Genesis PreviewE1 admittedSemantic Snapshot implemented500 Atlas identitiesQuery Lab: liveRead-onlyNo arbitrary URLsHosted CI: startup blocked

Live read-only Genesis baseline plus protected E4.7 agent, ontology and model candidates. Public MCP, frontier-model execution and TNOE training are not deployed; no arbitrary URLs.

TWIRX Reference implementation and public service of the Typed Web Commons

Model program

Two forms of model intelligence. Neither becomes semantic authority.

A frontier model will consume TWIRX through MCP. WSIM and the future TNOE learn candidate mappings inside TWIRX. The deterministic data plane remains between them, and every training or deployment claim is admitted separately.

The model is not one thing

  1. HumanquestionIntent and policy
  2. Frontier modelquery proposalGeneral reasoning
  3. TWIRXdeterministic stateExecution and proof authority
  4. WSIM / TNOEmapping candidatesInternal learning
  5. Reviewersigned decisionCanon admission
The frontier model consumes TWIRX. TNOE helps compile TWIRX. The deterministic data plane prevents either model from granting itself authority.

Current truth

The frontier gateway exists but has made no paid provider call. WSIM-C0 is an experimental candidate ranker. TNOE is an architecture and graph scaffold; no trained TNOE model exists.

What already augments a frontier model

The protected twirx_context binding removes deterministic assembly work from the model. One typed query produces answer-shaped source claims, the relevant native-to-semantic bindings, proof identities, execution metrics and explicit freshness, mapping and authority state. Agent Answer 0.2 then prevents the model from narrating stale, provisional or archive evidence as something stronger.

WSIM Seed: measured candidate retrieval

5 retrieval systems were compared on 70 examples, including 10 cases that required abstention.

Strongest direct result on the small E4.6 set
SystemBGE-M3
MRR0.953
Top-1 accuracy0.917
Encoder fine-tunedNo
AuthorityCandidate-only

WSIM-C0 abstained on all held-out unknowns but also rejected answerable examples; calibration error remained 0.259 and 0.293 across the two held-out folds.

These results cover two source families and do not establish Web-wide mapping performance.

TNOE-Align: dual-graph semantic induction

The implemented scaffold aligns a Native Schema Graph with an Ontology and Frame Graph. A future trained model will rank concepts and roles, classify mapping relations, detect compatibility and identity risk, and abstain under uncertainty. 

Native schema encoder            Ontology graph encoder
 field / path / type             + concept / frame / role
 value shape / neighbors         + relation / context / constraints
 language / source family        + reviewed mapping evidence
             \                    /
              cross-graph alignment
                        ↓
  concept · role · relation · compatibility · abstention
Native nodes
11
Ontology nodes
18
Ontology edges
31
Candidate alignments
20

Graph: sha256:f440e4dfe10cba9ea6edd5a86cf58dbbc0fc69f2eceb1e8b36c6ca71097ede6b

Roadmap to a powerful TWIRX-native model

  1. R0 — Agent-facing contract hardening Implemented — not admittedNine bounded MCP tools, including a compact task-ready agent context; evidence-bound Answer 0.2; effective query templates; proof escalation; readable telemetry; full Semantic Query schemas; and public explanatory surfaces. Admission: founder review required.
  2. R1 — Frontier agent pilot NextPublic read-only MCP, 20-scenario cost pilot, same frontier model across conditions, evidence-bound answers and published losses. Admission: budget, credential, public-endpoint and operator-security approval required.
  3. R2 — Ontology corpus and TNOE data admission PlannedFive source families, 250 unique mapping units, 500 signed decisions, 1,000 reviewed hard negatives and four useful universes. Admission: human policy and signed semantic review required.
  4. R3 — TNOE-Align 0.1 training candidate PlannedDual-graph alignment model, concept and role ranking, relation classification, calibration, abstention, identity-risk and security-anomaly evaluation. Admission: R2 floors and source-family-held-out evaluation required.
  5. R4 — Continuous diverse Web expansion PlannedWorld State expansion, Crossref, APIs.guru, NVD and Federal Register; packet, frame, delta, diversity and query-utility releases. Admission: per-source policy and bounded evidence acquisition required.
  6. R5 — Powerful TNOE research line ResearchReplace frozen text features when justified, add temporal change learning and active-review proposals, and evaluate generalization without granting model authority. Admission: measured data, compute, safety and independent evaluation gates required.

The path scales model power by increasing semantic diversity, reviewed decisions, hard negatives and held-out source families—not by treating a million repeated rows as a million independent meanings.

Invariants that do not disappear when the model improves

  • Source-native terms and lexical values remain recoverable beneath every semantic candidate.
  • Mapping relation remains separate from human admission status.
  • Confidence cannot establish identity, policy, publisher authority or canonical meaning.
  • Training and evaluation split by source family or origin to prevent vocabulary memorization from masquerading as generalization.
  • Unknown, withheld, disputed and incompatible states stay explicit.
  • Model artifacts, code revisions, data digests, calibration and failures are published together.

TNOE readiness report Ontology Workbench