# Verification vs syntax — constructed demo

Tiny local illustration for a reporter sidebar on the AI Engineer talk *From coding to Knowledge work agents* (Karan Vaidya). The talk is the story. This folder is not.

**This is not** a recreation of that system, **not** an agent benchmark, and **not** evidence of performance improvement. It uses synthetic task records and proposed JSON edits. No mail is sent. No network is used.

Point: a well-formed JSON edit can still be the wrong action. Schema checks, project/revision policy, and recorded history are separate. Local restore does not recall an already-dispatched external side effect.

## Command

From this folder, stdlib only:

```bash
python3 demo.py
```

The script prints a case table, asserts the final state, and writes `observed-result.json` next to itself.

## Files

| file | role |
|---|---|
| `demo.py` | one executable: schema, policy, mutate/reject, revert, assert |
| `fixture.json` | two synthetic tasks, five proposed cases |
| `observed-result.json` | machine-readable result from an actual run |

## Observed

Command above, recorded in `observed-result.json` (`ran_at` 2026-09-09T07:13:37Z, Python 3.9.6), `assertions_passed: true`. Cases run in order.

| case | schema | policy | changed |
|---|---|---|---|
| a-in-scope-update | pass | pass | yes (task + irreversible notice marker) |
| b-out-of-project | pass | reject: outside allowed project | no |
| c-stale-revision | pass | reject: stale revision (expected 2, current 5) | no |
| d-malformed | fail | skipped | no |
| e-revert-local | n/a | pass: restored from history; 1 irreversible notice marker(s) remain | local restored; external notice not recalled |

After revert, task `T1` matched the fixture (`status: open`, `revision: 4`). Task `T2` never moved. The synthetic outbound-notice marker from case (a) remained, with `recallable_by_local_restore: false`.

Case (c) is schema-valid JSON against a live in-scope task. It was rejected because it cited revision 2 after case (a) had already committed revision 5. That is the history check, not a second schema check.

## Bounded conclusion

On this machine, schema-valid JSON was not enough to change a task: an out-of-project target and a stale revision were rejected before mutation. A truncated body never reached policy. An accepted local edit could be restored from recorded prior state. A notice marker recorded at accept time could not.

That is all this demonstrates. It does not measure agents, recreate a vendor stack, or support a product claim.

## Limits

- Synthetic records in one process.
- The “outbound notice” is a local log marker. Nothing was sent, and this cannot prove how real mail systems behave.
- Local restore cannot recall already-sent mail or other external side effects. The marker only records that boundary.
- Five cases, matching the assignment. The count is not a quality score.
- Subordinate to a sourced video feature; do not promote this demo to the main evidence.