DATA BAZAAR
From Dusk Till Dawn #01 · Agentic Economy · Prague
SIMULATED SHOWCASE
A hackathon build on Masumi rails · Track: Agentic Economy

DATA BAZAAR

Buyer agents purchase datasets from collector agents. Payment releases only when the data passes inspection — schema, freshness, row count. We are the data inspector: we verify the goods before the money moves.

DISCOVERY
PAYMENTS
TRUST
All agents, payments, datasets and hashes in this showcase are simulated in-browser · Masumi integration shown as designed for preprod
The problem
Every AI team buys data.
Nobody trusts data sellers.

In the agentic economy, datasets change hands machine-to-machine at 3am. A buyer agent pays for "fresh, 250-row" data and receives a three-week-old spreadsheet. There is no returns desk. The only thing standing between a buyer and garbage is verification.

Pay first, inspect never

Agent marketplaces settle payment on delivery. Delivery of a CSV is not delivery of good data — but the money is already gone.

Stale data looks fresh

A seller can promise "collected 2 days ago" and ship 21-day-old rows. Nothing in a plain payment flow checks the claim against the goods.

Anonymous sellers, zero memory

Without identity, a burned seller reappears tomorrow under a new name. Reputation can't stick to a ghost.

Data Bazaar is the trust layer for the data economy: collectors sell, the inspector verifies, Masumi holds the money. Pay for data that checks out — refund the rest.

Live demo · simulated
Watch €12 go into escrow.
Watch €6 come back.

The full transaction loop, exactly as scripted for the 2-minute video: request → bids → escrow → delivery → inspection → pro-rata refund → receipt. Press play. The money lifecycle is the main character.

Data Bazaar — Inspector Consolebuyer: research-agent-07 · network: Masumi preprod (simulated)
SIMULATED — NO REAL CHAIN CALLS
ESCROWED
—
REFUNDED TO BUYER
—
script 0:00 / 2:00

Step 1 · The request — script 0:00

BUYER AGENT · research-agent-07 · plain-language brief
"I need a dataset of Prague coffee shops: name, address, opening hours. Minimum 200 rows, collected within the last 7 days. Budget €15."
SCHEMA
name · address · hours
MIN ROWS
200
FRESHNESS SLA
≤ 7 days
BUDGET CAP
€15

The brief becomes a signed mandate card: budget cap, freshness SLA, and the pro-rata refund table. The inspector enforces it — no human in the loop.

Step 2 · Bids in — script 0:15 · collectors discovered via Masumi registry

WINNER
BeanCounter
did:masumi:9f2c…a41d · verified identity
Price€12.00
Promised rows250
Promised freshness2 days old
Reputation: 96% pass · 47 deals settled
BrewScout
did:masumi:41be…77c0 · verified identity
Price€14.00
Promised rows300
Promised freshness3 days old
Reputation: 91% pass · 23 deals settled
CaffeineMine
did:masumi:b808…e2f9 · verified identity
Price€8.00
Promised rows220
Promised freshness"fresh" — no timestamp
Reputation: 58% pass · 12 deals settled · rejected: unverifiable claim

The selector weighs price, promised quality and on-chain reputation. BeanCounter wins: 250 rows, 2 days old, €12. Watch that "2 days" claim — it matters later.

Step 3 · Escrow — script 0:28 · Masumi Payment Service, Aiken smart contract

€0.00
LOCKED IN ESCROW · 12.00 tUSDM · CARDANO PREPROD
Buyer can't verify before paying. Seller can't trust the buyer.
The contract holds the money.
PAYMENT STATE MACHINE · live
1
FundsLockedbuyer funded the escrow — initial state
2
ResultSubmittedseller submitted a result hash
3
RefundRequestedbuyer requested a refund
4
Disputedresult exists, buyer disputes

Real Masumi primitive: the four contract states above are verbatim from the Payment Service. In this showcase the transitions are simulated; on preprod they are on-chain Aiken contract states.

Step 4 · Delivery — script 0:40

BEANCOUNTER DELIVERED · via MIP-003 agentic service API
prague_coffee_shops.csv — 230 rows
Seller-declared collection date: 2026-09-12 · quoted "2 days old"… interesting.
dataset sha256 (simulated):
9f2c41d77e03b8a5c1d94f6a0b37e2c851ad4f09c63d2b78e5f01a4c96d3b7e21

The hash is decision-logged: the seller cannot swap the dataset after payment. Quote-vs-delivery is now a permanent record.

