Skip to main content
Authority: Experimental explanatory reference · Status: Architecture scaffold only · Applies to: E4.7 · Verified against: current checkout · Source: internal/tnoealign, ADR 019, reports/e4-7-tnoe-align.md

​
Purpose

TNOE is the future TWIRX-native system for proposing correspondences between source-native structures and the explicit ontology. It is separate from both the frontier conversational agent and the deterministic data plane.

​
Data model

The experimental language-neutral contracts are:
  • schemas/cddl/tnoe-alignment-graph.cddl;
  • schemas/cddl/tnoe-candidates.cddl;
  • schemas/json/tnoe-candidates.schema.json.
The current deterministic builder creates native schema and field nodes, ontology module/concept/frame/role nodes, structural edges, and candidate alignment edges. Every alignment edge preserves the Semantic Mapping Unit, mapping relation and admission status as separate fields. Candidate heads are bounded to concept, frame role, mapping relation, compatibility, abstention, identity risk, and security anomaly. A candidate artifact must state authority: candidate_only and canonical_admission: false.

​
Current graph

The tracked Opportunity scaffold contains:
These counts describe a graph contract, not a trained neural model.

​
Roadmap to a trained candidate

TNOE proceeds through evidence gates rather than a calendar-only promise:
  1. Harden the compact MCP and evidence-bound agent-answer contracts.
  2. Run a bounded same-frontier-model AgentBench pilot after spend and endpoint approval.
  3. Reach at least five source families, 250 unique Semantic Mapping Units, 500 signed decisions and 1,000 reviewed hard negatives.
  4. Train TNOE-Align 0.1 with source-family-held-out evaluation, calibration and explicit abstention.
  5. Expand across World State, research, API, vulnerability and public-rule structures before claiming broad generalization.
  6. Replace frozen text features only when measured data and compute justify an independently evaluated TWIRX-native encoder.
The powerful-model objective is correspondence and structural learning across the Web. It never changes the invariant that explicit ontology, source evidence and signed review—not model weights—carry canonical authority.

​
Invariants and failure behavior

  • Model output never modifies a mapping decision or ontology module.
  • Relation and admission status remain orthogonal.
  • Graph arrays are deterministic and strictly ordered.
  • Unknown fields, duplicate identities and invalid authority fail closed.
  • Frozen encoder output, when later admitted, is a feature—not a label.
  • Signed Workbench decisions are the only eligible review labels.

​
Security considerations

Future model code runs offline, without secrets, network, deployment access or canonical write authority. Artifact hashes and exact model revisions must be checked before loading. Source text remains untrusted data and cannot issue instructions to the worker.

​
Conformance and implementation status

Run:
The graph and candidate parsers have malformed-input fuzz targets. TNOE training is currently blocked: only 15 unresolved Semantic Mapping Units, zero signed mapping decisions and zero reviewed hard negatives exist. No TNOE accuracy, calibration or generalization result is claimed. The public status and full gate sequence are presented at /model/; semantic diversity and review state are presented at /ontology/ on twirx.org.
Last modified on August 13, 2026