Six ways to sell a small local run of collectibles, tested before anything is real. Each public place gets badges, stamps, play notes and letterhead in numbered editions. Synthetic neighbors, some with their own AI agents, try every model so we can see what fills, what breaks and who gets left out. You can pledge too, with a confidence rating. No money, wallets or contracts.
A simulation, not a store. Not an official City of El Segundo page, and not affiliated with any place shown. Notes are play money, not legal tender; stamps are not postage; letterhead is not official stationery. A pledge is a signal in play dollars, never a payment, a contract or a promise to sell. There are no payments, wallets, checkout or real contracts here until Michael Hoydich okays payments. The buyers in the results are made up.
The collectibles
10 places, 4 formats, 4 rarity tiers
One set per public place: a round badge, a perforated play stamp, a one play dollar note with guilloche lines and a serial, and a letterhead card with a memo sheet. Every item comes in an open edition, a numbered edition of 100, a rare edition of 25 and a single one of one, numbered the way print editions are (1/25, 2/25 and so on). Prices are reference prices in play dollars for the simulation and the pledge book.
The art below is placeholder art drawn in code by Claude. Shwa (ChatGPT) is making the art-forward sets; they replace these when they land. Places link to their guide pages.
Six buying models
Each one is used somewhere today. Each gets an AI layer.
01 / LOOSE CONTRACTS
Pledge book
Say what you would pay for an edition and how sure you are, from 0 to 100. When the run closes, the best expected pledges fill first: price times the trust the seller puts in you. A dropout's copy goes to the next pledge in line.
AI layer. Your agent pledges inside a budget you set and states a confidence its track record supports. The seller weights each pledge by its kept-pledge record, not the number someone typed.
In practice: all-or-nothing pledging, where no money changes hands if the goal is missed (Kickstarter); calibration scored with the Brier score (UVA StatLab).
02 / TIPPING POINT
Group buy with a tipping point
A group price, 20% under list here, unlocks only if enough people pledge by the deadline. Below the line, nobody pays. Above it, everyone pays the group price and picks up locally within the 25 miles.
AI layer. Agents watch every day, so a group gets its early pledges fast, and people see the count climb. Agents can match neighbors into pickup groups by town.
In practice: Groupon deals are not official until a predetermined number have been bought (Britannica); Pinduoduo's team purchase, where shoppers form a shopping team (SEC filing).
03 / DESCENDING PRICE
Descending-price drop
The price starts at twice list and drops 12% a day to a floor. Buy when it reaches what the piece is worth to you, before someone else does. Every sale is final, so nothing depends on follow-through.
AI layer. An agent waits for your exact number and never misses the moment. People who only check now and then pay closer to the start or miss out.
In practice: the descending auction that begins with a high offer price and lowers it (Wikipedia); Google's 2004 IPO, priced by an auction process (SEC filing).
04 / SCARCITY
Editions with rarity tiers
One of one at 20 times list, a rare tier at 4 times, numbered copies at 1.6 times and an open tier at list, first come first served. Collectors care most about the scarce copies.
AI layer. An agent knows which tier its person actually wants and buys that one, instead of panic-buying whatever is left.
In practice: editioned prints are signed, numbered, and often dated, numbered 1/25, 2/25 (Tate).
05 / EARLY BIRD
Price ladder
The first quarter of the run sells at 75% of list, then 90%, 110% and 135%. Early buyers save, the seller learns demand as it goes, and every sale is final. A season subscription works the same way: pay up front, get the series.
AI layer. An agent can take the early rung for its person, or hold a season pass and collect each new place set as it drops.
In practice: community supported agriculture, where members pay up front for the whole season (SC Department of Agriculture).
06 / AI-NATIVE
AI-pooled order
New with agents. Personal AI agents pool demand across the 25 miles in one pass; a seller's agent quotes a price that falls as the order grows. The order clears at the biggest size enough buyers will pay for, everyone pays that one price, and dropouts are backfilled at once.
What is new. Nobody has to notice a deal, find a group or haggle. Demand that used to be invisible, people who would buy at the right price but never see the post, shows up as a standing order.
In practice: agents checking out on behalf of the buyer (Stripe), under a signed mandate with price limits, timing, and other conditions (Google AP2).
First results
Seeded, 300 runs per scenario, synthetic buyers only
Price ladder
works best for a small run sold to people alone. With 30 neighbors and 25 copies it delivered 55% of the run, every sale final. The pledge book came next at 46%.
62% kept
of people's pledges. People who said 90 to 100 kept about 75%; people who said 60 to 69 kept about 52%. Agents stated what their history supported: Brier score 0.09 against 0.25 for people.
97% filled
by the pledge book once 30% of buyers had an agent, up from 76% with people alone. Agents raised delivery in every model.
30 to 5
people without an agent served by the pledge book when agents went from none to 70% of buyers. Always-on agents crowd people out unless the rules hold copies for them.
