<!-- Canonical: https://myoctopus.ai/trust/transparency-report -->

1.  [Home](/)
2.  /[Trust](/trust)
3.  /Transparency Report

Trust Center

# Transparency Report

Edition 1

Period: H1 2026

Published September 2026

This report documents how the Octopus AI platform was governed during the stated period, and what changed.

It is published on a recurring basis. Each edition covers a defined period and keeps the same section structure so editions can be compared.

Where a figure is not yet independently verifiable, this report describes the practice rather than stating a number.

## Scope of this report

| Item | Detail |
| --- | --- |
| Period covered | 1 January 2026 to 30 June 2026. |
| Systems covered | The Octopus AI platform and all production AFI capabilities documented in the model system cards. |
| What is measured | Governance and control practice: approval gating, access scope, logging, incident communication, and regulatory mapping. |
| What is not measured | Customer-specific outcomes, forecast accuracy for individual customers, and any figure that would identify a customer. |
| Reporting cadence | Semi-annual. Each edition is published on this page and dated. |

## Governance changes in this period

*   Model system cards published for the Variance Commentary, Forecasting, and Knowledge Q&AFI capabilities.
*   The Octopus AI Constitution published at version 1.0, with technical enforcement and verification stated per principle.
*   Public trust hub introduced, consolidating system cards, the constitution, this report, security documentation, and system status.
*   Public status page introduced for external communication during service interruptions.

## Approval gates and human oversight

Octopus AI initiates contact with budget owners. Every such message passes an approval gate before it sends.

*   Outbound messages requiring named human sign-off: all of them. There is no unattended send path in production.
*   Approval records retained: approver identity, message content, recipient, and timestamp, stored together as one evidence chain.
*   Configuration changes covered by approval: thresholds, recipients, templates, cadence, assumptions, and metric definitions.
*   AFI capabilities can be paused by the customer without pausing the platform.

## Service interruptions

Service availability and the current status of the platform are published on the public status page, together with severity definitions and response expectations. Interruption history is retained there.

*   No customer-data loss events were recorded in this period.
*   Interruption communication follows the published process: status page update first, then direct notification to affected customers.

## Third-party model providers and subprocessors

Octopus AI uses third-party language models for narrative generation only. No provider computes a reported figure.

*   Language models are used for phrasing, interpretation of free-text answers, and narrative drafting.
*   Deterministic computation runs inside the Octopus AI platform, not at a model provider.
*   The current subprocessor list, including model providers and hosting regions, is available to customers and prospective customers on request and in the security documentation.
*   Data residency is configured per customer contract. Customer data is not used to train third-party foundation models.

## Access scope and audit logs

| Control | Practice |
| --- | --- |
| Source-system access | Read-only, scoped per source, provisioned with read-only credentials. |
| Authorisation | Role-based access control, applied before data retrieval rather than filtering results afterwards. |
| Action logging | Every action, refusal, approval, and version change is logged with who, what, and when. |
| Lineage | Every figure traces to the source line items it aggregates, back to the general ledger. |
| Auditor access | Logs, approvals, and version history are exportable for internal audit and external auditor review. |

## Regulatory mapping

| Framework | How the architecture addresses it |
| --- | --- |
| SR 11-7 (model risk) | Reported figures are produced by deterministic code, not model output, so model drift cannot change a number. Documented intended use, limitations, and evaluation are published per AFI capability. |
| DORA | Approval-gated, logged, read-only architecture with published incident communication supports operational-resilience and third-party-risk documentation. |
| COSO and ICFR/SOX | Approval gates map to control activities, the audit trail and lineage to information and communication, and the evaluation and escalation practice to monitoring. |
| GDPR and data protection | Scoped access, contractual data residency, and no use of customer data for third-party model training. |

## Commitments for the next edition

*   Publish a system card for each additional AFI capability before it enters production.
*   Report constitution version changes with dates and rationale.
*   Report the count and category of service interruptions for the period.
*   Report refusal and escalation categories in aggregate, without customer-identifying detail.

## Questions and answers

### What does the Octopus AI transparency report cover?

Governance changes, approval-gate practice, service interruptions, third-party model providers and data residency, access scope and audit logging, and how the architecture maps to SR 11-7, DORA, and COSO/ICFR.

### How often is the report published?

Semi-annually. Each edition covers a defined period, keeps the same structure, and is published and dated on this page.

### Is customer data used to train foundation models?

No. Customer data is not used to train third-party foundation models. Language models are used for narrative generation only.

### Do third-party model providers calculate reported figures?

No. Deterministic computation runs inside the Octopus AI platform. Model providers handle language only.

### Where can an auditor see the underlying evidence?

Logs, approval records, lineage, and version history are exportable from the platform for internal audit and external auditor review.

Subprocessor lists, hosting regions, and certification reports are available under NDA on request. Figures marked as practice rather than counts will be reported quantitatively in the next edition.

## Related pages

*   [Model System Cards](/trust/system-cards)
*   [The Octopus AI Constitution](/trust/constitution)
*   [Security & Compliance](/security)
*   [System Status](/status)
*   [Trust Center](/trust)
