Safety and Security Framework | Dweve Trust Centre

Inspect Dweve model, tool, deployment, and operator control layers, the safety lifecycle, threat register, and coordinated disclosure route.

Documented|||Internally confirmed against a current Dweve record, implementation or owner-provided fact as of the displayed date; not a claim of independent audit or certification

Pre-release|||Built or prepared for the planned external release

Future event|||The method and evidence contract are current; the named real-world event has not yet occurred

Ask for missing information, report a concern, challenge a copyright decision, or tell us where a public record is unclear. The message enters the same recorded contact route.

Trust is not a promise|||it is built into the system

The public evidence behind Dweve Loom, its data route, operating boundaries and release controls. Documented means internally confirmed as of the displayed date; independent assurance is named separately when it exists.

Inspect the public record|||Open the Loom record

The commitment is public|||the evidence stays up to date

Dweve is a voluntary signatory to the EU General-Purpose AI Code of Practice and participates in the Signatory Taskforce. This dossier explains what that commitment means in Dweve's architecture and where the evidence is kept.

Read our position|||Inspect the change record

One model family|||one accountable record

Dweve Loom 1.0 is the only Dweve model. The registry separates its adaptive model identity, captured states, access routes and factual release state without treating Loom as one static hash.

Open the registry|||Read the Loom record

A model is one thread|||Loom is the weave

Loom is built around constraint learning and a typed execution graph. It combines perception, memory, retrieval, domain specialists, solvers, and verification. This record states what changes, what is sealed, and what can be replayed.

Read the architecture|||Inspect training content

Every source family is named|||every admission still has to prove itself

The register publishes 772 supplied training and knowledge-source entries across 19 groups. Each entry states its licence gate, usable slice, and purpose. Exact releases enter Loom 1.0 only after their rights, provenance, privacy, and evidence checks pass.

Inspect the source register|||Read the copyright policy

Training register|||772 sources · current

Collection starts with rules|||not with a fetch

Winnow checks source policy, robots.txt, rates, targets, and request records. Spindle governs provenance and promotion. Loom consumes governed material and can use Winnow as a tool.

Inspect the controls|||Contact the rights desk

The model record travels|||with the integration

Downstream teams need current capabilities, limits, interfaces, evaluation context, and changes. Dweve publishes the common record first and scopes private details only where necessary.

Inspect the package|||Request more information

Every release carries|||a verifiable origin

Dweve's release-bound policy covers Loom text and code output plus the image output Dweve ships. Loom does not generate video. C2PA Content Credentials carry the machine-readable provenance record, and route-specific validation runs before an external route opens.

Inspect the marking policy|||Report a model issue

Output transparency|||Release-bound controls

Capability needs authority|||before action

Dweve tools are built for legitimate, governed work. A technically possible action is not automatically permitted. Purpose, authority, evidence, human responsibility, and deployment policy remain part of the execution boundary.

Send the concern once|||we route the record

Copyright, downstream information, model safety, security, privacy, and model-quality concerns enter one recorded route. Choose the topic, preserve useful evidence, and receive a submission reference.

Choose a reporting route|||Return to the Trust Centre

No silent revisions|||no missing history

Material changes to the model record, source policy, public controls, and machine-readable files are listed here with their effect. Human and machine feeds carry the same version.

Read the current release|||Open the JSON record

GPAI Code signatory|||Dweve signed voluntarily and publishes the implementation path

Signatory Taskforce|||Dweve participates in the working forum for signatories

External access|||Invite-only external access is planned for 1 September 2026

Training source register|||772 supplied source entries are public; exact release admissions remain evidence-bound

Full voluntary scope|||Transparency, copyright, safety, and security measures are included

The hub is organised by the question a reader needs answered. Model identity is separate from training content. Copyright controls are separate from acceptable use. Each record exposes its own status.

How to read the status|||Public evidence map|||Ten records|||Missing fields stay visible

AI transparency|||Signatory position, voluntary scope, governance, and claim boundaries

Loom model record|||Architecture, modalities, state, replay, and operating data policy

Training content|||772 supplied source entries, deterministic admission, and release-specific evidence status

Downstream documentation|||The common integration and evaluation information package

Generated content|||Output scope, trace attachment, and marking controls

Acceptable use|||Purpose, authority, refusal, review, and appeal boundaries

Reporting desk|||One route for copyright, safety, security, privacy, and model concerns

Change record|||Version history in HTML, JSON, and Atom

Compliance cost|||Scope, period, inclusions, and exclusions for the compliance programme

From an outside source|||to an accountable output

Winnow acquires under a declared source policy. Spindle preserves provenance and decides what becomes governed knowledge. Loom uses that state and can call Winnow for current information.

Three control layers|||Source-to-output route|||Typed hand-offs|||No hidden scrape path

Access policy|||Robots, rates, target safety, and operator rules apply before collection

Provenance|||Source, address, time, status, and content identity stay attached

Runtime boundary|||Loom receives governed material or invokes Winnow through a tool boundary

Winnow|||Discovers, fetches, parses, and records according to source rules

Spindle|||Assesses provenance, evidence, quality, disagreement, and promotion

Loom|||Selects the relevant cognitive weave and records what it used

Output record|||The answer carries its trace, sources, state, and execution context

Loom is currently in internal-only testing. External invite-only access is planned for 1 September 2026 across individual and combined product routes. The public record makes the remaining release-readiness steps explicit.

Internal now|||Invite-only next|||Versioned after

Current posture|||Release status line|||Three states|||Dates are not hidden

Internal testing|||Dweve teams operate the current pre-release system internally

External invite-only|||Testers, partners, investors, and selected participants enter by invitation

Post-market record|||Usage, incidents, evaluations, and changes continue after access opens

Now|||Internal-only pre-release testing with public records being prepared

Before 1 September|||Complete the external-access controls and keep each admitted item bound to its controlled Spindle record

From 1 September|||Invite-only external operation with versioned monitoring and change publication

A public record needs|||a public way to challenge it

The reporting desk separates the subject while keeping one submission route. Sensitive material should be described first, not pasted into a public or general-purpose field.

Reporting posture|||Choose the right desk|||Recorded route|||One submission, clear owner

Public entry point|||Every route starts from the same accessible form

Submission record|||Successful delivery returns a reference from the contact system

Subject ownership|||The selected topic determines the internal owner and evidence request

Rightsholder request|||Raise a source, reservation, removal, or copyright concern

Downstream request|||Ask for integration, evaluation, capability, or limitation information

Model safety|||Report harmful behavior, unexpected capability, or a failed control

Security|||Report a vulnerability without exposing exploit details publicly

Voluntary participation|||without invented classification

Dweve records Loom as a general-purpose AI model under Article 3(63). The Commission's 10²³-FLOP criterion is indicative, not a safe harbour. No public designation of Loom as a GPAI model with systemic risk has been found as of 1 August 2026.

Signatory|||Taskforce|||Voluntary full scope

Status boundary|||Dweve public position|||Current record|||Precise, not overstated

Voluntary signatory|||Dweve has chosen the Code path before external release

Taskforce participant|||Dweve contributes through the official signatory working forum

No public designation found|||No public Commission record currently states that Loom has been designated as a GPAI model with systemic risk

Code implementation|||Public model, training, copyright, downstream, safety, and monitoring records