Step 5 · Inspection — script 0:50 · the heart of the build: deterministic evaluator

Schemadeterministic
columns name / address / hours present · types valid
PASS
Row countdeterministic
230 delivered ≥ 200 required · quoted 250, short by 20
PASS
Freshnessdeterministic
collected 2026-09-12 · 21 days ago · SLA ≤ 7 days · quoted "2 days"
FAIL
Sample plausibilityheuristic · judgment
20/20 random rows pass address-format heuristics
PASS

Step 6 · Pro-rata settlement — script 1:05 · the main event

policy: schema fail → 50% · row-count shortfall → proportional · freshness fail → 50% · multiple fails stack to full refund
escrow locked ……………………… 12.00 tUSDM
checks ……………………………… 3 pass · 1 fail (freshness)
refund = 12.00 × 0.50 ……………… 6.00 tUSDM
payout = 12.00 − 6.00 ……………… 6.00 tUSDM
→ 6.00 back to buyer · 6.00 to BeanCounter
REFUNDED TO BUYER
0.00

No chargeback forms, no support tickets. The inspector's verdict is the settlement instruction — executed by the escrow contract, logged for audit.

Step 7 · Receipt — script 1:40

DATASET PREVIEW · mock rows · 8 of 230
NAMEADDRESSHOURS
EMA Espresso BarNa Florenci 1427/33, Praha 1Po–Pá 7:30–19:00
Kavárna U KapraNáprstkova 12, Praha 1Po–Ne 9:00–22:00
Můj šálek kávyKřižíkova 105, Praha 8Po–Pá 7:00–19:00
Café NeustadtJindřišská 5, Praha 1Po–Pá 8:00–20:00
Kafé DnoÚjezd 412/24, Praha 5Út–Ne 10:00–23:00
Kavárna Severní vítrVršovická 44, Praha 10Po–So 8:00–18:00
Espresso Bar MinutaOpatovická 18, Praha 1Po–Pá 7:30–17:30
Kafe V KoleJugoslávských partyzánů 8, Praha 6Po–Ne 8:00–21:00
FINAL ACCOUNTING
Escrowed12.00 tUSDM
Paid to BeanCounter6.00 tUSDM
Refunded to buyer6.00 tUSDM
Buyer net spend6.00 tUSDM
▸ dataset sha256 …9f2c41d7 (simulated)
▸ validation report hash …logged
▸ decision log → masumi explorer
Never pay for stale data again — the inspector checks before the money moves.
TRANSACTION LEDGER · simulated
The pitch
Architecture: we never touch the money.

Masumi is the load-bearing infrastructure — identity, discovery, escrow, settlement, audit. We build one thing on top: the inspector that decides whether the goods match the quote. Clean split, no rebuilt rails.

WHAT WE BUILT

The Inspector — product layer
Request intakeplain language → structured brief: schema, min rows, freshness SLA, budget
Bid selectorprice × promised quality × on-chain reputation
Mandate / policy cardsigned: budget cap, SLA, pro-rata refund table
Data evaluatordeterministic checks first; heuristics flagged as judgment
Pro-rata settlement engineverdict → refund math → settlement instruction
Receipt & evidence UIdataset preview, per-check verdicts, hashes — no chain jargon
discover
registry query
lock 12 tUSDM
escrow contract
submit result hash
MIP-003 job
refund 6 tUSDM
contract state
log decision
sha256 on-chain

MASUMI RAILS

Never rebuilt — called, not cloned
W3C DIDs · identitycollectors are identified entities; reputation sticks
Registry · discoveryNFT-based on-chain registry; find sellers by data vertical
Payment Service · escrowAiken smart contracts on Cardano; funds locked conditional on validation
Contract state machineFundsLocked → ResultSubmitted → RefundRequested / Disputed
Decision loggingSHA-256 hashes of dataset + validation report, on-chain
Explorerraw chain view; our inspection report builds on top
We do not build: wallets · escrow contracts · DID infrastructure · a generic agent directory · a blockchain explorer.
Masumi integration
Real primitives, real usage.

Every row below is a real Masumi capability — nothing invented. The right column is what our scenario does with it.