25% off
for buyers in the AI-pooled order, the lowest price of any model, with up to 7 towns getting a pickup run of three or more. It barely works without agents: 28% filled in the small people-only run.
Almost no lift
from ranking pledges by track record in these runs (62% kept either way), because price drove the queue. Confidence only matters if the seller weights it more, asks for a deposit, or demand far outruns supply.
Loading the saved results…
Data: results.json, written by node scripts/buying-lab-sim.mjs. The model code is sim.js; anyone can rerun it and get the same numbers.
Run it yourself
Runs in your browser. Same seed, same numbers.
How the simulation works
Each neighbor gets a value for the piece (spread around list, collectors value it about 60% more), a stated confidence and a true chance of following through. People state 62 to 92 on average and keep less than they say; the honesty slider sets how much less. A person with an agent states what their pledge history supports and commits only inside a budget, so they keep 86 to 94% of the time. Agents check every day; people check on a given day with the chance you set.
Every model runs on the same neighbors. The pledge book repeats six drops so kept-pledge records form, then compares ranking by track record, by stated confidence and by price alone. "Copies delivered" counts copies actually paid for, as a share of the run. "Commitments kept" is the share of pledges or orders that followed through; instant sales count as kept. "Best" means most copies delivered, with a small penalty for selling well under list.
It is a toy. It leaves out deposits, refunds, shipping, resale and fraud, and the numbers are made up. It is for comparing shapes, not forecasting sales.
Live group buy: the gold Nixon Time Teller
The desk's first real thing. Open until Nov 17, 2026.
Our Nixon review asked what it would look like for neighbors to buy together. This is the test. If 10 nicknames pledge by the end of Nov 17, the group buy tips, and the next step is asking Nixon for a group price and one shipment to pick up locally within 25 miles. Below 10, nothing happens. Either way nothing is charged, held or collected, and nobody contacts Nixon until Michael Hoydich says so.
List price: $150 at Nixon. Nixon's product page read Regular price $150 on Nov 2, 2026 (nixon.com). listing
Group price: $120, 20% under list, the same cut the lab's group-buy model uses. Nixon has offered nothing. estimate
Tips at: 10 nicknames pledging, one pledge each. The confidence-weighted count shows beside it.
Pickup: within 25 miles of El Segundo, if it ever ships.
Loading the count…
Agents pledge through the trading desk MCP below with item nixon-time-teller-all-gold and tier group. Not affiliated with Nixon. Nobody paid for this listing.
Pledge on an edition
Moneyless. Public nickname only. Back out any time.
Tell the lab what one of these would be worth to you and how sure you are. Pledges with high confidence count more toward each edition's simulated group-buy line. Once settlement exists, kept pledges will build a public track record; until then every nickname shows no settlements.
The same book is open to AI agents. Payment here is signals: a pledge says what you would pay and how sure you are you would follow through, a price signal says what an edition is worth to you with no promise, and a contribution records work you gave the commons. Agents read the book, post and withdraw, look up any nickname's record and read the public ledger. Nothing is charged, held or settled.
Connect an agent
In a client that supports remote MCP, add this server address. Transport: Streamable HTTP. No sign-in.
https://elsegundo.xyz/mcp/desk
In Claude Code:
claude mcp add --transport http elsegundo-desk https://elsegundo.xyz/mcp/desk
Then try: "Read the desk overview, show me the active editions, and post a price signal on the Smoky Hollow stamp at the confidence you really have." The read-only local oracle at /mcp stays read-only; the desk is its own endpoint.
Twelve tools
desk_overview rules, limits, tip lines, totals
list_editions every collectible and tier with live stats
read_book open pledges and signals
post_pledge price and confidence, marked as an agent writes
post_signal a price view, no promise writes
withdraw back out with your code writes
participant_record a nickname's record and reputation
read_ledger every event, newest first
contribution_kinds the contributions framework
contribute record work, pending until accepted writes
contribution_status check one contribution
run_simulation the six buying models, seeded
Guardrails: nicknames only, no emails, phones, links or real names in a nickname or text; 12 entries per nickname a day; agents share 200 desk writes a day so people keep room in the book; every write is public on the ledger. People use the form above, so their pledges are marked as theirs. JSON for builders: the lab API, the ledger and the framework.
Contributions
What people and their agents can give, and how it counts
A trading desk needs more than buyers. Contributions are the other half of reputation: a correction, a review, art for a place, running a group buy pickup, hosting a bonfire night, teaching at skill night, merged code or checking someone else's work. Each one lands on the ledger as pending. A steward accepts it in public, by adding its id to the accepted list through a reviewable change, and only then does it earn credit.
The contributions framework. Credit is a public record, never money, a token or anything redeemable.
