The open typed interface layer for the agentic web.
TWIRX compiles public web representations into deterministic, typed interfaces whose native meaning, evidence, and derivation remain inspectable.
Follow one real value Gate E1 evidence Documentation Support Genesis
- Origin 162 B
- Evidence sha256:e194e7dfd8…
- Observation sha256:98025ae34c…
- Verified Go + C
- Native usd
- Transform trim → uppercase
- Typed result USD
usd; the agent receives
USD; the derivation between them is part of the
result. Open the provenance explorer to inspect every stage.
Evidence before interpretation
TWIRX does not certify that a provider's statement is objectively true. It records what an origin returned, what was observed, where the source statement appeared, and how declared transformations and mappings produced the agent-facing value.
That is a narrower claim than most systems in this space make, and it is the reason the rest of the architecture holds together. A website may be accurate, mistaken, outdated, satirical, compromised, or in disagreement with another source. Typing a statement does not turn it into a fact, and committing a hash to a blockchain does not make the external proposition true.
What the system claims fidelity to
- Origin representationThe bytes a public origin actually returned, at a recorded time, under a declared access policy.
- ObservationAn immutable record binding those bytes to their retrieval metadata, encoded deterministically.
- Source statementWhat the provider said, in the provider's own term and lexical value, at an identified locator.
- Extraction assertionThe declared operation that located the statement, executed offline against stored evidence.
- Declared transformationEvery normalisation applied, named and versioned, never silently.
- Semantic interpretationA versioned mapping into shared meaning, recorded as a claim rather than an equivalence.
- Agent-facing resultThe typed value, carrying the whole chain above it as field-level provenance.
TWIRX claims fidelity to this derivation. It does not claim that the external
proposition corresponds to objective reality. Where a field cannot be interpreted
safely, the correct result is unresolved — not a plausible invention.
Engineering Gate E1 is implemented Implemented
E1 demonstrates a controlled public representation, bounded observation, SHA-256 content-addressed evidence, deterministic CBOR, shared Go and C validation, fail-closed JSON extraction, native-plus-semantic output, provenance-bearing unresolved states, atomic per-file publication, fuzzing, sanitizer checks, race testing, static analysis, and offline replay.
| Metric | Value | Scope | Source |
|---|---|---|---|
| Named Go test events | 940 failed | Seven tested packages, one recorded host | reports/gate-1-genesis.md |
| Shared Go/C observation vectors | 162 accepted, 14 rejected, in both implementations | Committed vectors; identical inputs and expectations | conformance/observation/vectors.json |
| Extraction vectors | 115 accepted, 6 rejected | Controlled fixtures only; not arbitrary websites | conformance/extraction/vectors.json |
| Independent implementations | 2Go and restricted C | Of the observation envelope and evidence check only; the adapter runtime still has one | reports/gate-1-genesis.md |
| Sanitizer-instrumented C libFuzzer runs | 5,0000 ASan/UBSan findings | Parser-only harness, one recorded host; a smoke run, not a campaign | reports/gate-1-genesis.md |
| Published unresolved risks | 7the complete recorded set | Every risk in the E1 report, including the inconvenient ones | reports/gate-1-genesis.md |
| Third-party runtime dependencies | 0Go standard library only; no go.sum | Protocol implementation; the website build is also dependency-free | reports/gate-1-genesis.md |
| Hosted CI executions | 0check-suite startup failed before any job existed | No hosted CI execution is claimed anywhere on this site | reports/gate-1-genesis.md |
The implementation is deliberately narrow. It does not yet expose arbitrary websites, browser discovery, model inference, actions, payments, or public multi-tenancy.
The exact CI position
Engineering Gate E1 passed the complete local validation suite recorded in the Gate E1 evidence report. GitHub-hosted Actions failed during check-suite startup before creating any job; no hosted CI execution is claimed.
Read the full Gate E1 evidence record, including all seven unresolved risks, or open the provenance explorer.
Agents should not have to rediscover the Web one browser session at a time
The Web's dominant interface was designed for human perception: pages, layout, navigation, forms, visual hierarchy, browser state, gestures. That interface transformed civilisation. It was not designed for a reader that arrives a thousand times a day and needs to explain where its answer came from.
So software agents improvise. They imitate human browsing, infer meaning from presentation, repeat expensive navigation, collapse observation and interpretation into a single opaque step, and frequently lose the evidence that would let another person or machine check the result. Extraction built on presentation is brittle by construction: it breaks when a layout changes, and it fails silently when a selector still matches but now means something else.
The alternative to that is not obviously better. A machine-readable Web is arriving regardless, and the default path leads to a small number of private interfaces controlled by a few model providers, browser runtimes, and clouds — meaning inferred invisibly, provenance reduced to a link, independent agents dependent on proprietary interpretation, and publishers facing uncontrolled automation with no credible way to declare what machines may read or do.
TWIRX is an attempt at the open path: a semantic and evidentiary layer that can be implemented independently, inspected locally, and eventually published directly by the websites it describes.
