About
    • AMMDI is an open-notebook hypertext writing experiment, authored by Mike Travers aka mtraven. It's a work in progress and some parts are more polished than others. Comments welcome! More.
Search
Full
Incoming links

Protocolized Ontology

22 Sep 2026 - 22 Sep 2026
Open in Logseq
    • Gonna feed this right into a Claude project.
    • Traditional ontologies:
      • Hard, fixed, crystalline logical structures. Hard to use because people aren't that logical.
      • Bascially authoritarian / high-modernist
      • The advantages of a trad ontology: creates a common standard, everyone is using the same concepts and so no translation costs.
    • Local ontologies (new-ontologies or ghetto-ontologies – need a better name):
      • Created locally for local purposes
      • Note: every software project with a database schema basically is doing this, whether they know it or not! So this is the default.
      • Postmodern / poststructualist
      • Early explorations in BioBIKE work
      • A lot more suited to the squishy world of LLM-based software
    • So the question here: can you take a world of local, inconsistent ontologies, and design protocols that make them work together better? What would that mean?
    • Ask Claude
      • Here's a speculative sketch, roughly from plausible to strange. Handshakes instead of standards. When two systems meet, they run something like a TLS handshake for meaning. Each exposes its local schema with glosses and, crucially, exemplars: "here are five things I call cytosol." An LLM-mediated negotiator proposes a bridge mapping, and both sides validate it by round-tripping shared instances. That amounts to unit tests for semantics. The result is a pidgin scoped to that interaction, not a global truth. Pidgins creolize into standards. Most bridges are ephemeral. The ones that get reused across many pairs harden into shared vocabularies. This is your paper's "convergence as asymptotic endpoint" made mechanical. Standards aren't decreed; they're what the network's traffic has worn smooth. BFO-style upper ontologies would still exist, but as popular creoles rather than legislation. Mappings are first-class citizens. The interesting artifact isn't the ontology but the mapping, which is versioned, forkable, and carries provenance ("negotiated by agents A and B on date D, verified on 40 instances"). It also carries scope ("cytoplasm ≈ cytosol, for transport reactions in bacteria") and loss annotations ("this translation collapses a distinction the source cares about"). The cytosol/cytoplasm UI trick from your example becomes an explicit, inspectable object instead of a buried display-name hack. Equivalence is pragmatic, not essential. Two terms are "the same" if they produce the same downstream consequences for the task at hand, which is James operationalized. You test equivalence by running both systems' queries and diffing the results. Two terms can be equivalent for one purpose and distinct for another, and the protocol is fine with that. Perspective-tagged answers. Queries across the network don't return one answer. They return answers indexed by viewpoint: "Under BioCyc boundaries, this pathway has 6 reactions; under KEGG, 25." Contradiction is data, not an error state, and Hewitt-style inconsistency-robust logic is the native reasoning mode. Lazy, just-in-time alignment. Nobody aligns whole ontologies. You align only the fragment a query touches, at query time, and cache it. Most of the semantic web's failure was trying to do integration eagerly and globally. Jurisdictions and treaties. This is where the marriage example gets fun. Ontology communities act like states. They have Full Faith and Credit agreements ("we accept your diagnosis records as ours, with these exceptions"), DOMA-style refusals, and extradition rules for data. Meta-protocols govern how conflicts between treaties get resolved. The politics your paper describes stops being hidden and becomes explicit, auditable, and contestable. Economics and power don't go away. Translation has costs, so hubs emerge as Schelling points, the way English or Wikidata did. Whoever controls a hub vocabulary has soft power. You might see mapping brokers, reputation scores for bridges, and "exchange rates" between vocabularies (how lossy is going through X?). Decentralization reshapes the alpha ontologist rather than eliminating them. Ontological evolution. Terms behave like genes: they mutate through local tweaks, spread through reuse, and go extinct when nothing maps to them. You could run phylogenetics on concepts, tracing that "sample" in lab X descended from a LIMS schema from 2014 and diverged when they started doing CITE-seq. The ontology's history is its meaning, like a Wikipedia talk page promoted to the core data structure. Wild edge. Agents negotiating directly will drift toward private, compressed dialects that are efficient and illegible to humans. So the protocol probably needs a legibility requirement: every negotiated term must carry a human-readable gloss and exemplars, and mappings must be explainable on demand. Otherwise you get a machine Babel that works until it silently doesn't. Failure modes worth designing against:
      • Telephone-game drift: meaning degrades across chains of lossy translations. Mappings would need composable loss accounting, so A→B→C can report that it's worse than a direct A→C.
      • Laundering: a sloppy mapping picks up authority by being reused.
      • Ossification: a popular creole becomes the new authoritarian standard by inertia.
      • Adjudication: when two bridges disagree, someone still has to decide, and the protocol only makes that choice visible. The core shift is that the unit of ontology stops being the schema and becomes the negotiation. The world has many local ontologies plus a living, versioned web of partial, scoped, tested agreements between them. LLMs are what make this newly plausible: they're good at the fuzzy, exemplar-driven, prototype-style matching that formal alignment tools always choked on.
       (in Project, not here)