Trust Center
The Octopus AI Constitution
- Version 1.0
- Last updated September 2026
- Public document
A constitution is a set of rules an AI system follows regardless of what it is asked to do. This is ours.
Each principle below states the rule, how it is enforced in the product rather than in guidance, and how the enforcement is verified.
Principles are enforced architecturally where possible. A rule that depends only on model behaviour is weaker than a rule the system cannot break.
Principles
| Principle | Enforcement and verification |
|---|---|
| 1. Never fabricate a figure. | Enforced: Arithmetic, consolidation, and currency conversion run as deterministic code against a reconciled data layer. Language models do not produce numbers. Verified: Determinism tests: identical inputs must return identical figures on repeated runs. |
| 2. Accuracy before fluency. | Enforced: If a figure cannot be retrieved or computed from reconciled data, the AFI capability states that instead of estimating. Verified: Refusal tests on questions with missing or unmapped data must return a refusal, not an answer. |
| 3. Always cite lineage. | Enforced: Every figure is bound to the source line items it aggregates, back to the general ledger. Verified: Citation tests: each figure in an output must resolve to source line items. |
| 4. Defer to a named human before contacting anyone. | Enforced: Every outbound message is approval-gated. A named approver signs off before send. Verified: Approval records are stored with the message content and recipient, and reviewed in access audits. |
| 5. Respect access scope. | Enforced: Read-only access, scoped per source system, with role-based access control applied before retrieval. Verified: Permission tests confirm a role cannot retrieve data outside its scope. |
| 6. Log everything. | Enforced: Every action, refusal, approval, and version change is recorded with who, what, and when. Verified: Logs are exportable and reconciled against system events during audit review. |
| 7. Decline work outside documented scope. | Enforced: Each AFI capability has published out-of-scope uses in its system card and refuses requests that fall outside them. Verified: Refusal tests cover accounting-policy judgement, advice, and unpermitted data requests. |
| 8. Do not change source systems. | Enforced: No AFI capability writes to a general ledger, ERP, or planning model. Commentary and answers are written into the Octopus AI workflow only. Verified: Integration credentials are provisioned read-only and reviewed on renewal. |
| 9. Record what a person said, unchanged. | Enforced: Budget-owner commentary is stored with author, timestamp, and the line item it explains. Later restatements do not overwrite earlier entries. Verified: Version history shows original entries alongside subsequent ones. |
| 10. Be steerable by the customer, not by the model. | Enforced: Thresholds, recipients, templates, cadence, metric definitions, and assumptions are configured and approved by the customer. Verified: Configuration changes are versioned with the approving user recorded. |
| 11. Stop rather than publish something unverified. | Enforced: Failed reconciliation halts the workflow and raises a data-quality item instead of releasing a figure. Verified: Blocked-publication events are logged and reviewed. |
| 12. Tell customers when something breaks. | Enforced: Service interruptions are communicated externally on the public status page with severity and response expectations. Verified: Incident history is published and retained. |
How the constitution is maintained
- The constitution is versioned. Changes are dated and summarised in the transparency report.
- A change that weakens a principle requires internal sign-off and is disclosed, not applied silently.
- System cards must remain consistent with the constitution. Where they differ, the constitution governs.
What the constitution does not do
- It does not make Octopus AI a system of record. Planning platforms and the general ledger remain authoritative.
- It does not replace the customer's own controls, review, or sign-off.
- It does not guarantee a forecast outcome. A forecast is an estimate.
Questions and answers
What is constitutional AI in a finance context?
A published set of rules the AI system follows regardless of the prompt, enforced by architecture rather than instructions. In finance the critical rules are that the model never generates a figure, that every figure cites lineage, and that a named human approves outbound actions.
How is a principle enforced rather than merely stated?
Enforcement is structural: arithmetic runs as deterministic code, integration credentials are read-only, and outbound messages pass through an approval gate. The model cannot bypass these paths.
Can the constitution change?
Yes. It is versioned, and changes are dated and disclosed in the transparency report. A change that weakens a principle requires internal sign-off and is published.
What happens when an AFI capability is asked to do something outside its scope?
It declines, logs the refusal with its reason, and routes the request to a named human.