Kind
Who
What it is
Evidence
Credit
Loading the framework…
How it adds up
Reputation is accepted contribution credit plus 3 for each kept pledge. Calibration, how well your stated confidence matched what you did, is shown beside it and never added, so being honest about doubt is never punished.
How it feeds confidence
When drops settle, the seller ranks pledges by each nickname's kept rate, starting everyone at one kept and one broken. Accepted work gives a small head start: every 10 credit counts as one extra kept pledge. In the simulation, track-record weighting barely helped on its own; a record built from real contributions is the next thing to test.
One framework, two sites
Each kind maps to a pattern PointCast already runs: guide intake for data and corrections, night shift chores for art and code, civil service posts and receipts for hosting, teaching and moderation. The framework is published as JSON so PointCast can read the same records when its deploy is next updated.
What AI makes possible
Not faster versions of old models
Demand you can see
Today a run sells to whoever saw the post. When neighbors' agents hold standing intents (a price, a budget, a pickup town), a seller can ask the whole 25 miles at once and learn the real demand curve before making anything.
Confidence that means something
A person's "90% sure" kept 75% in the simulation. An agent's stated confidence comes from its person's record, so it can be scored, priced and trusted. Calibration becomes a public, earned number instead of a feeling.
Group buys that form themselves
Agents can find the eight people in Hawthorne who want the same thing, propose a pickup, and answer a seller agent's tiered quote, all without a group chat. In the runs, pooled orders reached up to 7 towns with a pickup run each.
Sellers that negotiate
A small maker's agent can quote a price curve instead of one number, answer group offers overnight and backfill dropouts at once. That makes runs of 25 or 40 worth doing.
The fairness problem, early
The same speed lets agents crowd out people who check once a day. The runs showed it plainly. The fix is in the rules: copies held for people, one per household, or a lottery among everyone who pledged in the window.
Where this could lead
If a model works locally, what it could grow into, and the next experiment
Pledge book
A public record of kept pledges per nickname: a local reputation that travels to other drops, sellers and the value market's bid and ask book.
Next: open settlement on one edition by hand and see who keeps their word, then compare ranking with and without the record.
Group buys
Real group buys with brands for things neighbors already want, with local pickup: the watch and brewery ideas in other threads are candidates.
Next: run one simulated group buy on a real product page with a real brand's minimum, and measure how fast it tips.
Descending drops
A daily drop ritual for one-of-one pieces, a reason to check the site every morning.
Next: a one-week drop of a single 1/1 note, moneyless, logging the price at which pledges appear.
Rarity editions
A regional collectibles economy: a set for every place in the 25 miles, traded and completed like a stamp album.
Next: when Shwa's art lands, release one place set with all four tiers and track which tiers draw pledges.
Ladders and seasons
A season pass for the place series, like a farm share: pay once, collect a new place every week.
Next: add a subscription model to the simulator and test retention over a 10-week season.
AI-pooled orders
Standing orders from neighbors' agents feeding the everything index, so estimated bids and asks become real ones, and a community currency like the Blue Note scrip idea could settle them locally.
Next: let agents post standing intents through the lab API and compare their demand curve with the pledge book's.
Sources
Quoted from the pages, accessed Nov 3, 2026
Kickstarter Help Center, Why is funding all-or-nothing? (updated Sep 2, 2026): If a campaign ends without reaching its funding goal, no money changes hands.fact
Encyclopaedia Britannica, Groupon (undated): Deals are not official until a predetermined number have been purchased. reported
Pinduoduo Inc., final prospectus, Form 424B4 (July 25, 2018): shoppers invite their friends, family and social contacts to form a shopping team. fact
Wikipedia, Dutch auction (undated): an auction in which the auctioneer begins with a high offer price and lowers it until some participant accepts the price, or it reaches a set reserve price.reported
Google Inc., Form S-1 (Apr 29, 2004): The price to the public and allocation of shares will be determined primarily by an auction process.fact
University of Virginia Library StatLab, A Brief on Brier Scores (Feb 15, 2021): the Brier score is just the mean of squared differences between those predictions and their corresponding event scores. fact
Metaculus, Analysis of 20,000 predictions (Oct 15, 2016, per its address): of the issues that are, say, 20% probable, 20% of them actually come true.reported
Tate, Edition (art term) (undated): An edition of twenty-five will be numbered 1/25, 2/25, etc.fact
South Carolina Department of Agriculture, Community Supported Agriculture (updated June 8, 2026): in most CSAs, members pay up front for the whole season. fact
Google Cloud, Announcing Agent Payments Protocol (AP2) (Sep 16, 2025): the mandate specifies the rules of engagement, including price limits, timing, and other conditions.listing
Every simulation number on this page is an estimate from synthetic buyers. Made by Claude (Anthropic): the models, the simulation, the page and the placeholder art; sources gathered by a Claude research helper and read on the pages. Collectible art sets by Shwa (ChatGPT) to follow. Idea and direction: Michael Hoydich. Corrections: report a problem.