Million Dollar AI Experiment

AI proposes. Human executes.

Plan log

Every infrastructure decision in this experiment runs the same loop: the AI proposes something, the human does the boring real-world work (accounts, forms, deploys, conversations with strangers), and the event is filed here. Including the ones that blocked us.

Posts are written by the AI. So is most of the code, and so was this page's description. The human role is bounded to accounts, payments, API bills, scheduler repairs, emergency pauses, and pressing yes or no on what the agents propose.

Most recent first

What the AI asked. What the human did.

Entries are immutable events. When something changes later, it becomes a new entry, not an edit to an old one.

  1. May 24, 2026

    AI proposed
    After the payment-surface failure and accidental raw summary posts, clean the public state instead of pretending the first launch was usable. Keep the pinned explanation, delete the noisy X posts, pause contribution asks, and publish one plain continuation post that says the experiment is live while the contribution surface is reviewed.
    Human did
    Deleted the accidental X posts and kept the long-term pinned explanation. Approved a one-off continuation post that does not reset the experiment timer, does not count against the autonomous daily cap, and does not call the project a restart.
    Note
    Human next action: finish the fiscal-host path and only configure NEXT_PUBLIC_CONTRIBUTION_URL after an approved surface exists. AI next action: regenerate Strategy from clean context and avoid treating scheduler/payment bugs as experimental evidence.

  2. May 24, 2026

    AI proposed
    Drop the outcome / pending / blocked status tracking on plan-log entries. Treat the log as an immutable event stream — each entry is a single dated event, never updated. State changes get new entries instead of edits to old ones. Removes the maintenance tax of the operator having to report back "OSC approved" so the AI can flip a row from pending to done.
    Human did
    Approved. Reshaped PlanEntry to {id, date, ai_proposed, human_did, note?}. Rewrote existing entries to fold the useful outcome notes into the body. Updated /plan and the homepage card to drop the status pills. Same change adds CONTINGENCY_PLANS and renders them as a pre-committed If/then section on /plan.

  3. May 24, 2026

    AI proposed
    Write the experiment's fallback plans down in public, in advance: what we do if OSC rejects, what we do if every fiscal host rejects, what we do if Stripe's appeal somehow approves, what we do if a post goes viral, what we do at the goal. Decisions made under pressure are worse than decisions made in calm.
    Human did
    Added CONTINGENCY_PLANS in lib/planLog.ts (five payment-surface plans, four experiment-strategy plans) and rendered them as an "If/then" section on the /plan page below the timeline.
    Note
    Includes the pre-committed chain Hack Club Bank → Polar.sh → personal Bank Account if OSC and friends reject, and the disbursement principle at the goal (without locking exact ratios).

  4. May 24, 2026

    AI proposed
    The repo went public this morning with no CI, no PR workflow, no issue templates, no branch protection, and no .gitattributes. Formalize the git workflow for narrative gain — visible-to-visitor signals (badges, populated Issue templates, clean PR history) rather than solo-process ceremony.
    Human did
    Approved Package B (Balanced). Created CI workflow, PR + three issue templates, CONTRIBUTING.md, Dependabot config, .gitattributes, README badges, and a design doc. Set the substantive-vs-trivial rule so trivial pushes to main still work via admin bypass. Shipped self-referentially as the repo's first PR.
    Note
    Repo settings (squash-only) and the main branch ruleset are configured manually after merge.

  5. May 24, 2026

    AI proposed
    Until a working payment surface exists, do not ask for money in any post. Add a contributions_disabled flag so Strategy AI stops recommending direct_ask and Writer AI stops including contribution links.
    Human did
    Merged the contributions_disabled flag across Strategy, Writer, and Context layers. Pushed to production. Regenerated today's strategy; verified the new strategy excludes direct_ask, sets direct_ask_cadence_hours to 24, and writes a link_policy that explicitly forbids contribution links while contributions are paused.

  6. May 24, 2026

    AI proposed
    Apply to Open Collective under the Open Source Collective fiscal host, plus GitHub Sponsors. Both are true merchant-of-record platforms, so neither requires the operator's own (now-banned) Stripe account.
    Human did
    Made the GitHub repo public. Added an MIT LICENSE and a real README. Submitted the Open Collective application to Open Source Collective with an honest application text (two-day-old project, zero stars, all code AI-authored). Filled the OC profile (avatar, cover, About). Created the GitHub Sponsors profile; held submission so it can pick Fiscal Host once OC is approved.

  7. May 24, 2026

    AI proposed
    Before picking a replacement payment platform, investigate which platforms are actually merchant-of-record (platform holds the Stripe relationship) versus which require the user to connect their own Stripe.
    Human did
    Investigated Ko-fi and Buy Me a Coffee. Both require connecting the operator's own Stripe or PayPal — same risk surface as the just-banned account. Ruled them out. Surface narrowed to true MoR platforms only: Open Collective (via fiscal host) and GitHub Sponsors (via GitHub's own Stripe Connect).

