MATCH · EXPLAIN · NEGOTIATE
Agent marketplace: compare eligible offers transparently
Inspect seven synthetic registry entries. Apply hard eligibility gates before ranking and proposing a compatible local contract.
Scroll the diagram horizontally to inspect every label. The panels below give the same state as text.
Candidate inspector
Your experiment
Compare synthetic offer cost, latency and observed success without treating any as real marketplace prices. Disable the curator gate to expose why a compelling score is not approval.
Ranked eligible candidates
Proposed local contract
Explain the exact score
Latency term = 1 − fixture p95 / 500. Cost term = 1 − fictional cost / 10. Observed-success term = (successes + 1) / (observations + 2). Weighted score = sum(weight × term) / sum(weights). All fixture terms are in [0,1]. If all weights are zero, every score is zero; agent ID gives a deterministic tie-break.
None of these terms can override tenant scope, curator policy, required skill/format/delivery, metadata TTL, cost budget or minimum observations. Inspection playback only reveals the computation; it does not advance metadata age. Use the separate age control for that experiment.
Offline registry-selection policy with synthetic entries, latencies, observations and fictional cost units. This is not a production marketplace, ADK registry API, real Agent Card schema validator, signature verifier, authentication system or remote-agent invocation. Curator approval is a fixture flag, not cryptographic proof. A high historical success estimate does not establish trust, safety or future reliability; no statistical confidence claim is made. Metadata freshness is not a liveness check. A2A does not prescribe this ranking formula or a standard curated-registry API. The local proposed contract is descriptive JSON, not an A2A wire request. Agent discovery and curated registry concepts.
Eligibility first, then a disclosed ranking
A marketplace answers two separate questions: which offers may serve this request at all, and in what order the rest are shown. Seven fictional listings pass five hard gates before a sixth stage ranks the survivors. With the opening settings only scout and ranger remain. atlas fails both the budget (7 cost units against 5) and the evidence gate (1 observation against 10), echo offers neither retrieve nor JSON, shade carries 180-second-old metadata against a 60-second limit, unreviewed lacks curator approval, and tenant-b belongs to Team B.
The opening ranking is price only
The page opens with Latency weight 0, Cost weight 100 and Observed-success weight 0, so the score is 1 minus cost divided by 10. ranger (1 unit) scores 90.00, scout (3 units) 70.00, and the proposed contract names ranger. Common baseline settings does not restore these values: it sets the weights to 40, 30 and 30, and scout then wins 77.31 to 77.06, a margin set entirely by the weights. Latency weight 100 with Cost weight 0 puts scout ahead 76.00 to 52.00. Observed-success weight 100 alone puts ranger ahead 97.52 to 86.36: the success term is (successes + 1) / (observations + 2), and ranger has 196 of 200 against 18 of 20.
Ties and offers that should stay hidden
Raise Maximum metadata age (seconds) to 180. shade passes at the inclusive boundary and ties ranger at 90.00, and the tie is broken by agent ID, so ranger wins. A transparent marketplace must disclose that rule, because it decides the sale. Return the limit to 60 and clear Require curator-approved fixture: unreviewed also ties at 90.00. Restore approval and set Advance metadata age (seconds) to 30. ranger, 40 seconds old, now reads 70 and expires, leaving scout alone at 70.00. At 50 every listing is stale or excluded and no contract is proposed.
Every remaining control
Tenant Team B leaves only tenant-b, at 60.00. Required skill Summarize leaves scout. Translate with application/json leaves nobody, because echo returns only text/plain; choose text/plain under Required output format and echo wins at 80.00. Require streaming delivery removes ranger. Fictional cost-unit budget 1 removes scout. Budget 7 with Minimum fixture observations 0 admits atlas at 30.00; add Latency weight 100 and Cost weight 0 and atlas wins at 82.00 on a single observation. Minimum fixture observations 200 leaves only ranger; 250 excludes everyone. Inspect candidate only chooses which record is shown.
Playback
Play reveals one gate every 10 seconds of the 60-second inspection, and Pause holds it: scope at 10 seconds, ranking and the contract only at 60. Next gate jumps to the next boundary, Restart inspection rewinds, Speed runs from 0.25x to 2x, and Inspection time scrubs. The remaining list follows fixture order, not rank order. The preview states the winner at once; editing an input recomputes the run and replays it.
Assumptions and limits
The listings are static fixtures, not an ADK API, an A2A registry or real vendors. Approval is a flag, not a verified review, and success counts are invented. There is no bidding, sponsored placement, negotiation, payment or identity check, and the score is a linear weighted sum. A real marketplace must also label paid placement and publish its ranking parameters.