Reconstructing the task list
The first inventory listed six deliverables. Reading the original requests alongside the repository history revealed two more: the GitHub organization and private repository, and the first standards meeting’s agenda draft. We fixed the unit at one deliverable before assigning any actor. Ten language versions counted as one website. Repeated verification attempts counted as one X registration. Both unfinished registrations stayed in the ledger.
| Deliverable | Prepare | Execute | Check | Deliver |
|---|---|---|---|---|
| Assessment protocol and kit | Shared | AI | AI | AI |
| Website and domain | Shared | AI | Shared | AI |
| GitHub organization and repository | AI | AI | AI | AI |
| Project email and daily check | AI | Shared | AI | AI |
| LinkedIn page | AI | AI | AI | AI |
| X account | AI | Shared | Shared | Unfinished |
| Nonprofit incorporation | Shared | Shared | AI | Unfinished |
| Meeting agenda draft | Human | AI | AI | AI |
Attributing human participation
A founder may give permission while an agent still does the work. For LinkedIn, the founder authorized representation and accepted the terms; the agent prepared the form, submitted it and checked the page. For the project mailbox, the founder personally logged in, so execution was shared. For X, the founder completed the face check and corrected the inbox the agent was inspecting: execution and checking were both shared. For incorporation, the founder personally paid for the registered agent. Only the state filing receipt would close that task.
Reading the G1 result
- AI · 21
- Shared · 8
- Human · 1
- Unfinished · 2
With equal task and stage weights, 21 stages were AI-only, eight shared, one human-only and two unfinished. Counting none of the shared work toward AI gives 21/32 (65.6%); counting all of it gives 29/32 (90.6%). G1 requires more than half. These are bounds on a proxy for workload. Actual hours and offline human work were unavailable.
If each task can weigh one to two units, the lowest AI share is 56.8%. Counting GitHub and LinkedIn permissions as shared execution brings that lower bound to 52.3%. Counting the domain purchase permission the same way brings it to 47.7%. Under that combined assumption G1 is uncertain.
What would establish G2?
G2 requires accepted AI end-to-end deliveries, without substantive human rescue, worth more than half of all due work, with at least one in every preregistered core workflow. A live website is a candidate. Its four stage labels do not show acceptance against a criterion fixed in advance. Neither X nor incorporation reached its endpoint. This window lacked preregistered workflows and workload weights, so G2 coverage cannot be calculated. The retrospective scorer evaluates only G0 and G1.
For the next window, one outcome closer to the project’s mission would be an external team completing an assessment with its own records and independently reproducing its report. Register that acceptance condition, the workflows and weights before the window opens, and record any human takeovers.
Repeat the assessment
- Fix a team, a window and an acceptance condition for each deliverable. Export requests as well as completed work; ask someone who can see human-only work to flag omissions.
- Group records by deliverable. Keep failed, pending and human-led items. For each stage, cite a dated extract; record missing attribution as unknown.
- Unzip the public case and run the command below in that folder with Python 3.9+. Compare the new report with expected-report.json. For your team, use AGENT.md in the general kit to build a new manifest and test alternative weights.
python3 evaluate.py manifest.json my-result