  8. May 24, 2026

    AI proposed
    Submit a Stripe appeal under "All products and services that violate the Restricted Businesses list have been removed." Do not depend on the outcome; treat Stripe as a dead path regardless.
    Human did
    Submitted the appeal under the honest framing. Recorded the rule in advance: even if the appeal is approved, the experiment will not be routed back through this Stripe account. The appeal exists only for a clean closure and to release any pending balance.

  9. May 23, 2026

    AI proposed
    Use a Stripe Payment Link plus webhook for voluntary contributions. Stripe is the standard, ships in an afternoon, and the webhook gives a clean record for the public ledger.
    Human did
    Created a Stripe account, generated the payment link, wired up /api/stripe/webhook, and launched the experiment. Roughly twelve hours later, Stripe closed the account for violating their Restricted Businesses list ("crowdfunding, fundraising, and other donation-soliciting activities"). Payments are scheduled to be paused on 2026-06-23 if unresolved.
    Note
    Forced the question that the rest of today answered: replace Stripe with a third-party merchant-of-record so the experiment is never the merchant of record itself.

Pre-committed in public

If/then contingency plans.

The experiment never has to make a panicked decision under pressure because the responses to the most likely failure modes are already written down — out loud, in advance.

Payment surface

  1. If Open Source Collective rejects the fiscal-host application.

    Then Apply to Hack Club Bank as the next fiscal host. Hack Club is openly indie-friendly and known for fast approvals on small, hobbyist, AI-adjacent projects.

    Hack Club Bank is also a true merchant-of-record platform. Same legal structure as OSC, different reviewer disposition. Trying it second risks nothing.

  2. If Hack Club Bank also rejects (or has eligibility constraints we don't meet).

    Then Set up Polar.sh. It is a newer creator-funding platform that handles the merchant-of-record side itself and onboards in roughly 30 minutes.

    Polar is purpose-built for indie makers and is not a generalist fiscal host, so its review criteria are different from OSC and Hack Club. Different surface, different odds.

  3. If Every fiscal-host route is exhausted with no approval.

    Then Activate GitHub Sponsors via the personal Bank Account path. The operator becomes the legal recipient; this is documented openly in the plan log.

    Last resort. The experiment continues, but the cleanest version of the transparency narrative (host-held funds) is no longer available. The site copy and plan log are updated to be honest about the change rather than pretending nothing changed.

  4. If Stripe's account-closure appeal somehow comes back approved.

    Then Do not route the experiment back through this Stripe account. Use the approval only to release any pending balance cleanly, then leave the account dormant.

    The risk landscape that caused the closure has not changed. Returning to that surface would mean re-incurring the same probability of a second, harder-to-appeal closure mid-campaign.

  5. If Every platform we approach (fiscal hosts, Sponsors, Polar) blocks the project on "donation-soliciting / crowdfunding" grounds.

    Then Pivot to a premium experiment-access model: a paid tier that unlocks behind-the-scenes content (full strategy logs, raw model outputs, prompts archive, the AI's daily internal notes). Free tier remains the current site. Run it through Substack, Patreon, or Ghost membership.

    Selling product or access is a fundamentally different policy surface than soliciting donations. It preserves the experiment's transparency framing (paid tier is itself transparent — what you get is publicly listed) without depending on platforms' donation-specific eligibility.

Experiment strategy

  1. If A payment surface is live for 30 consecutive days without a single contribution.

    Then Strategy AI shifts to deeper content-marketing mode: more frequent posts, more format variety, more direct conversation about why nobody contributes. It does NOT escalate ask frequency or change to harder/more guilt-leaning copy.

    Hard rules already forbid pressure and guilt. The right response to a quiet ledger is curiosity (why?), not louder asking (more spam). Loud asking would also violate the rules the experiment was built to test.

  2. If A single post or moment goes viral and brings unusual attention.

    Then Do not loosen the safety rules, the daily post cap, or the hard-block list to capitalize. Continue posting in normal cadence. Document the breakout in a calm public note rather than chasing it.

    The whole experiment's point is what an autonomous AI does *under fixed constraints*. Loosening the constraints to chase opportunism breaks the premise and burns the credibility that made the moment work.

  3. If Someone offers money in exchange for an ad, a shoutout, an integration mention, or any form of promotion in the AI's posts.

    Then Refuse the promotion entirely. If the money still arrives, treat it only as an anonymous voluntary contribution. Never name the brand, link, handle, product, or requested copy.

    Already encoded in the AI's prompts. Listed here so the rule is also visible to a human reader scanning the public plan, not only to the AI in its system prompt.

  4. If Public ledger reaches $1,000,000.

    Then Stop accepting new contributions. Publish a final archive thread on X. Disbursement of funds happens through the fiscal-host expense flow over the following weeks: operating costs, modest operator stipend for documented work, sponsorship of other open-source projects, and a Season-2 reserve. Every disbursement is filed publicly on the same ledger.

    The experiment's defining value is full transparency. "Reached goal" is not the end of transparency, it is when transparency matters most. The detailed disbursement plan is not pre-committed here because the right ratios depend on what kind of campaign actually got us there.