VoxShield Console

Detect. Decide.
Respond.

The Console is where a verdict becomes a decision. It runs on one laptop, offline, with one process and one GPU worker — and it is honest on screen about which of its answers it actually trusts.

How access works The pipeline behind it

There is no hosted Console to sign in to. It binds to 127.0.0.1:8000 on the machine that runs it, by design — that is what “offline” means here. This page is the tour; the running system is on the demo laptop.

Inside the Console

What is on screen today

The v1.0 console, built and tested. Each view is plain HTML and vanilla JS.

Dashboard

Built

Live call cards in all five verdict states, signal bars ordered by fusion weight, a sparkline with real gaps where windows were refused, and the risk-profile selector.

Alerts

Built

Verdicts that crossed a threshold, with the reason code and the prescribed action beside them.

Analytics & reports

Built

Verdict distribution over time, including how often the system refused to answer — which is a number worth watching.

Settings & audit

Built

Enrolments, retention, and a verify-audit panel that reports either an intact chain or the sequence number where it broke.

System

Built

Model hashes, device and VRAM, offline indicator, and the health endpoint's report of which alert mode is live.

Assistant

Built

Grounded Q&A with citations, plus a ? control on each verdict card that pre-fills the question for that card's reason code.

Enterprise layer

The layer that wraps the core

Multi-tenant sign-in, mobile recordings, incidents and mitigation. It never edits the detection core and never widens a claim the core makes. Dashed links below are specified, not built.

Built Detection core, unchanged by this layer Specified, not built
Sign-in, roles and tenant scoping are implemented. Mobile recording ingest, risk events, incidents and the alert outbox are written up in the enterprise specification and are not built yet — so they are drawn dashed rather than shown as working.

Enterprise sign-in

Four steps, checked in the backend

Seven auth endpoints are implemented. Authorisation is resolved in a dependency before the handler body runs — a hidden button is a convenience, not a control.

  1. Enterprise identify Built

    An email domain resolves to its enterprise. The lookup is one of only two cross-tenant queries in the system, and it is marked as such in the source.

  2. OTP request and verify Built

    Codes are generated with secrets, compared with hmac.compare_digest, rate-limited and expiring. The plaintext code is never stored. A timeout is a failure, never a pass.

  3. Employee login Built

    Password (scrypt) or OTP, then a session token that is stored only as a hash. Sessions can be revoked, and there is an absolute cap on lifetime.

  4. Role-based console Built

    Platform admin, enterprise admin, security, bank and agent. Every tenant query takes the enterprise id first, and CSRF is checked inside the principal resolver.

The full permission matrix

Build state

What exists, honestly

The project distinguishes code that is written and tested from work a human has accepted against its criteria. That distinction is kept here rather than flattened into “done”.

AreaStateDetail
Detection pipelineBuiltIngest, VAD, gate, signals, fusion, verdicts, profile actions.
Audit chainBuiltHash-chained, with three corruption paths covered by tests.
API surfaceBuilt17 endpoints — 14 REST, 3 WebSocket — one identical JSON object.
Console viewsBuiltDashboard, alerts, analytics, reports, settings, system, assistant.
AssistantBuiltGrounded retrieval with citations, refusal discipline, 28 tests.
Enterprise schema & tenancyBuilt16 tables, idempotent migration, every repository function scoped by enterprise.
Auth, sessions, RBACBuilt7 endpoints, scrypt, CSRF in the principal resolver, 36 security tests.
Alert outboxIn progressCore built and carrying OTP delivery; incident templates and the redaction test remain.
Admin APISpecifiedEnterprises, employees, domains and contacts. Not built.
Mobile device ingestSpecifiedDevice tokens and recording upload into the existing pipeline. Not built.
Mitigation engineSpecifiedConsumes the existing action matrix rather than defining a second policy. Not built.
Verification & incidentsSpecifiedStep-up OTP, incident status machine, append-only notes. Not built.
Human acceptance passOutstandingCode is tested, but the walk-through against the acceptance criteria has not been signed off.

Access

How to actually open it

There is nothing to sign up for. Download the Console with its models included and run it locally.

Download Console (160 MB, models included)
# Unzip voxshield-console.zip, then from that folder
make setup      # virtual environment + dependencies (models are already bundled)
make run       # 127.0.0.1:8000, one worker

make test       # 351 passed, 1 skipped
make coverage   # 91.97%
make audit      # 21 permissive deps, 15/15 model hashes

At the demo

In person

A mock telephony tool replays a WAV in real time into the Console, which produces one verdict per second, renders the correct card state, and appends every decision to the audit chain — with the network disconnected.

For evaluators

Reproducible

Nine benchmark harnesses regenerate every number this site publishes, on the reference machine named on the benchmarks page.

Next

See the evidence behind the screen

Measured numbers, dataset by dataset — including the ones that argue against us.

Benchmarks Security & privacy