Conclusion (provisional)

This project began as a strong hypothesis and is reported here at the honest size the evidence supports — no larger.

Initial hypothesis. BH introduces a new, general paradigm of representation.

Result of the investigation. That hypothesis was not confirmed. The principles it leaned on — a shared substrate stored once, and selective reading over it — already appear, in mature form, across many domains (DICOM, COG/STAC, lakeFS, CRAM/tabix, S-LoRA, MAM, …). As a general paradigm, BH is not novel.

What remained. A recurring property around the representation of concurrent, possibly contradictory interpretations over one shared substrate. It appears in some specific domains — notably RDF named graphs (+ provenance) and standoff annotation — but has no uniform, named treatment across them. We give it a working name: the First-Class Interpretation Representation (FCIR).

Current status. There is not yet sufficient evidence that this property constitutes a new concept. The most defensible contribution at this moment is to treat it as a proposal of synthesis and classification — a cross-domain name plus a falsifiable test — subject to further validation. Where existing systems already implement it, this report says so (see BH_PRINCIPLE.md).


The conclusion, in three lines

This is a strong conclusion precisely because it is compatible with everything observed. It does not overstate what was found, and it does not discard the work done.

What the research is, regardless

A hypothesis was pursued, pressure-tested, and cut down to what survived: hypotheses tested, counter-examples sought, scope reduced when the evidence demanded it, contrary evidence incorporated rather than explained away. The honest win is the method as much as any single finding — and a measured trail with its boundaries and its prior art named in public.

Responding to external critique (DeepSeek review, 2026-06)

An external model-based review raised five points. Two were already addressed in this repo (immutability is not new → the prior-art confrontation in BH_PRINCIPLE.md; the 20-domain method → applicability/). Three were fair and are addressed here:

(a) "No formal algebra → a pattern, not a model." Addressed in BH_ALGEBRA.md: the operators ( coexistence, Δ conflict, σ projection, α adjudication, precedence, composition), the laws, and FCIR stated formally as the decoupling of coexistence from adjudication (⊕ ⊥ α). It is a specification, not yet a verified theory.

(b) Trade-offs (the costs, honestly). The family is not free: immutability → monotonic storage growth, needing compaction/GC and content-addressing; co-registration → interpretations must address substrate elements stably, so substrate re-addressing breaks layers (the standoff "offset-drift" problem); preserving conflict → read-time adjudication is O(layers) per address. None of these are solved here — they are the known costs of the immutable-substrate family.

(c) Comparison to real tools, not naive baselines. The prototype numbers are vs a naive flat store, not vs SOTA. Said plainly: against the real tools the prototypes do not win. bhanno is standoff annotation (brat / Label Studio / Web Annotation do it better and at scale); bhckpt's selective load is what safetensors / S-LoRA already ship; bhmem sits below Mem0 / Zep. The prototypes are illustrations of the property, not competitors; the claim is the property + the test, not the implementations.

(d) Where it would actually help. The most defensible use case is preserving and learning from disagreementhuman label variation in ML. Pipelines today collapse many annotators into one gold label and discard the disagreement, which a growing literature argues is real signal (ambiguity, subjectivity, uncertainty) being thrown away. FCIR's narrow but concrete value: keep every annotator's reading first-class and adjudicate at read-time, per consumer (one model trains on majority, another on the full distribution, an auditor inspects the conflict) — without forking the dataset. The usual gap is not storage (standoff already stores layers) but the pipeline collapsing them to gold. That gap, not the bytes, is the opportunity.

This does not make FCIR novel (see the prior-art confrontation). It makes the critique part of the record.

Paths from here (open)

None of these requires the original hypothesis to have been right:


Companion to BH_MASTER.md (the measured study), applicability/ (the domain sweep) and BH_PRINCIPLE.md (the FCIR confrontation). Provisional — it should be revised if stronger evidence, for or against, appears.

Hierarchical Bits · © 2025–2026 Márcio M. Carvalho · code under Apache-2.0, docs under CC BY 4.0 · repository