Pack definition
The public role, available evidence, permitted actions, required approvals, and expected result come from the same pack record Cadre uses.
Methodology
What a Cadre pack shows.
Cadre publishes the design of each agent pack alongside representative cases, process placement, and, where available, frozen model runs. This makes the pack easier to examine. It does not establish production fitness, comparative performance, or regulatory compliance.
Evidence vocabulary
How Cadre distinguishes design, examples, and deployment evidence
- 01
Operating design
What the pack is intended to do and where its authority ends.
- 02
Evaluation case
What behavior a representative, borderline, or edge case is designed to test.
- 03
Representative capture
What happened in one frozen execution; not a repeatability claim.
- 04
Paired policy capture
What changed across two separate examples with an illustrated control.
- 05
Institution-validated deployment
Evidence an institution establishes in its own data, policy, controls, and operating environment—not claimed by this site.
From pack to page
The published view follows the implemented pack.
Published examples
Published examples come from selected evaluation runs. Sensitive data and private instructions are removed before publication.
Operating-line definitions
Process roles and handoffs are matched to the pack library. Work without a corresponding pack remains visible as an open role.
Published
What can be inspected
Operating design
The role, inputs, tools, action boundary, structured handoff, and guardrails.
Evaluation cases
Scenario inputs and the behaviors each case is intended to exercise.
Representative captures
A sanitized sequence of evidence checks, policy decisions, and the recorded outcome.
Paired policy captures
Separate examples with and without an illustrated policy so the intervention and next action can be inspected.
Operating-line maps
Where packs, people, systems, checkpoints, and open roles meet in a wider process.
Withheld
What never crosses the boundary
Prompt templates
System prompts and prompt-bearing context payloads are removed before publication.
Credentials and infrastructure
Secrets, environment data, internal hosts, account identifiers, and local paths never publish.
Customer data
The public examples use representative domain cases, not a bank's production records.
Live controls
The public site cannot execute an agent, call a tool, or modify a run. Replays are frozen records.
Reading a replay
One run is a close-up, not a verdict.
Context
The replay records that context was assembled, while the system prompt and raw model context stay withheld.
Model turn
Published turns show the material retained for the replay. They are not rewritten after capture.
Tool use
Inputs and results show what the pack checked, subject to the publication sanitizer.
Policy
Where policy intervened, the trace records the decision, policy reference, and resulting change of course.
Handoff
The structured result can be read against the pack's published expected result.
Limits and corrections
A capture keeps its context and its rough edges.
A captured run can contain a failed lookup, a weak citation, or an inconsistent line of reasoning. Published model output is not silently rewritten to make a pack appear more capable. A new run becomes a new record; the old record remains attributable to its captured date.
Representative cases are designed to exercise financial-services work, but they are not a substitute for an institution’s data, policies, model validation, legal analysis, or production controls. Aarvion, Cadre’s sister governance product, can be shown at the action boundary in selected captures; that is an illustration of the control pattern, not a certification. Corrections can be sent to hello@aarvion.ai.