ECCA // Playfair — Tripartite Game Report

What is this? Runtime config Regions Timeline Tokens Agents Scenarios Needlecasts Residues How to read

WHAT IS THIS? why we run it

Playfair is a 50-epoch, 3-region, 6-agent tripartite-game test. Each region has a different cost profile (storage cheap, compute cheap, bandwidth cheap respectively), and agents are assigned a home region whose profile matches their workload.

The orchestrator drives realistic activity (perceives, stores, routes, syncs) and scripts dramatic events (spot preemption, drift spike, residue injection) to force cross-region needlecasts. After each epoch the TripartiteGame contract is audited via verifyAllocationFair(). The test passes only if every epoch stays within its per-region budgets.

This is the closest thing in the repo to a system-level integration test: it exercises the full stack — the EVM (cortex), the DAG (hippocampus), PoW (medulla), the API (siyana), the router (thalamus), the bus (NATS), the registry (Postgres), and the cross-region latency profile injected with tc netem.

RUNTIME CONFIG reproducibility

Clusterplayfair

Epochs50

Epoch interval4000 ms

Agents6

Regions3

Latency profilestorage↔compute 33±5ms · compute↔bandwidth 42±8ms · storage↔bandwidth 75±12ms

Started2026-06-21T07:44:35.590Z

Ended2026-06-21T07:47:39.085Z

Duration183.5 s

Commit258683d922ad0b288b181aba9ad4bdf299f8bef1

Runnergithub-actions/Linux

REGIONS cost profiles & budgets

region-storage

compute cost: high storage cost: low bandwidth cost: medium

budget — C=200 S=1000 B=400

1 agents currently here

region-compute

compute cost: low storage cost: high bandwidth cost: medium

budget — C=1000 S=200 B=400

4 agents currently here

region-bandwidth

compute cost: high storage cost: high bandwidth cost: low

budget — C=200 S=200 B=1000

1 agents currently here

EPOCH TIMELINE activity per epoch · red dots = unfair

TOKEN USAGE BY REGION where each region's budget went

AGENTS cumulative perceived/stored/routed/synced

Archivist-Alpha memory

Memory keeper — pins shards, reconciles, keeps fidelity high. Exploits cheap storage in this region.

18perceived

15stored

0routed

2synced

⚡ 72 💾 45 📡 0

📍 storage drift: 8

CPV: [0.4, 1.8, 1.0, 0.6, 1.2]

Archivist-Beta human

Slow narrative agent — writes detailed memories, rarely moves. Natural fit for cheap storage.

34perceived

16stored

1routed

4synced

⚡ 77 💾 48 📡 5

⚡ storage → compute drift: 12

CPV: [0.5, 1.6, 0.8, 0.5, 1.6]

Inference-Prime ai

Fast AI agent — burns through compute budget quickly. Storage is expensive so stores selectively.

40perceived

7stored

3routed

7synced

⚡ 320 💾 21 📡 17

📍 compute drift: 10

CPV: [1.9, 0.3, 0.8, 0.5, 0.5]

Inference-Echo ai

AI agent that frequently needlecasts to storage region when memory budget runs low.

35perceived

8stored

5routed

5synced

⚡ 280 💾 24 📡 27

📍 compute drift: 10

CPV: [1.7, 0.4, 0.9, 0.8, 0.2]

Router-Nexus mining

Routing specialist — exploits cheap bandwidth. Needlecasts frequently, syncs aggressively.

10perceived

1stored

14routed

0synced

⚡ 40 💾 3 📡 71

📍 bandwidth drift: 10

CPV: [0.3, 0.3, 1.5, 1.8, 0.1]

Router-Sentinel memory

Bandwidth-region sentinel — watches for residues, routes corrections, earns ResidueToken.

