A return to first principles. Transforming the token from an isolated statistical fragment into a verifiable, situated unit of intelligence.
AI inherited Peirce’s word but abandoned the architecture it was designed to carry.
tokum restores the token as a situated sign for an open, collaborative network: connected to an object, interpreted by an accountable observer, and bounded by what can actually be verified.
A return to first principles. Transforming the token from an isolated statistical fragment into a verifiable, situated unit of intelligence.
AI inherited Peirce’s word but abandoned the architecture it was designed to carry.
tokum restores the token as a situated sign for an open, collaborative network: connected to an object, interpreted by an accountable observer, and bounded by what can actually be verified.
The AI industry inherited Peirce’s token, then reduced it to a statistical fragment disconnected from reality. tokum restores the triadic architecture of the sign—transforming the token into a situated, portable, and verifiable unit of meaning.
The AI industry inherited Peirce’s token, then reduced it to a statistical fragment disconnected from reality. tokum restores the triadic architecture of the sign—transforming the token into a situated, portable, and verifiable unit of meaning.
The extension is encoded in the word itself. tokum transforms the token from an isolated computational fragment into a situated, portable, and verifiable unit of meaning.
A discoverable and verifiable address for a concept. Carries context, provenance, observer and validation state. Returns a Semantic Zero when no verified claim exists within its present Semantic Light Cone of Care.
A situated unit that incorporates the observer. Restores the Peircean triadic relation through which meaning is produced. Its observer-dependent relations curve Semantic Spacetime.
The TCP/IP of meaning. Operating as a protocol of meaning, it seamlessly exchanges verifiable units of meaning on top of standard data packets—moving fluidly across agents, models, and domains.
tokum extends Saussure’s dyad into Peirce’s triad, restoring the missing leg of meaning: the interpretant, situated in an accountable observer.
The extension is encoded in the word itself. tokum transforms the token from an isolated computational fragment into a situated, portable, and verifiable unit of meaning.
A discoverable and verifiable address for a concept. Carries context, provenance, observer and validation state. Returns a Semantic Zero when no verified claim exists within its present Semantic Light Cone of Care.
A situated unit that incorporates the observer. Restores the Peircean triadic relation through which meaning is produced. Its observer-dependent relations curve Semantic Spacetime.
The TCP/IP of meaning. Operating as a protocol of meaning, it seamlessly exchanges verifiable units of meaning on top of standard data packets—moving fluidly across agents, models, and domains.
tokum extends Saussure’s dyad into Peirce’s triad, restoring the missing leg of meaning: the interpretant, situated in an accountable observer.
Semantic Zero is not merely a conversational refusal. It is the formal architectural state recognizing that a holon has reached the boundary of its verified knowledge. Reaching this boundary triggers seamless interoperability across the network rather than statistical hallucination.
No verified claim within the present Semantic Light Cone of Care→Semantic Zero.
Semantic Zero is not merely a conversational refusal. It is the formal architectural state recognizing that a holon has reached the boundary of its verified knowledge. Reaching this boundary triggers seamless interoperability across the network rather than statistical hallucination.
No verified claim within the present Semantic Light Cone of Care→Semantic Zero.
This is an open call rather than a closed demonstration. The proposed experiment validates a foundational horizontal protocol layer that seamlessly integrates with any system—LLMs, World-Models, or Agentic AI—without replacing the underlying architecture. It offers the same intercompatibility for verifiable knowledge exchange as the internet did for documents and data. We explicitly invite research labs, open-weight model developers, and academic institutions to build open-source testbeds, execute the five-arm ablation, and independently certify or contest the results.
This invitation extends to groups working on world-model architectures, abstraction-and-reasoning benchmarks, and AI safety and causal-reasoning research. Full falsification details are available in Section 11 of the preprint.
| Arm | Condition | Mechanism |
|---|---|---|
| 1 | Baseline | Unmodified open-weight model, prompted normally. |
| 2 | RAG-Only | Ordinary retrieval over the same corpus without cryptographic provenance or explicit trust weights. |
| 3 | Semantic-Zero | Dedicated NIL-class or abstention head returning a bounded refusal below a preregistered threshold. |
| 4 | CCI/CTI | Retrieval results delivered as CCI-addressed, cryptographically signed CTIs with explicit trust weight, observer identity, timestamp, and context. |
| 5 | Hybrid | Model uses both the NIL-class Semantic Zero and CCI/CTI trust-weighted retrieval. |
These mechanisms are not expected to produce a substantial increase in genuinely novel-relation task performance. If they unexpectedly produce a substantial increase in discovery, or fail to reduce unsupported assertions relative to RAG-only, the central claims require revision.
A system that answers 'Semantic Zero' to everything scores perfectly on safety but possesses vacuous utility. Falsification must strictly distinguish between 'checked and refuted' claims (validator disagrees) versus 'unverifiable/NIL' claims (validator unavailable).
This is an open call rather than a closed demonstration. The proposed experiment validates a foundational horizontal protocol layer that seamlessly integrates with any system—LLMs, World-Models, or Agentic AI—without replacing the underlying architecture. It offers the same intercompatibility for verifiable knowledge exchange as the internet did for documents and data. We explicitly invite research labs, open-weight model developers, and academic institutions to build open-source testbeds, execute the five-arm ablation, and independently certify or contest the results.
This invitation extends to groups working on world-model architectures, abstraction-and-reasoning benchmarks, and AI safety and causal-reasoning research. Full falsification details are available in Section 11 of the preprint.
| Arm | Condition | Mechanism |
|---|---|---|
| 1 | Baseline | Unmodified open-weight model, prompted normally. |
| 2 | RAG-Only | Ordinary retrieval over the same corpus without cryptographic provenance or explicit trust weights. |
| 3 | Semantic-Zero | Dedicated NIL-class or abstention head returning a bounded refusal below a preregistered threshold. |
| 4 | CCI/CTI | Retrieval results delivered as CCI-addressed, cryptographically signed CTIs with explicit trust weight, observer identity, timestamp, and context. |
| 5 | Hybrid | Model uses both the NIL-class Semantic Zero and CCI/CTI trust-weighted retrieval. |
These mechanisms are not expected to produce a substantial increase in genuinely novel-relation task performance. If they unexpectedly produce a substantial increase in discovery, or fail to reduce unsupported assertions relative to RAG-only, the central claims require revision.
A system that answers 'Semantic Zero' to everything scores perfectly on safety but possesses vacuous utility. Falsification must strictly distinguish between 'checked and refuted' claims (validator disagrees) versus 'unverifiable/NIL' claims (validator unavailable).
We use cookies to provide you a better user experience on this website. Cookie Policy