Hierarchical Bits — Presentation Pitch

In one sentence: a representation model where multiple — possibly contradictory — interpretations share one immutable substrate and stay queryable, without duplicating the data and without forcing them into a single truth.

SLIDE 1 — The problem

Don't duplicate the world every time someone disagrees with it.

Start from a common base — one dataset, one building, one document set, one conversation history. Over that base, multiple interpretations naturally arise: annotators label it, disciplines read it differently, hypotheses compete over it, versions accumulate. They coexist.

Today, holding them forces a bad choice:

• copy the base once per interpretation → K readings cost K copies of the world; or
• merge to one → a dominant representation wins and the rest is lost.

SLIDE 2 — The idea (a model, not a file format)

BH is a representation model for shared immutable substrates and concurrent interpretations.

One substrate, stored once and immutable. Each interpretation is a first-class, co-registered entity over it. The reader picks the lens at read time — adjudication is deferred and optional, never baked in.

SUBSTRATE   stored once, immutable, shared by every reading
LAYERS      each interpretation is a first-class, co-registered entity
READINGS    one lens / the matrix / the majority / the disagreement
            — your choice, at read time, not baked in

We give this property a working name — the First-Class Interpretation Representation (FCIR): interpretations kept as persistent, addressable, first-class entities over a shared substrate, rather than temporary versions or conflicts to be resolved away. (Working name — see Slide 4.)

SLIDE 3 — The distinguishing test (falsifiable)

Given two interpretations that disagree about the same element — can both remain, neither marked wrong, until a reader chooses (or declines) to adjudicate?

Many systems end up converging to a dominant representation, or isolating each interpretation into an independent copy/version. BH keeps them co-registered over one substrate and lets adjudication wait. That is the differentiator — stated as a test you can run on any system, not a boast.

SLIDE 4 — What we are NOT (the honest positioning)

We surveyed 20 data domains. The result killed the easy claim that "BH is universally new":

Storing the substrate once + reading it selectively is already mature SOTA — DICOM, COG/STAC, lakeFS, CRAM/tabix, S-LoRA, MAM. BH does not claim to invent that.

Our 20-domain sweep identified the First-Class Interpretation Representation (FCIR) — keeping rival readings as preserved entities instead of resolving them away — as the property that best distinguishes BH from the approaches evaluated. That is a result of the investigation, not a universal claim; FCIR is a working name, and the property may yet prove broader. Saying plainly what BH is not — and how far the claim reaches — is what survives the skeptical engineer.

And FCIR is not unique to BH: RDF named graphs and standoff annotation already implement it in their own domains. So FCIR is a synthesis and a cross-domain name, not a new mechanism — judged as a synthesis, not an invention (full confrontation in BH_PRINCIPLE.md).

SLIDE 5 — Where it fits, and where it doesn't (the useful limit)

FITS     multiple readings of ONE base object:
         · annotation with annotators who disagree
         · agent memory with conflicting versions over one history
         · BIM/CAD — disciplines reading one building (not five copies)
         · legal / eDiscovery — rival readings of one document set
         · science — competing hypotheses over the same raw data

DOESN'T  dense signal (photo / audio / embeddings) → delegate to codecs
         single-truth goals (consensus, gold labels) → already solved

A pitch that states its own limit is the opposite of vapor.

SLIDE 6 — The evidence (the principle is reproducible)

The same model showed up — independently — in four completely different domains. That reproducibility matters more than any single number:

One principle, four instances, each measured and tested — correctness as a gate, honest baselines, public self-corrections, a Zenodo DOI. The numbers exist (4.6×, 35×, 1,779×, 9×); the point is that the principle held every time.

SLIDE 7 — The state and the ask

• It is a principle with measured instances, not a finished product. The sweep found its useful limit — and a useful limit is where a serious product starts; without one, it's religion.
• The next question is product, not novelty: of the domains where it fits, which is the first seed worth building for real?

Don't duplicate the world every time someone disagrees with it.