Verify
The orchestrator routed you here. This atom checks the final deliverable against the commitments made at the start of the workflow.
Inputs
You receive:
- Commitments from the command's parse step -- the promises made about what the deliverable would contain
- Compose outputs -- the finished artifact(s) and their metadata
Composition is optional: raw desktop screenshots/state and the driver's evidence handoff can be the deliverable. Verify the task's postcondition, not merely the action's exit status or effect field.
Status vocabulary
Every step, check, and postcondition gets exactly one of these; commands and driver atoms use the same words.
| Status | Meaning | Counts as |
|---|---|---|
PASS | The postcondition was observed | Met |
FAIL | The postcondition was observed to be false | Not met — this is a finding |
BLOCKED | The postcondition could not be observed: capture unavailable, permission wait unresolved, missing binary or connection, environment prevented the step | Neither. Name what blocked it and what unblocks it |
BLOCKED is never rolled into PASS or FAIL. A deliverable with BLOCKED steps is incomplete unless the user accepts the partial evidence. /verify's claim verdict (CONFIRMED / REFUTED / INCONCLUSIVE) is about evidence sufficiency, not step status: any BLOCKED step that the claim depends on makes the verdict INCONCLUSIVE, but INCONCLUSIVE also covers steps that all ran and still did not decide the claim.
Video deliverables
Technical checks
Run ffprobe on the final .mp4, then a full decode:
ffprobe -v quiet -print_format json -show_format -show_streams <video>
ffmpeg -v error -xerror -i <video> -f null - # -xerror: exit non-zero on the first decode error
| Check | Pass condition |
|---|---|
| Exists and decodes | ffprobe exits 0 with duration > 0, and the ffmpeg -xerror decode exits 0 with empty stderr |
| Resolution | Matches the fidelity compose resolved: 1920x1080 for compact/standard, 2560x1440 for inspect (the default for side-by-side), or the explicit width/height in the props |
| Pixel format | pix_fmt=yuv420p with color_space=bt709 (a yuvj420p stream means the render bypassed render-showcase.sh) |
| File size | Under 5 MB for GitHub embeds (25 MB hard limit) |
| Duration | Equals 4s + longest_clip / speed + 3.5s (compose's duration checkpoint) and falls within the compose pacing table's target range for this demo type: 30-45s (single feature), 45-75s (side-by-side comparison), 60-120s (multi-phase). Below the minimum is a failure — re-compose with a lower speed factor or re-capture with more steps. |
| Filename | Includes PR number or meaningful identifier |
Commitment checks
Walk through each commitment from the parse step:
| Commitment | How to verify |
|---|---|
| Title card | Video starts with a static frame showing PR info (check first 5s) |
| Side-by-side layout | Video shows two panels with a divider |
| Showcase polish | Resolution matches the resolved fidelity, window chrome and rounded corners visible |
| Keystroke overlay | Pill overlays appear at interaction points |
| Effects | Effects matching the committed tier are present (utilitarian: zoom/keystroke; full: spotlight, zoom, callout, keystroke) |
| Speed note | Title card mentions playback speed |
Content checks
Metadata cannot show what is on screen. For every proof point, extract the frame and look at it:
ffmpeg -v error -y -ss <t> -i <video> -frames:v 1 "${RUN_DIR}/proof-<t>.png" # then Read the PNG
- Every claim from the "what to prove" analysis has visible evidence in the video, confirmed in an extracted frame at the timestamp you report
- Both states (before/after, input/result) appear on screen
- The final state: clips play from
t = 4.0stot = 4.0s + longest_clip / speed, fully visible; after that the last frame is held under the outro crossfade. Extract the final proof frame just before the clips end (e.g.t = 4.0 + longest_clip / speed - 0.1), not from the crossfade - In a side-by-side, a shorter clip holds its final frame; confirm the held panel shows the intended final state, not an unfinished step
- No dead time longer than 3 seconds without visible activity
Screenshot/snapshot deliverables
For proofs
| Check | Pass condition |
|---|---|
| Evidence exists | Screenshots/snapshots at every claimed proof point |
| Environment stated | Driver, terminal/browser, OS identified |
| Conclusion present | Evidence explicitly supports or refutes the claim |
| Before/after paired | If comparison, both branches shown at same capture points |
For QA reports
| Check | Pass condition |
|---|---|
| Step coverage | Every defined test step has a status (PASS / FAIL / BLOCKED) |
| Evidence attached | Screenshots/snapshots at every step |
| Failures documented | FAIL steps have evidence and description; BLOCKED steps name the blocker and what unblocks it |
| Report structured | Markdown report follows the QA template |
Failure handling
If any check fails:
- Identify which stage produced the problem (capture or compose)
- Report the specific failure: "Side-by-side layout was committed but the output is a single panel"
- Go back to the failed stage and fix it
- Re-verify after the fix
Do not report a deliverable as complete until every commitment is met.
Output
## Verification
### Technical
- Decode: ffmpeg -xerror exit 0 ✓
- Resolution: 2560x1440 (inspect, side-by-side) ✓
- Duration: 52.5s = 4 + 45 + 3.5 ✓
- Size: 3.2 MB ✓
- Format: yuv420p / bt709 ✓
### Commitments
- [x] Title card with PR info
- [x] Side-by-side comparison layout
- [x] Showcase hero preset applied
- [x] Keystroke overlay visible
### Content
- [x] Fork creates independent session (frame at 0:18 read: new session id in header)
- [x] History diverges after fork (frame at 0:32 read: differing last messages)
All commitments met. Deliverable ready.