PJC Workflow Integration Note
Purpose: explain how PJC fits into a human–AI agent workflow.
Boundary: these are engineering constraints, not evidence of effectiveness.
1. Role separation
| Role | Responsibility |
|---|---|
| Judgment subject | Confirms proposition, rationale, confidence, and falsification rule |
| Agent | Supports research, execution, organisation, and checking |
| Record service | Enforces timestamps, versions, and append-only originals |
| Outcome source | Supplies the pre-specified external observation |
| Settlement reviewer | Applies the rule and handles disputes |
One agent should not create the judgment, control the outcome source, and settle the record alone.
2. State model
Draft
↓ commit
Committed
↓ observation window opens
Awaiting Outcome
├─ settle → Settled
├─ insufficient evidence → Inconclusive
└─ contamination or rule failure → Invalidated
Every transition records actor, time, and reason.
3. Data invariants
Originals are append-only
After commitment, the proposition, confidence, outcome source, and falsification conditions remain unchanged. Corrections are appendices.
Time order is auditable
Keep creation, commitment, human and agent appendices, first outcome availability, and settlement times.
Outcome sources are bound in advance
Store an identifier, query, or adjudication protocol at commitment. A source change at settlement invalidates the record unless a fallback rule was already specified.
Permissions are minimal
Agents receive only the permissions needed for the current step. An executing agent cannot rewrite the original or delete a failed case.
4. Recommended events
| Event | Minimum content |
|---|---|
| judgment.created | record_id, subject, created_at |
| judgment.committed | original-field digest, committed_at, signature |
| evidence.appended | source, digest, actor, timestamp |
| outcome.observed | source, value, observed_at |
| judgment.settled | status, evidence, reviewer, settled_at |
| judgment.reviewed | dispute, adjustment, version |
The log should be exportable for independent review.
5. Agent rules
Agents may identify missing fields, propose alternatives, test settleability, execute authorised work, and organise outcome evidence.
Agents may not submit without human confirmation, backfill original reasoning from outcomes, delete contradicted records, switch criteria without disclosure, or treat their own natural-language assessment as an external outcome.
6. Security and privacy
Judgment records may contain plans, personal performance information, and unpublished data. Minimise data, separate public cases from internal records, define retention, audit access, anonymise before publication, and prohibit covert employee surveillance.
7. Minimum integration
A complex platform is unnecessary at first. Existing versioned tools are sufficient if they can verify pre-outcome commitment, prevent silent overwrite, pre-register outcome sources, retain every committed record, and export settlement states.