Presentation is not an interface
An agent reading rendered layout is reverse-engineering a human artefact. The meaning it recovers is an inference, and inferences need provenance.
Enclosure is the default outcome
Whoever defines the machine-readable schema of the Web defines what agents can see. That should not be a private decision.
Publishers have no lever
Today a publisher can block automation or tolerate it. There is no standard way to say: here is the authoritative typed interface to my content.
Two paths, deliberately separated
Compilation and execution have opposite requirements. Conflating them is what makes browser-driven agents both slow and unauditable. Every gate that adds intelligence adds it to the left-hand path, behind admission — never to the path that runs a million times.
Compilation — the slow path
Runs rarely. Allowed to be expensive, and to be wrong, because a human admits its output.
- Isolated browser discovery PlannedE4
- Model-proposed candidate structures ResearchE5
- Semantic and mapping proposals Specified—
- Review and admission Implementedby hand, today
Execution — the fast path
Runs constantly. Must be boring, deterministic, and readable by a reviewer.
- Direct structured access ImplementedE1
- Compiled extraction plan ImplementedE1
- Bounded fail-closed validation ImplementedE1
- Provenance assembly ImplementedE1
- Generated agent bindings PlannedE2
No browser, model, open-ended reasoning, or dynamic code generation exists in the fast path today, and none is planned for it. Whatever cleverness produced an adapter, the thing that actually runs is a declared plan a reviewer can read. See the full architecture.
Founding properties
Six commitments that constrain every later gate. A gate may not weaken one without a public protocol decision and a migration path.
Evidence-native
Every resolved value identifies the observation, source bytes, locator, adapter, transformations, mapping, and semantic closure that produced it.
Native meaning preserved
The provider's original term and lexical value remain available beside the shared semantic view. Normalisation is a view, not an erasure.
Language-sovereign
Specifications, canonical encodings, and public vectors define the protocol. No implementation language, runtime, or package ecosystem owns it.
Hostile-world security
Untrusted origins, adapters, browsers, models, and contributors may observe, extract, infer, and propose. They cannot grant themselves canonical authority.
Local-first
The Genesis proof runs and replays locally with no cloud service, browser, model API, chain, or live origin after observation.
Open commons
Self-hosting, competing implementations, federation, extension, and the right to fork remain part of the design. There is no protocol token.
What exists, and what does not
The fastest way to lose a reader's trust is to describe an accepted architectural direction as though it were already running. Both columns below are drawn from the same machine-readable file the pages are built from.
Implemented now Implemented
- Controlled read-only observation
- Content-addressed evidence
- Deterministic observation encoding
- Independent Go and C verification
- Bounded fail-closed JSON handling
- Deterministic manual adapter
- Native and semantic views
- Provenance-bearing unresolved fields
- Adversarial conformance corpus
- Atomic per-file publication
- Offline replay
Not implemented Planned
- Arbitrary website compilation
- Isolated browser discovery
- Model-assisted schema induction
- Public adapter registry and admission
- Publisher verification and federation
- MCP, WebMCP, and OpenAPI generation
- Typed Web intermediate representation
- Semantic module compiler and kernel
- Scoped actions, receipts, and payments
- Signed adapters and release provenance
- Blockchain anchoring
- Production multi-tenancy
Source: implementation-matrix.json. Each row on the right carries its own status and a note on the progress page.
Sustain the next public gates
TWIRX is seeking USD 6,000 in transparent Genesis support for maintainer runway, coding-agent and model access, compute, isolated worker infrastructure, evidence storage, monitoring, documentation, security work, and administration.
Contributions fund public work. They do not purchase equity, protocol ownership, governance rights, private technical votes, preferential canon admission, or a financial return. There is no token.
Maintainer runway is infrastructure
The first infrastructure dependency of this project is sustained human work. Maintainer food, housing, communications, and uninterrupted engineering time are therefore represented openly in the budget rather than hidden inside a server invoice. Private household records remain private.
Help build an open agentic Web
Specific help is more useful than general encouragement. The project needs:
| If you are | Start here | What would help most |
|---|---|---|
| An engineer or independent implementer | Proof, then Architecture | Reproduce Gate E1 locally and tell us where it fails on your host. |
| A security or parser reviewer | Security | Attack the bounded parsers and the URL policy. The corpus is deliberately small and wants adversaries. |
| A semantic-web or ontology researcher | Semantics | Review whether mappings-as-claims survives contact with real vocabularies. |
| A publisher or public-data institution | Governance | Tell us what a publisher would actually need in order to declare an authoritative interface. |
| An infrastructure sponsor or funder | Funding | Compute, storage, security tooling, a grant introduction, or direct Genesis support. |
| A documentation or accessibility reviewer | Documentation | Find the places where this project explains itself badly. |
A machine-readable web will exist. TWIRX exists to help ensure that it remains open, evidentiary, plural, independently implementable, and accountable to the people and publishers who make the web worth accessing.