Tools

Prompts

Connectors

Resources

Knowledge Graphs

Skills

Extensions

Components

Context

Context window composition for the active profile.

Updated when profile changes or after each query.

Workspace

EVIDENTIA

Police Investigation Platform

Intelligence-Led Investigation.
Chain-of-Custody Trust.

Governed Investigation Data

Ask questions of your force's data products in plain language. Every answer comes from governed, signed data — never from guesswork.

Chain-of-Custody Trust

Every answer carries a cryptographically signed trace of how it was produced — verifiable offline, exportable as a receipt, and honest about what the data does and does not support.

Protected by Design

Personal data stays protected at the database itself — role-enforced views, automatic redaction, and guardrails that screen every answer before it is delivered.

Live Status

Configuration

No Model/Prompt Loaded

Last Stmt

In: 0

Out: 0

Last Turn

In: 0

Out: 0

Session Total

In: 0

Out: 0

Filter

Waiting for a new request...

Awareness

Your agents’ session operations

Auto-refresh

Total Sessions

Total Tokens

0 in / 0 out

Success Rate

Est. Cost

$0.00

Velocity (24h)

Model Distribution

No data available

Most Token-Intensive Sessions

No session data available

Most Token-Intensive Questions

No question data available

Session Roster

Status

Sort

Loading sessions...

Session Details

Total Turns

Total Tokens

Model

Planner Examples

Execution Timeline

Loading timeline...

Trust

Your agents’ trust, quality & reliability

Live scoring: —

Agents Evaluated

Healthy

Needs Attention

Live Turns Scored

Gating Failures

Loading agents…

Components

Your agents’ instruments & connectors

Category:

Status:

Ontology

What your world knows, who you can ask, and what to trust — always in sync.

Marketplace

Discover and subscribe to community-curated collections

Execution patterns and strategies for proven task completion Reference documents and domain knowledge for planning context Entity-relationship graphs for database topology, business concepts, and domain knowledge Pre-built profile bundles with knowledge repositories for instant deployment User-created post-processing extensions for structured output Pre-processing prompt injections for task specialization and context enrichment Ready-made component instances — flows and dashboards — to fork into your own workspace

Search Collections

Sort By

Visibility

Loading marketplace collections...

Intelligence

Your agents’ knowledge and expertise

Auto-refresh

Champion Strategies

Available for RAG

RAG Activation Rate

Tasks Enhanced

--%

RAG-Enhanced: --

Total Tasks: --

Plan Quality

Output Token Efficiency

$--

-- tokens saved

Speed Boost

From Token Efficiency

--%

With RAG: --s

Without RAG: --s

Planner Repositories

Execution strategies and planning patterns captured from successful agent interactions

Active only

Loading collections...

Planner Repository Constructors

Populate Planner Repositories using predefined templates

Inspect Collection

Loading rows...

ID ↕	User Query ↕	Strategy ↕	Efficient ↕	Feedback ↕	Output Tokens ↕	Timestamp ↕

Inspect Knowledge Graph

Anatomy

How your agents are built — composition & setup

Administration

User Tier Management

Assign users to profile tiers (User, Developer, Admin)

Username	Email	Current Tier	Consumption Profile	Status	Actions
Loading users...

User Tier

Developer Tier

Admin Tier

User Token Consumption

Monitor and manage user token usage and consumption profiles

Username	Consumption Profile	Tokens Consumed	Monthly Limit	Usage %	Status	Actions
Loading consumption data...

Total Input Tokens

December 2025

Total Output Tokens

December 2025

Active Users

With token usage

Avg Tokens/User

December 2025

Feature-to-Tier Configuration

Configure which tier is required for each feature

Feature	Description	Required Tier	Category
Loading features...

User Tier Features

Basic access features

Developer Tier Features

Advanced features

Admin Tier Features

System administration

Pane Visibility Configuration

Configure which tiers can access which panes

Pane Name	Description	User	Developer	Admin
Loading panes...