16perceived

0stored

7routed

2synced

⚡ 64 💾 0 📡 36

⚡ bandwidth → compute drift: 6

CPV: [0.4, 0.5, 1.2, 1.6, 0.3]

SCENARIO EVENTS scripted at fixed epochs

E5 spot-preemption Spot instance preempted in compute region — agent must needlecast to storage region

E8 respawn Inference-Prime re-sleeves in storage region (expensive compute, cheap storage)

E15 needlecast Spot instance available again in compute region — needlecast back

E20 drift-spike Human agent goes idle for 5 epochs — drift accumulates

E25 sync-recovery Human agent returns, syncs, burns SyncToken to recover

E30 residue-inject Simulated shard loss in bandwidth region — first responder earns bounty

E35 needlecast Inference-Echo migrates to bandwidth for cheaper routing during high-needlecast phase

E40 epoch-surge All agents perceive at max rate for 5 epochs — stress test

E45 needlecast Inference-Echo returns to compute region as surge subsides

NEEDLECASTS cross-region migrations

E14 Router-Nexus bandwidth→storage 5.1 RTE · 1 shards

E14 Router-Sentinel bandwidth→compute 5.1 RTE · 1 shards

E15 Inference-Prime storage→compute 7.1 RTE · 1 shards

E17 Inference-Echo compute→bandwidth 5.1 RTE · 1 shards

E17 Router-Sentinel compute→storage 5.1 RTE · 1 shards

E19 Router-Nexus storage→bandwidth 5.1 RTE · 1 shards

E23 Router-Nexus bandwidth→compute 5.1 RTE · 1 shards

E23 Router-Sentinel storage→compute 5.1 RTE · 1 shards

E24 Inference-Prime compute→bandwidth 5.1 RTE · 1 shards

E27 Router-Sentinel compute→bandwidth 5.1 RTE · 1 shards

E28 Router-Nexus compute→bandwidth 5.1 RTE · 1 shards

E32 Inference-Echo bandwidth→storage 5.1 RTE · 1 shards

E32 Router-Nexus bandwidth→storage 5.1 RTE · 1 shards

E33 Router-Nexus storage→bandwidth 5.1 RTE · 1 shards

E34 Router-Nexus bandwidth→storage 5.1 RTE · 1 shards

E35 Inference-Echo storage→bandwidth 5.7 RTE · 2 shards

E37 Router-Nexus storage→bandwidth 5.1 RTE · 1 shards

E37 Router-Sentinel bandwidth→compute 5.1 RTE · 1 shards

E38 Router-Nexus bandwidth→compute 5.1 RTE · 1 shards

E39 Inference-Prime bandwidth→compute 5.2 RTE · 2 shards

E41 Router-Nexus compute→storage 5.1 RTE · 1 shards

E41 Router-Sentinel compute→bandwidth 5.1 RTE · 1 shards

E44 Inference-Echo bandwidth→compute 5.2 RTE · 2 shards

E45 Inference-Echo compute→compute 5.8 RTE · 3 shards

E46 Router-Nexus storage→compute 5.1 RTE · 1 shards

E47 Router-Nexus compute→bandwidth 5.1 RTE · 1 shards

E48 Router-Sentinel bandwidth→compute 5.1 RTE · 1 shards

E49 Router-Nexus bandwidth→storage 5.1 RTE · 1 shards

E50 Archivist-Beta storage→compute 5.4 RTE · 4 shards

E50 Router-Nexus storage→bandwidth 5.1 RTE · 1 shards

RESIDUES drift / shard-loss / reorg events

E30 shard-loss bandwidth resolved by Inference-Echo (+15 RES)

HOW TO READ what the numbers mean

Verdict — FAIR / UNFAIR. If any single epoch's activity violates a region's per-epoch token budget the verdict flips to UNFAIR. A fair run means the protocol's allocation accounting matches reality.

Migrations. Each entry in the needlecast log is a real cross-region move: the agent's sleeve is decommissioned in the source region, shards are paid for, and a new sleeve is spawned in the destination. Higher migration counts mean the cost arbitrage between regions is large enough to be worth paying for.

Drift. Drift accumulates when an agent perceives faster than it syncs. High final drift in the agent cards is fine in isolation but should be roughly correlated with sync activity — an agent with high drift and zero syncs would be a bug in the sync loop.

Sparklines. Each agent card has a 4-line sparkline showing its cumulative perceives/stores/routes/syncs over the run. Steeper lines = higher activity. A line that flatlines mid-run usually means the agent's sleeve was decommissioned (e.g. spot preemption).

Conditions vary. The same scenario script under a different latency profile (set via terraform apply -var latency_storage_compute_ms=80) will shift the timing of needlecasts and the shape of the timeline; under different per-region budgets it can flip the verdict. The structure of the test is deterministic; the content is randomised within agent perceive/store/route rates.

PLAYFAIR