DreamNet Evidence Lab

Inspiration

An agent saying "done" should be the start of inspection, not the end. We want a small workflow that another builder can run, reopen and challenge without access to DreamNet's private infrastructure.

What it does

Evidence Lab is a public demonstration of an existing Cerberus-to-Memory Weaver integration. It inspects synthetic repository text as inert data, retains source manifests and denied decisions, builds portable source-linked context and checks the saved result. This is a reproducible slice of DreamNet, not a new claim that the entire agent ecosystem runs through this demo.

The demonstrated workflow

Create two synthetic text files; do not execute repository code.
Inspect them with Cerberus and record source digests.
Change an input and show the earlier digest no longer describes it.
Inspect an inert install-hook example, which receives Red / denied.
Compile the original sources and inspection context with Memory Weaver.
Save the bundle and archive, reopen them, and verify source, receipt and semantic consistency.

Synthetic fixture -> Cerberus -> inspection context
                                      |
                                      v
                               Memory Weaver
                                      |
                                      v
                         bundle + portable archive
                                      |
                                      v
                         reopen + independent replay

What was actually executed

The October 18 local run verified three sources, three inspection receipts and three relationships. Its archive was reopened in the public Memory Weaver application. The focused regression suite passed 12/12 tests, including source drift, altered receipts, missing provenance and changed engine bytes.

The fixture sources stay labeled synthetic. The experiment is real execution of deterministic public modules, not a recording of a model-led agent or a production fleet. Review the linked public artifacts and reproduction instructions below.

How we built it and how to reproduce

The public runner uses Node.js with reviewed JavaScript/TypeScript engine modules and source locks. Start with the full reproduction guide. The guide includes exact engine revisions, setup and test commands, the saved example JSON, the Capsule and an openable workspace.weave.json archive.

No account, API key, wallet, provider call or package installation is needed for the local reproduction. Initial cloning downloads public source; the demo itself runs offline. The browser application is a separately hosted interface.

Challenges and lessons

Missing authenticated origin must stay missing. The initial fixture receives Orange / denied despite having no static findings. An unkeyed hash can establish internal consistency, not an authenticated publisher or the truth of a claim. Recomputing the outer bundle hash must not conceal altered internal receipts.

Limits and event status

Receipts in this demonstration are unsigned; browser FNV-1a identifiers are not cryptographic signatures. This is not a Google Cloud deployment, Gemini runtime, model benchmark, trading result or measured credit-savings demonstration.

This project was an unfinished draft associated with All Things Agentic. Its submission window is closed. This October 18 portfolio completion does not establish a timely contest entry or satisfy the event's separate model, framework, cloud and video requirements. The event record remains separate.

Next steps

Use the lab to review additional authorized fixtures and real-repository handoffs, then add paired quality and resource measurements. Keep consequential execution behind explicit permission checks rather than promoting a report into authority.

Links

The private dream-net core is not needed and remains private.

Built With

typescript