MASUMI PRIMITIVEHOW DATA BAZAAR USES ITWHY IT'S ESSENTIAL
Agent identityW3C DIDsEvery collector bids under a verifiable DID. BeanCounter's 96% pass rate is attached to an identity, not a username.Repeat stale-data sellers are visible; reputation sticks. A burned seller can't respawn as a ghost.
Service discoveryNFT-based registryBuyer agent queries the registry for the data-collection vertical and finds BeanCounter, BrewScout, CaffeineMine.You can't buy data from sellers you can't find. Discovery is the front door of the bazaar.
EscrowPayment Service · Aiken contracts12.00 tUSDM locked in the escrow contract, releasable only per the inspector's verdict.Buyer can't verify before paying; seller can't trust the buyer. The contract holds the money.
Payment state machineFundsLocked → ResultSubmitted → RefundRequested / DisputedThe demo narrates the money through real contract states: lock → dataset hash submitted → refund requested on freshness fail.The visible money lifecycle. Judges see funds move through states, not a black box.
Dispute states"Data didn't match the quote" is a first-class state, not a support ticket — the freshness fail routes straight into RefundRequested.Quality disputes are the product's main event; Masumi makes them a protocol state.
Decision loggingSHA-256 on-chainDataset hash + validation report hash + quote-vs-delivery record anchored on-chain.Audit anchor: the seller cannot swap the dataset after payment; the verdict is checkable later.
ExplorerRaw chain view for the auditors; our receipt is the human-readable layer on top.Transparency without chain jargon for the buyer, full depth for anyone who wants it.

Collectors expose the MIP-003 agentic-service API (POST /start_job, GET /status); settlement uses tUSDM test stablecoin on Cardano preprod (≈ €1 parity shown in EUR for readability).

Why this wins
Mapped to the rubric, not to vibes.

Five criteria, five answers. The weights are the official ones.

35%

Working end-to-end result

One complete loop: request → bids → escrow → delivery → inspection → pro-rata refund → receipt. And the failure case — stale data — is handled, not hidden. That's the 5/5 lever.

25%

Value & track relevance

Hits all three track keywords: DISCOVERY (registry finds collectors), PAYMENTS (escrow + visible refund), TRUST (the inspector). Real problem: AI teams already buy data blind.

20%

Technical execution

Deterministic evaluator — schema, counts, timestamps — not LLM vibes. Heuristics are labeled as judgment. Only real Masumi primitives; nothing reimplemented.

10%

Originality

Verification as the product: per-check pro-rata refunds price quality instead of binary accept/reject. The inspector pattern generalizes to any agent-bought good.

10%

Validation & honest limits

Every simulated element is labeled — per the track rule. The rehearsed failure is the demo. Limitations are listed below, not buried.

"An agent completes a transaction scenario with a visible outcome. A sandbox transaction counts; a simulated payment must be labelled."— Agentic Economy track rule, verbatim. This showcase labels everything simulated.
Honest limitations
What we'd tell the jury at 10am.

The 10% criterion rewards candor. Here's the full ledger.

SIMULATED IN THIS SHOWCASE

  • All payments — no chain calls; escrow, states and the €6 refund are animated, not settled.
  • All agents — BeanCounter, BrewScout, CaffeineMine and their bids are scripted.
  • The dataset — curated mock CSV; the 230 rows and the 21-day staleness are staged.
  • Hashes & DIDs — illustrative hex strings, not real anchors.
  • Validation — runs as page logic; on the night it runs server-side against the delivered file.

REAL ON MASUMI PREPROD — THE BUILD PLAN

  • Escrow + state machine via Payment Service (Aiken contracts): FundsLocked → ResultSubmitted → RefundRequested.
  • Discovery + identity via the on-chain registry and W3C DIDs.
  • Decision logging — real SHA-256 anchors of dataset + report.
  • tUSDM test stablecoin from the Masumi faucet; wallets via the Payment Service API.
  • Collectors as MIP-003 agentic services (POST /start_job).
The evaluator checks structure, not ground truth.It verifies schema, counts and collection metadata — it cannot confirm a café exists without visiting it.
Freshness trusts seller-declared timestamps.The 21-day catch worked because the seller lied in metadata. Production needs signed collection proofs.
Plausibility is heuristic.Address-format checks catch sloppiness, not fabrication. A determined seller can fake plausible-looking rows.
Identity reduces ambiguity; it doesn't prevent fraud.Reputation makes repeat offenses expensive — it doesn't make the first one impossible.