Interpret the Web once; reuse the result with proof
TWIRX is a Semantic Data Plane for public web representations. It continuously separates what an origin represented from how that representation is interpreted, then publishes reusable typed state that agents can query and trace.
It is not a search engine, browser wrapper or universal truth oracle. Search and browsers may help discover representations. TWIRX’s distinct job is to compile admitted representations into versioned semantic state without discarding their source-native meaning.
The end-to-end flow
- Atlas and policyIdentify an origin, record interfaces, effects and access constraints, and require an explicit decision before retrieval.
- ObservationA route-bound worker retrieves bounded bytes. Evidence is stored before parsing.
- Native extractionPreserve the source term, exact lexical value and locator.
- Semantic compilationCreate typed packets whose mappings, context, versions and proof links remain explicit.
- Immutable log and snapshotPublish packets, views and deltas under a canonical manifest digest.
- Query fabricResolve typed queries against admitted materialized state, retaining disagreements and unresolved results.
- Agent interfacesExpose a compact query, trace, explain, compare and subscribe surface rather than one tool per origin.
Authority only moves through admission
Untrusted inputs may propose
- Origin responses and archive captures
- Adapters and mapping candidates
- Browser or model output in future gates
- Publisher capability declarations before verification
Only admitted artifacts may serve
- Explicit policy decision and bounded work order
- Validated observation and evidence digest
- Reviewed adapter, mapping and semantic closure
- Canonical manifest published last
No component promotes itself
An origin, parser, adapter, model, browser, payment signal or verifier can produce or reject a candidate. None can grant itself policy approval, publisher authority, semantic canon status or execution authority.
Compilation is separated from query execution
Compilation path
Rare, reviewed and potentially expensive.
- Policy and route admission
- Bounded retrieval or archive acquisition
- Native structure and schema analysis
- Mapping review and packet generation
- Delta and view construction
Query path
Frequent, deterministic and read-only.
- Typed semantic request
- Materialized-state lookup
- Trust, freshness and source filtering
- Conflict-preserving merge
- Proof-linked result
The recorded cross-origin query returns 4 rows from TWIRX and World Bank materialized state with 0 origin calls. That is a small proof of this separation, not yet a broad production benchmark.
Current deployment model: immutable snapshots
The current shared VPS is treated as a replaceable edge, not the authoritative mutable data plane. Canonical state is compiled off-host into a deterministic snapshot, verified by its manifest digest, and activated atomically by a read-only runtime.
approved evidence
→ local deterministic compiler
→ packets + deltas + views + proof index
→ canonical manifest
→ snapshot_id = SHA-256(manifest)
→ read-only query runtime| Snapshot ID | sha256:54739822257ef617b136454285a8fd47802f0960c7cf53a49abd2d5d1f1389c5 |
|---|---|
| Public-source packets | 15 across 3 origins |
| Materialized views | 2 |
| Deltas | 1 real origin delta |
| Runtime writes | None to semantic state |
The protocol is language-neutral
Normative behavior lives in specifications, schemas and conformance vectors—not in one implementation language. The current trusted implementation uses Go; independent bounded verification uses restricted C. Those are implementations of the protocol, not the protocol’s authority.
Deterministic CBOR/CDDL artifacts, shared positive and adversarial vectors, and explicit error behavior keep independently written implementations comparable.
Deliberately deferred
Continuous production state
No authoritative PostgreSQL is deployed on the current shared host. Durable independent infrastructure and restore evidence come first.
Browser discovery
No browser runs in the trusted execution path. Future discovery requires separate isolation.
Model induction
No model creates canon or execution authority. A future model may rank candidates only after an adjudicated corpus exists.
Actions and payments
Commercial metadata may be described; no production purchase, payment, booking or write action is authorized.
Arbitrary URLs
Public callers cannot supply destinations. Retrieval requires an admitted origin and route-bound work order.
Federation
Signed packages and multi-node registries remain future work after the single-node proof is durable.