QwenCloud Hackathon · Track 3 · Agent Society
Tarmac is an airline irregular-ops agent society where conflict is the architecture: five agents with opposed objectives negotiate sealed-bid claims on a row-locked seat ledger, and a mediator resolves mechanically detected deadlocks with Ed25519-signed, regulation-cited rulings. An ablation bench proves the society beats a single planner — measurably.
SCENARIO storm_dfw PAX 180 / SEATS 139 OFFLINE zero keys REPLAY byte-identical
0→0
Protected passengers stranded — single planner vs the society
docs/BENCH.md · 10 seeds
0%→0%
Special-needs SLA met — minors, medical, wheelchair
docs/BENCH.md · 10 seeds
0
Tests, all green, in ~5 seconds
pytest · CI on every push
0%
Coverage across ledger, bids, deadlock, mediator, invariants
pytest --cov=tarmac_society
The one number
Same storm, three conditions, ten seeds, fully deterministic. The society doesn't just beat a single planner — remove its mediator and the society becomes worse than one planner, because contested claims get quarantined with no one to rule. The mediator is load-bearing, not decorative.
Protected passengers stranded (medical courier · minor · wheelchair · tight connections — lower is better)
Stranded overnight — no seat at all (lower is better)
Seat supply is physically fixed — 139 seats for 180 passengers — so raw strandings barely move. The society's win is in who flies: the single planner seats a flexible high-fare passenger on the last compliant nonstop while stranding the medical courier; the society doesn't.
| Metric | Single planner | Society − mediator | Full society |
|---|---|---|---|
| Protected pax stranded (SLA-failed) ↓ | 17 | 14 | 3 |
| Stranded overnight (no seat) ↓ | 52 [50–53] | 96 [81–109] | 50 [49–52] |
| Tight connections saved ↑ | 0 | 3 | 9 |
| Special-needs SLA met (%) ↑ | 0.0 | 0.0 | 100.0 |
| Crew duty violations (invariant I2) ↓ | 1 | 0 | 0 |
| Rounds to quiescence ↓ | 1 | 6 | 5 |
| Contest stake (credibility) ↓ | 0 | 60 | 40 |
Medians across 10 seeds; [q1–q3] IQR. Offline deterministic policy agents; identical scenario per seed across all three conditions. Reproduce: tarmac bench.
The protocol
Contention has physics: a claim is a typed mutation against a real ledger. Two threads racing the last seat — exactly one wins. Tested.
In contested rounds every agent posts SHA256(claim ‖ nonce) before anyone reveals — rivals can't adapt a bid after seeing another's. The ledger rejects any reveal that doesn't re-derive its digest (invariant I4). Negotiation that's provably leak-free.
A wait-for cycle or a resource contested ≥2 rounds triggers mediation. Never vibes-based — always reproducible.
Every mediator ruling is signed and cites a regulation — a portable artifact you can verify with the public key, without trusting the database.
SQLite row locks enforce capacity and exclusivity (I1). Double-claims are rejected at the storage layer, and every event extends a hash-chained log — replay re-derives the identical manifest (I5).
The protocol ships as tarmac-society — ClaimLedger, DeadlockDetector, Mediator. examples/meeting_rooms.py reuses it for room booking in 20 lines. Not airline-shaped.
Honest by design
Every claim on this page is committed to the repo and checked in CI. Here's the fine print, up front.
Offline default
The default run uses deterministic policy agents — no network, no key, byte-identical every time. They genuinely negotiate, deadlock, and resolve on the fixture. Add --live + DASHSCOPE_API_KEY to swap in the real Qwen agents (qwen3.7-plus roles, qwen3.7-max mediator).
Wire-tested, not vaporware
The LiveQwen DashScope transport is exercised by 19 deterministic tests against a stub client — model routing, structured output, one-retry, the mediator's thinking flag, embeddings. --live executes an already-verified path; it adds tokens, not code.
A simulator, stated plainly
Synthetic reservation data, one curated storm plus a seeded generator. Decision-support and rehearsal — not a live booking integration. ECS deployment is documented in infra/ecs/, not claimed as deployed.
Flight information
The ablation's middle column is the proof: run the identical society without adjudication and contested claims are quarantined at the round cap — 96 passengers stranded overnight versus 52 for a lone planner and 50 for the full society. Deleting one component makes the system worse than the baseline; that's an existence proof the component does the work.
Yes — that's the designed judge path. python scripts/verify_offline.py disables all sockets, runs the full society on the committed fixture, replays the manifest from the hash-chained log, and re-checks invariants I1–I5. Exit 0 means the run reproduced itself byte-for-byte with zero network.
A credibility currency: contesting a claim costs points, winning refunds them plus a premium, losing burns them. Argument is economically bounded — the bench reports the stake spent (40 for the full society vs 60 without a mediator), and rounds to quiescence stay finite.
The society must be measured, not just demoed. qwen3.7-plus keeps five role personas affordable across ~60 turns per run; qwen3.7-max with thinking adjudicates five conflicting position papers; structured output makes claims mechanically executable; context cache reuses the 4k-token storm prefix every turn; the Batch API prices a 60-run ablation. Remove that stack and the benchmark becomes economically impossible.
No. The negotiation core — ClaimLedger, DeadlockDetector, Commitment, Mediator — is domain-agnostic. examples/meeting_rooms.py rebuilds meeting-room booking on the same protocol in about 20 lines. Anywhere agents contend for scarce, exclusive resources, the physics transfer.
327 tests, 99% coverage, green in ~5 seconds — covering ledger locking and double-claim rejection, commit→reveal mismatch rejection, the deadlock detector on synthetic wait-for graphs, credibility math, the ruling flow, invariants I1–I5, and a fixture guarantee that storm_dfw always produces at least one genuine deadlock. CI runs ruff, mypy, pytest, CodeQL, pip-audit, TruffleHog, and a wheel-install smoke test on every push.
Cleared for departure
One command verifies the whole society offline — sockets disabled, invariants replayed, exit 0.