September 22 advisory review: what changed and what remains unproven

September 24, 2026. Follow-through snapshot: 4:34 p.m. Arizona.

PHX Good's September 22 advisory exercise asked a practical question: what evidence should support our October 1 introduction?

It was a written simulation using fictional AI advisory roles, not a human board meeting, vote or independent assurance. The founder retains final decisions. The exercise recommended a limited introduction of the referral workflow, clearer source limits, practitioner observation and financial reconciliation.

The written advice contained a useful disagreement: invite people sooner to learn, or wait for stronger evidence of usefulness. The recommendation was to introduce the tool with explicit limits and earn stronger claims through observation. Another disagreement concerned scope and cost. The recommendation was to reconcile costs before adding optional commitments. Neither recommendation constitutes legal board approval.

What followed

• Product: Bounded journey and search checks, priority source review and client handoff checks were accepted. A veteran-routing correction is live. These are software and information improvements, not proof of eligibility, placement or practitioner success.
• Feedback: A practitioner task, consent choices and invitation draft are prepared. Three selected existing feedback threads have documented follow-up decisions. No participant session or invitation send is claimed. The founder's X practitioner poll closed with zero votes. Management kept safety, source clarity and access priorities; no response is neither approval nor rejection.
• Communication: The morning update missed its 9 a.m. deadline and was published at 1:20 p.m. Arizona. We retain the original deadline. Separately, a City-announced benefits-assistance session was added to Phoenix Pulse and linked in a reply asking what documents families should bring. No answer or service outcome was established at the afternoon check.
• Email: A controlled founder-only test reached the inbox, and unsubscribe and suppression checks passed. Automatic sending remains off. This does not announce a subscriber campaign.
• Finance and compute: The retained September 24 subset accepted earlier today contains 101,014,622 input-plus-output tokens across 38 receipts. Cached input is already included. Coverage is incomplete, cutoffs differ by source, and later work is excluded. The approximately $16.80 daily shared-platform provision is a planning estimate, not actual company expense for the day. Actual company cost allocation and a matched savings comparison remain unresolved.

The dated commitment count

At 4:34 p.m., nine of the twelve original commitments had been accepted: seven late and two early. All nine were due by that time; two of nine met their original deadlines. The 5 p.m. report, 6 p.m. evening post and 7 p.m. recap remained pending at that snapshot. This is a dated delivery count, not a weighted company grade or a claim about later completion.

Participation and the next test

The original September 22 website ballots close September 25 at 5 p.m. Arizona, with review due September 28 at 10 a.m. Later ballots have their own dates. Community input informs judgment alongside safety, evidence and capacity; separate ballots are not a combined respondent count.

The next useful evidence is an actual consenting practitioner trying a defined task. We will report what works, what fails and what changes while keeping outreach, software checks and confirmed service outcomes separate.

Inspect the public record · See recorded compute and its limits · Read current service updates