCHOR. They proved the generalization (one envelope, many domains) honestly — but the sweep shows their economic win is largely already-solved. That is the method working: it refuses to let us overclaim. - Controls behaved. Image/audio codecs (25) and vector DBs (30) sit at the bottom as
DELEGATE— the score discriminates dense-signal from structure-dominant.
Recommendation
Two moves, in order. First, formalize the principle (see BH_PRINCIPLE.md): the sweep shifts BH from a file format to a representation model — multiple concurrent, possibly contradictory interpretations sharing one immutable substrate and remaining queryable without forced adjudication. That definition is what separates BH from the ANCHOR systems; it is the real contribution this sweep surfaced.
Then, if a measured test is wanted, the natural candidate in this sample is CAD/BIM — the only domain here classified BUILD — a .bh where the building object-graph is the substrate, with rival discipline/version overlays + clash annotations co-registered, selective branch reads, and the deferred-adjudication face.