About Pane Visibility

Control which user tiers can see and access different panes in the application. Users will only see panes they have access to based on their assigned tier.

User Tier: Basic users with limited access
Developer Tier: Advanced users with development features
Admin Tier: Full system access and administration

Note: The Administration pane is always visible to admins only.

Consumption Profiles

Configure token limits and rate limiting for different user tiers

Profile Name	Description	Prompts/Hour	Prompts/Day	Input Tokens/Month	Output Tokens/Month	Users	Active	Actions
Loading profiles...

Total Profiles

Configured consumption profiles

Active Profiles

Currently enabled

Assigned Users

Users with profiles

User Interface

Configure panel visibility, startup behavior, and MCP classification

Startup & Panel Defaults

Default panel visibility and welcome screen settings at startup

These settings define the default visibility state for panels when users first load the application. Individual users can override these defaults using their personal panel toggles, and their preferences will be saved locally.

Session History Panel (Left Sidebar)

Panel is visible for the user

When enabled, the session history panel is available to users

Panel default visibility mode

Initial state when entering conversation pane

User can collapse/expand

When enabled, user can toggle the panel open/closed

Resources Panel (Right Sidebar)

Panel is visible for the user

When enabled, the resources panel is available to users

Panel default visibility mode

Initial state when entering conversation pane

User can collapse/expand

When enabled, user can toggle the panel open/closed

Live Status Window (Bottom)

Panel is visible for the user

When enabled, the live status window is available to users

Panel default visibility mode

Initial state when entering conversation pane

User can collapse/expand

When enabled, user can toggle the window open/closed

Always Show Welcome Screen at Startup

When enabled, the welcome screen will always be displayed in the conversation pane at startup or after browser refresh, regardless of whether the application is configured for the user.

Default Theme

Select the default theme that users will see on their first visit. Users can change their theme preference, and their choice will be saved.

Allow Users to Override Panel Settings

When enabled, users can toggle panels and save their preferences. When disabled, the above default settings are enforced for all users and panel toggle controls are hidden.

Document Upload Configuration

Configure document upload capabilities for LLM providers

About Document Upload

Control how documents are processed for each LLM provider. Some providers support native document upload (sending files directly), while others require text extraction.

Native Upload: Google, Anthropic, Amazon (better quality, images supported)
Text Extraction: OpenAI, Azure, Friendli, Ollama (fallback method)
Override settings to force text extraction or adjust file size limits

Provider	Capability	Enabled	Native Upload	Max Size (MB)	Status	Actions
Loading configurations...

Capability Levels

native_full Full document support (PDFs + images)

native_vision Vision models only (images)

text_extraction Fallback text extraction

not_supported No document support

MCP Resource Classification

Configure LLM-based classification of MCP capabilities

Enable Automatic Classification

When enabled, MCP capabilities are automatically categorized during configuration using the LLM. When disabled, generic categories are used for faster configuration.

Classification Status

Ready to classify

Security & Access

Configure rate limiting, session management, and security policies

Security Settings

Session and authentication parameters

Session Timeout (hours)

How long sessions remain valid (default: 24)

Token Expiry (hours)

Authentication token validity period (default: 168)

Rate Limiting

Configure system-wide rate limiting policies

Enable Rate Limiting

When enabled, enforces request limits to prevent abuse. Recommended for production and multi-user deployments. Disabled by default for single-user installations.

Global Override

When enabled, the global limits below override individual consumption profile settings for all users. Use for emergency rate limiting.

Per-User Limits (Authenticated)

Prompts per Hour

Prompts per Day

Config Changes per Hour

Per-IP Limits (Anonymous/Authentication)

Registrations per Hour

API Calls per Minute

About Rate Limiting

Rate limiting protects your application from abuse and ensures fair resource allocation across users.

