Reconciliation
Task, revision, and deliverable declarations, side by side.
| Deliverable | Declared in packet | Finding |
|---|
Findings and next steps
A matching declaration is not release approval. Keep required evidence; correct records from the actual work.
Preparing a synthetic example.
A clearer handoff
What was requested.
What was reported delivered.
ProofNib compares a work-order JSON record with an AI completion-packet JSON record. It checks task identity, input revision, required deliverable records, declared status, hash references, and evidence-reference links.
It shows discrepancies in the supplied declarations. It does not open referenced files, run tests, establish actual completeness, or certify software. The examples above are synthetic.
Free working preview. No account, subscription, or checkout.
Record format and comparison limits
Work order: task_id, input_revision, and a nonempty required_deliverables list. Each deliverable needs an id and kind. An optional expected_sha256 provides a declared hash to compare.
Completion packet: the same task and revision, plus an artifacts list. Each artifact declares id, kind, status (complete, incomplete, or failed), and sha256. Optional evidence_refs links to IDs in the packet’s evidence list. No reference is fetched.
Limits: 100,000 characters, 64 nesting levels, and 20,000 values per input; at most 500 entries per list, 1,000 checked evidence references across a packet, and 1,200 retained findings. Exceeding a comparison limit is reported as incomplete. The screen shows up to 60 findings; the JSON export retains the complete bounded report.
Results: declarations align, contradiction, or not established. Only the work order’s declared scope is compared. Evidence-reference resolution does not verify the evidence itself.