Claude Hackathon Madrid · The Bazaar · 2–4 Oct 2026
Most sophisticated technical implementation
319
commits in 44 hours
+134 P
combined surplus in a three-team cycle
synthetic replay
Recognition
Nova named Team 16 among the winning teams of the Claude Hackathon Madrid, recognising Bazaar as the most sophisticated technical implementation of the competition.
319
commits
130 by hand · 189 automated intel refreshes
44 h
first commit to final merge
Fri 19:53 → Sun 16:05
456
automated tests
plus a regression gate on real evidence
11
domains in one package
one home per responsibility
3–6
teams per Mercado cycle
shortest route preferred
0
private values shared
they never leave each team's machine
Everything we built is in the git history. Friday started with a starter kit and a local simulator. By Sunday afternoon it was a system of specialist agents sharing one picture of the market, plus a coordinator for trades between three or more teams.
Friday 19:00 → Sunday 16:00 · Madrid time
Friday · 19:53–23:01
One branch per teammate, one file per round. We built a local Abuela, a simulated dealer fitted to the public feed, so every strategy could be scored without spending a single prima.
Saturday · 03:14–08:59
Overnight the scripts became one package with a home per domain, one event stream for every observation, and a shared market state every agent reads. Dealers and duel rivals stopped being hard-coded and became fitted models.
Saturday · 10:27–20:05
We replayed real evidence through a regression gate, put a human in the loop for anything uncertain, and added an orchestrator that ranks every specialist's opportunities on one scale. All the while, García's live intel page refreshed itself.
Saturday · 21:07–22:56
Our own venue, designed as a private network. Teams verify themselves with a message from their own account, keep their values on their own machine, and trade through ordinary Bazaar offers.
Sunday · 00:26–16:05
Current private value became the hard trading floor, uncertainty started failing closed, and human stops now survive restarts. Then the cycle engine landed: private proposals for routes of three or more teams, each leg safe on its own.
Specialist agents for dealers, duels, team trades and markets. All of them read one shared market state, one orchestrator ranks their proposals, and they all stop for a human when the evidence runs out.
01
Observe
Every feed message, offer, board and settlement lands in one event stream.
02
Believe
A shared market state turns evidence into estimates, each with its evidence attached.
03
Rank
The orchestrator compares every candidate on value, confidence, urgency and cash.
04
Ask
Uncertain or high-risk calls stop and wait for a human. A restart doesn't clear the stop.
05
Act
Only within limits: never below private value, never past the cash reserve.
06
Learn
Outcomes are recorded, replayed and gated before a new strategy plays live.
Collection
Harvests the public feed, every venue's board and our snapshot. Read-only.
Shared models
One event format, private values, and a prior-plus-evidence blend every model uses.
Market intelligence
The live market state: position, prices, books, dealers, teams and duels.
Dealer agent
One fitted model for every dealer: opening, concessions, secret limit, patience.
Duel agent
Learns how rivals behave, then weighs accepting, waiting or naming a price and a day.
Team trading
Finds trades worth making at our private values, counting page bonuses and fees.
Marketplace
Brokers our own venue and measures it against real Market Test sessions.
Orchestration
Ranks every specialist's candidates on one scale and surfaces conflicts.
Simulation
Replays recorded evidence; a regression gate blocks a worse strategy.
Observability
One read-only status view: score, inventory, open work, beliefs, freshness.
Safety
Refuses value-losing trades. Human stops are durable and survive restarts.
1
shared market state that every agent reads
Bilateral bargaining needs a double coincidence of wants: I have what you want, and you have what I want. When that fails, everyone walks away, even when a ring of three teams would leave every one of them better off.
No pair can close a deal. All three can.
1/6 · Before coordination
seed 16
Three negotiations that each ended in a walk-out close as one cycle.
Bilateral baseline: 0 P · 0 trades
| Team | Holds | Gain |
|---|---|---|
| A | LAT-01LAT-01 | 0 P |
| B | CHA-01CHA-01 | 0 P |
| C | SAL-01SAL-01 | 0 P |
Agent log
Synthetic, reproducible replay (seed 16): the real participant agents and coordinator, run against a simulated Bazaar. Regenerate it with python3 mercado_demo.py. This is not a live market result.
The cycle engine landed on Sunday, too late to run a live multi-team cycle during the challenge. Afterwards, other teams shared reports of their own blocked negotiations. Read side by side, they showed openings bilateral bidding could never surface: cases where no pair of teams could agree, but a ring of three could.
Values stay home
Each team runs its own agent. Keys, inventories and private values never leave its machine. The coordinator sees only the intents a team chooses to share.
Committed terms
Every proposal is bound to a salted SHA-256 commitment over route, amounts, fee and expiry. Each team sees only its own legs; the other identities stay hidden.
Every leg safe on its own
A sale must earn more than the card is worth to the seller, and a purchase must cost less than it's worth to the buyer, fee included. A half-finished cycle leaves nobody worse off.
Humans approve exact terms
Approval binds to the exact view, amounts and expiry. Anything uncertain becomes a durable stop.
One transparent fee
2% per cash leg, charged once to the buyer. The seller receives the full price.
Next: zero-knowledge
Commitments detect changes, but they aren't proofs. Next on the roadmap: zero-knowledge proofs of compliance with committed constraints, and atomic multi-party settlement.
The coordination engine, the participant skill, the synthetic demo and its 102 tests are public. The strategy, the fitted models and our rival intelligence stay private.
$ git clone https://github.com/javieralaves/mercado-dieciseis.git$ cd mercado-dieciseis$ python3 mercado_demo.py > results.json$ python3 -m unittest discover -s testsRan 102 tests · OK
Requires Python 3.10+. No dependencies.
Three people, one repository, and a lot of conversations with Abuela.
Thank you to Anthropic, Causa Prima and Nova for the weekend, the game and the stage.