Changes take effect within 60 seconds (cache refresh)
Authenticated users tracked by user ID, anonymous by IP address
Uses token bucket algorithm for smooth rate limiting
Disable for single-user installations or trusted environments

Trust

The deployment's assurance controls, in the order you set them up: measure, decide what counts as failing, decide what acts, decide what reaches you.

Scoring

Whether turns are measured at all, and whether LLM judges take part. PROVEN's receipts ride the same hook — sealing is on exactly when automatic scoring is.

Automatic Agent Scoring

Deterministically score every executed turn and flow run (free, non-blocking) and feed the Agent Performance roster. Turn off to score only on demand.

Include LLM Judges

Add the LLM-judge panel to automatic scoring. Judges call an LLM (token cost), so they run only on a sampled fraction of turns plus all anomalous ones.

Judge sample rate: 0.05

Fraction of turns also judged (1.0 = judge every turn). Anomalous turns are always judged when judges are on.

Quality bands

Where a check counts as healthy, worth a look, or a failure. Every measurement is counted at JUDGED, grouped by the moment of the turn it measures.

Preset:

Quality bands — where green ends and amber begins

Green at or above Amber at or above

Blast radius: these two floors colour every pass rate in the product — the Agent Roster bars, the Trust Circle's JUDGED reading, and the cross-agent KPI charts. The other seven stations report what their mechanisms did and are not coloured by a measurement, so these floors do not move them. Raising the green floor will turn currently-green agents amber immediately; it changes how quality is reported, never how a turn is scored or gated. Defaults 0.80 / 0.50. Only the numbers are configurable: whether a gate acts (enforce vs observe), and worst-of aggregation across agents, are properties of the system rather than judgements about where a bar sits.

Plan

Retrieve

Act

Compose

The run

Enforcement

Which controls may act on a turn. Each agent's own mode is set in the profile editor's Trust Configuration; these master switches govern the whole deployment.

Grounding Enforcement

Master switch for the grounding gate. When on, each profile's Grounding mode applies (default observe: surface KG scope drift without changing answers; annotate/enforce are opt-in per profile). Off → the gate never runs. The checks that GRADE this guarantee are tuned under JUDGED, with every other measurement — a threshold tunes a measurement, and measurement is counted in one place.

Guardrails (P2)

Master switch for the deterministic guardrail. When on, each profile's Guardrail mode applies (default observe: detect outbound PII & toxicity without changing answers; annotate/enforce — PII redaction, toxicity withhold — are opt-in per profile). 0-token. Off → the guardrail never runs.

Governed Data (OBDA)

Master switch for the OBDA observability lane — the Governed data mechanism under GROUNDED. When on, each turn's data operations are classified as governed (through the semantic layer) vs free-formed (raw SQL) and the governed lane is rewarded. 0-token, observe-only. Leave off until you deploy a governed-metric layer: OBDA is a preview (validate/execute/provenance are still to come), so with no governed metrics every turn reads free-formed and would keep GROUNDED amber. Off → the scan never runs, Governed data reads not applicable, and its scorecard check abstains.

Behavioral Trust Score

The ONE aggregate: fuses each agent's Trust-Membrane integrity (signed / changed / unsigned) with its live conduct — a recency-weighted blend of the trust-critical checks with every remaining check, each dimension counted exactly once — into one composite state (Provisional / At-risk / Watch / Proven / Trusted). Integrity can only cap conduct, never promote it. Read-side only: never gates a turn; surfaces the seal+ring on the Agents roster and the Trust Posture matrix. Off → the ring is hidden.

Circuit breaker

Trips only on the volume-gated transition into At-risk (one alert per bad streak). Freeze touches unattended use only (Agent Authority autonomous=false + frozen_by_breaker) — interactive use is never affected; visible and reversible in the Agent Authority card. Incidents land in the Assurance Inbox (Agents view).

Assessment modules

Which module assesses each property, during the turn and after it. One assessment, up to two lanes — changing a row moves both the gate's input and the score, together. Each change is applied and audited on its own.

