Structured Outputs, Evidence, and Freshness

Home · Knowledge and Agents · Reports, AI Log, and Traceability

Objective

This page explains how to interpret a structured output in ProPM Agent to decide whether the result is directly usable, needs review, or should be transformed into a governed artifact.

Why It's Important

ProPM Agent does not return only free text. Observed runs can additionally expose:

• a summary;
• findings;
• decisions needed;
• actions;
• assumptions;
• missing information;
• watchpoints;
• evidence;
• metadata for freshness and confidence;
• artifact proposals, follow-ups, and notifications.

Where This Information Appears

Structured outputs and their metadata are visible in several surfaces:

• Agents, immediately after a run;
• AI Log, in the detail of a run;
• Reports & Artifacts, via the lineage between run, artifact, and PM Doc.

How to Read a Structured Output

Recommended Reading Order

1. Read the Summary;
2. Then open Decisions needed or Actions;
3. Check the Evidence section;
4. Review the Confidence and Source freshness;
5. Then decide whether the response can remain in the chat, become a draft artifact, or require human review.

Meaning of Sections

Section	What It's For
Summary	Shortest version of the result, ready for quick review
Findings	Facts, reasons, or observations highlighted by the run
Decisions needed	Points requiring human arbitration
Actions	Recommended follow-ups to execute or plan
Assumptions	Implicit conditions the response depends on
Missing information	Missing information that reduces the robustness of the result
Watchpoints	Risks, contradictions, or subjects to monitor
Evidence	References serving as proof or justification
Artifact status / Artifact proposals	Indications on the possible transformation of the result into a governed artifact
Follow-ups	Proposed continuation steps
Notifications	Drafts or notifications suggested by the flow

Simplified Reading Example

The example below is illustrative. It serves to show how to review a structured card before reuse:

• Summary: The critical batch appears to be two weeks late.
• Findings: The latest planning note is more recent than the governed calendar connector.
• Decisions needed: Should the sponsor milestone be rebaselined or should a planning revalidation be requested?
• Evidence: Planning note, excerpt from weekly report, state of the source connector.
• Source freshness: Knowledge fresh, planning connector stale.
• Watchpoints: Possible contradiction between project communication and the latest synchronized data.

Recommended reading in this case:

1. The Summary tells you what to review first;
2. The combination Evidence + Source freshness prevents you from publishing too quickly;
3. The presence of a Watchpoint and a decision to take pushes towards human review, then towards an artifact if the subject becomes formal.

Two Common Cases to Distinguish

Case	Practical Reading	Recommended Decision
Correct confidence but insufficient freshness	The reasoning seems coherent, but part of the sources is stale or aging	Do not disseminate without revalidation or documentary refresh
Low confidence but recent sources	The sources are recent, but the run still reports gaps, contradictions, or overly strong assumptions	Keep the output as exploratory work, then relaunch the review or the run before formalizing

This distinction avoids treating confidence as a single score. A response can seem solid while relying on outdated sources, or vice versa.

How to Read Evidence

Each piece of evidence or citation can expose several useful elements:

• a title or source name;
• a source URI or documentary name;
• a snippet;
• a page or section when available;
• a sync date;
• a source system;
• a freshness badge;
• an authority rank when exposed.

What to Check Before Reusing Evidence

1. Does the snippet properly support the displayed message?
2. Is the source identifiable and reopenable?
3. Is the freshness acceptable for the expected decision level?
4. Do multiple pieces of evidence tell the same story, or is there contradiction?

Freshness States

State	Practical Meaning	Recommended Reaction
fresh	Source sufficiently recent for normal use	Can be reused after normal review
aging	Source still usable but approaching a need for revalidation	Check quickly before broad dissemination
stale	Source too old to be considered reliable without additional control	Refresh, reimport, or confirm before decision
conflicting	The source contradicts another relevant source	Do not arbitrate automatically; review the evidence
unavailable	The source could not be confirmed or retrieved	Treat as an alert, not as usable evidence

Confidence

Confidence is a global signal of the result's robustness. It must be interpreted with freshness and evidence, never isolated.

Plan for additional human review if you see:

• low confidence;
• absent or non-specific evidence;
• stale, conflicting, or unavailable states;
• significant Missing information section;
• decision or action with external impact.

Review Level Based on Impact

Impact Level	Minimum Review	Recommended Escalation
Internal work draft	Check Summary, Findings, and at least one reopenable piece of evidence	Keep the output in the chat if it remains exploratory
Team coordination or internal project action	Check evidence, freshness, missing information, and proposed actions	Create an artifact if the result must be shared or archived
Sponsor decision, publication, external notification, or governed action	Review all evidence, arbitrate conflicting / stale states, retain technical IDs	Go through Reports & Artifacts, the diff, the lineage, and the AI Log before dissemination

Traceability IDs to Keep

Field	Usage
Trace ID	Locate precisely the run or event for support
Structured output ID	Identify the structured output actually produced
Context snapshot ID	Understand in which documentary or project context the run executed

These fields are particularly useful when you need to reconcile a run, an artifact, and an event from the AI Log.

When to Transform the Output into an Artifact

The typical observed path is:

1. Project question;
2. Structured response;
3. Creation of a draft artifact;
4. Review of the diff and lineage;
5. Approval or publication.

Transform an output into an artifact when:

• It must become a formal deliverable;
• It must go through validation or publication;
• You need to maintain explicit lineage to a run and its evidence;
• The content must leave the chat for external dissemination, approval, or durable traceability.

Simple rule: While exploring, the chat may suffice; as soon as a result must be shared, approved, published, or kept as a trace, go through Reports & Artifacts.

Checklist Before Reuse in a Deliverable

1. Does the summary accurately match the detailed findings?
2. Are the proposed actions consistent with the project situation?
3. Is the evidence sufficiently precise to justify the decision?
4. Are the freshness states acceptable?
5. Should an artifact be created or is keeping the output in the chat sufficient?

Common Issues

The Response Seems Clean but No Evidence Appears

Treat the result as an item to review before external use. The absence of visible evidence reduces the ability to justify the content.

Evidence is conflicting

Do not publish directly. Open the source, compare the conflicting references, and document the arbitration in the artifact or in governance. If the subject requires a formalized decision, proceed to Governance, Decisions, and Actions rather than leaving the conflict solely in the chat.

Evidence is unavailable

Consider that the justification is not stabilized. Check the source in Knowledge, the import, or the run detail in AI Log.

Next

• Knowledge and Agents
• Reports, AI Log, and Traceability
• Maintenance, Support, and FAQ