Appearance
Evidence and provenance
DutyGraph records not only what it knows about a company, but how it knows it. A fact from a website, a guess from an AI model and a statement from the person who does the work are different things. DutyGraph keeps them apart and never promotes one into another without a recorded human step.
Five kinds of standing
| Standing | Who produces it | What it means | What it does not mean |
|---|---|---|---|
| Research | Company research from public sources | A sourced finding, with its sources and a confidence | That the company agrees |
| AI proposal | An AI model, or code that proposes | A draft: a Task Card, a reading of an answer, an edit, an AI mode | A fact. Nothing changes until a person accepts it. |
| Participant confirmation | The person who answered | "This is what I meant." They confirmed DutyGraph's understanding of their own words. | That the company accepts it as true |
| Advisor approval | The advisor | The advisor reviewed and accepted the item, or approved a report for an audience | Authority inside the client's business |
| External authority | People in the company with that authority | A decision by the people who already approve access, recorded in the approval chain | That DutyGraph changed anything in a business system |
Evidence, approval and enforcement are different
Evidence says something is true. An approval says someone with authority decided. Enforcement says a system allowed or refused an action. None of them proves the others. A confirmed observation is not authority.
DutyGraph keeps these kinds of evidence and authority distinct:
| Kind | Who produces it | What it is | Status in DutyGraph |
|---|---|---|---|
| External research | Company research | Sourced findings from public pages | Built |
| Worker observations | The person who does the work | Their answer, stored as understood | Built |
| Manager confirmations | The person's manager | That the work is real and is the person's work | Partial: the Manager step of the approval chain is Built. A manager confirming a newly captured task is Planned. |
| Advisor reviews | The advisor | Acceptance into the company's record | Built |
| Application-owner validation | The owner of a business system | That a capability maps to specific entitlements | Planned. The role owner step of the approval chain is Built. |
| Internal Controls approval | Internal Controls | SoD review and any mitigating control | Built, as a step of the approval chain |
| IAM issuance | IAM | The decision to issue, recorded in DutyGraph | Built as a ledger entry. Creating the identity happens in the company's IAM. |
| External provisioning evidence | The company's IAM or provisioning system | That access was actually granted | Planned, through Event Import |
| Runtime observations | The gateway, a runtime or a business system | What the agent actually did | Built for calls through the DutyGraph gateway. Other sources are Planned. |
AI interpretations stay proposals until the right person validates them through the right workflow.
Evidence states on Repo files
Every file in the Company Repo carries one state, computed from its sources.
| State | Meaning |
|---|---|
| Confirmed | A person with authority confirmed it in DutyGraph. The confirmation is recorded. |
| Observed | At least one client source backs it, such as an accepted answer from the person who does the work. |
| Inferred | It rests only on research, a draft, a rule or an advisor entry. It needs a client source. |
| Disputed | Sources disagree. The best-supported value is shown. |
| Unknown | DutyGraph searched and found no usable evidence. |
| Proposed | A draft that waits for review. |
| Generated | DutyGraph builds it from other files, for example an index or the company brief. |
| Slot | Nothing has filled it yet. |
Research, AI and an advisor's own entries never make a file Observed on their own. A research-backed file also shows a confidence, for example "Inferred · 77%".
How each kind is recorded
Research
- Each finding keeps its sources, their type (filing, company website, job posting, public web and others) and a reason for its confidence, such as "Corroborated by independent sources" or "Single independent source".
- Unknown is not zero. A field with no evidence has no value and zero confidence. DutyGraph never fills it with a guess.
- A correction from a person is new evidence. It never deletes the research evidence it corrects.
- A refresh never changes a value that a person confirmed. When new research disagrees, the confirmed value stays and a flag goes to Repo › To review.
- A simulated run is labelled "Simulated (not real data)" on every finding.
AI proposals
- A drafted Task Card is Proposed and Inferred until the advisor accepts it.
- The DutyGraph Agent only proposes. Accepted text is labelled as an advisor draft and stays Inferred until a person at the company confirms it.
- A proposed link between files, for example between an initiative and a department, is marked Proposed and is not treated as a fact.
Participant confirmation: stored as understood
- The person answers an ask in their own words, by voice or text.
- Here's what we heard shows DutyGraph's understanding of the answer.
- The person confirms or corrects it, then sends it.
DutyGraph keeps the person's own words as evidence and the confirmed version alongside them. This is stored as understood. The confirmation binds the exact understanding the person saw. After the advisor accepts the answer, the Repo file shows the confirmed version, never the raw words.
Answers from a company picture review are held for the advisor. They become client evidence only when the advisor accepts them, because a private link does not prove who the person is.
Advisor approval
- Accepting an answer, a Task Card or a research correction is recorded with who did it and when.
- Approving a report binds its exact content and audience. Each delivery link is bound to the exact version and content hash.
- When a source the report used changes, the report is marked out of date, and it must be prepared again before it is sent.
External authority
DutyGraph does not hold authority inside a client's business systems. For a governed agent, the people who already approve access record their decisions in the approval chain. DutyGraph records each decision on the ledger and checks calls routed through the DutyGraph gateway. The company's IAM provisions the access. The business systems behind the gateway are simulated today. See Approval, provisioning and revalidation.
Versions and content hashes
- Every record change is a new version. Earlier versions cannot be changed or deleted by the application.
- Each version has a content hash over its canonical JSON plus the tenant, company, record id, kind, title and version.
- A confirmation binds an exact version and hash. If the described work changes, the person must confirm again. A change to advisor-only fields, such as the owner or a review date, carries the confirmation forward with a note.
- State changes are audit events, not new content versions. A hash identifies content, not every later review step.
- Retracting evidence keeps its history and marks the records that depended on it as stale.
Idempotency and concurrency
- Every command carries an
Idempotency-Key. Repeating the same key with the same request returns the saved result. The same key with a different request is refused. - A replay never shows a private link a second time, and never makes a second AI call.
- Each payment event from Stripe is processed once.
- Writes carry the version they expect. If someone else changed the record first, the write stops with a version conflict instead of overwriting newer work.
The ledger
Governance decisions have their own hash-linked, append-only ledger, with Verify chain. Hash-linking shows that a record was not changed after it was written. It does not show that the business claims in it are true. See Ledger and Audit evidence.
INFO
Status. Built: the five kinds of standing and their separate records, the evidence and authority kinds marked Built above, the eight Repo file states, research sources and confidence reasons, stored-as-understood answers, report approval bound to exact content, versioned records with content hashes, idempotent commands, version checks and the governance ledger. Simulated: the business systems behind the agent gateway. Partial: manager confirmations (the approval-chain step only) and runtime observations (gateway calls only). Planned: application-owner validation of entitlement mappings, external provisioning evidence through Event Import, and provisioning an agent's own identity in the company's identity system.
Source: gitbook/product/trust/evidence-and-provenance.md in the DutyGraph repository.