Third-party modules

Packs may add properties and ways of assessing them. They never add a station, an enforcement act, or a binding — and installing one assigns nothing: you bind it above, and that is audited. A pack must be signed before it can be bound.

Allow third-party modules

What reaches you

Each station's reporting level. Danger always surfaces — this filters what is presented in the Assurance Inbox, the badge and the ring, never what is recorded.

SIGNED born, then attested

GROUNDED truth · faithfulness

GUARDED rules · safety

BOUNDED authority, borrowed

JUDGED every measurement, graded after the turn

TRUSTED the badge breathes

REVIEWED a human decides

PROVEN receipts, not promises

System Operations

Maintenance tasks and cache management

System Maintenance

Cache management and system operations

Clear Application Cache

Clears temporary data and cached responses

Reset Application State

Resets runtime state (requires reconnection)

Document Context

Limits for uploaded document attachments sent to LLM context

Total Document Context Limit

Total character limit across all uploaded documents (~tokens/4) (default: 50,000)

Per-Document Character Limit

Individual document truncation limit (default: 20,000)

System Restart

Notify all connected users before a manual platform restart

Notification Message

Countdown (minutes)

0 = no timer

Estimated downtime (minutes)

0 = unknown

System Prompt Editor

Manage system prompts for different profile tiers

Select System Prompt

System Prompt Content

0 characters

Prompt Parameters

Configure parameters used in this prompt template

Parameter	Type	Default Value	Override Value	Scope	Actions
Select a prompt to view parameters

Version History

View changes made to this prompt over time

Version	Date	Author	Description	Actions
Select a prompt to view history

Compare Base vs Override

Cost Analytics

Auto-refresh

Total Cost

$0.00

Avg / Session

$0.00

Avg / Turn

$0.00

Sessions

Cost by Provider

Top 5 Expensive Models

Model Cost Configuration

Manage LLM model pricing for accurate cost tracking

Default Fallback Cost

Used for models not found in the pricing table

Input (per 1M)

Output (per 1M)

Show Deprecated

Show per page:

Provider	Model	Input Cost (per 1M)	Output Cost (per 1M)	Source	Last Updated	Actions
Loading cost data...

Showing 0 to 0 of 0 entries

Page 1 of 1

Intelligence-Led Investigation. Chain-of-Custody Trust.

Approaching Consumption Limit

Almost at Limit

Consumption Limit Reached

Governed Investigation Data

Chain-of-Custody Trust

Protected by Design

Awareness

Model Distribution

Most Token-Intensive Sessions

Most Token-Intensive Questions

Session Roster

Session Details

Execution Timeline

Planner Repository Cases

Trust

Components

Ontology

Marketplace

No Collections Found

Intelligence

Champion Strategies

RAG Activation Rate

Plan Quality

Speed Boost

Planner Repositories

Planner Repository Constructors

Knowledge Repositories

Knowledge Repository Constructors

Knowledge Graphs

Knowledge Graph Constructors

Inspect Collection

Corpus Health

Selected Case Details

Inspect Knowledge Graph

Anatomy

Configure MCP Servers

Vector Stores

Add Vector Store

Knowledge Graph Backends

Add KG Backend

Configure LLM Providers

Context Window Types

Installed Context Modules

Utilization Analytics

Agent Memory

Configure Agent Profiles

Skills

Extensions

Agent Packs

Advanced Settings

Access Token Management

Administration

User Tier Management

User Token Consumption

Total Input Tokens

Total Output Tokens

Active Users

Avg Tokens/User

Feature-to-Tier Configuration

User Tier Features

Developer Tier Features

Admin Tier Features

Pane Visibility Configuration

About Pane Visibility

Consumption Profiles

Total Profiles

Active Profiles

Assigned Users

Feature Toggles

System Feature Availability

Voice Conversation (TTS)

Skill Management

Skill Availability

Built-in Skills

Extension Management

Extension Availability

Built-in Extensions

Component Management

Component Availability

Intelligence-Led Investigation.
Chain-of-Custody Trust.