TSX Toronto Stress Test | accountability.ai

Results: Single AGI Agent · TSX Toronto Desk · March 2026

Throughput

253,807/sec

Decisions per second on real TSX symbols.

AKI Latency

3.94 µs

Average AKI capture latency. Python baseline. See Rust note below.

Reliability

100%

Zero dropped records across all 100 million decisions.

Integrity

BLAKE3

Final Merkle root intact and forward-secret. All records fully verifiable.

⚡ Rust Implementation: 0.62-0.95 µs Latency

The 3.94 µs figure above reflects the Python SDK baseline (agdr-aki via PyPI). The AgDR-Phoenix Rust kernel now achieves 0.62-0.95 microseconds AKI capture latency: matching the Zero-Latency Floor benchmark on the homepage. Horizontal scaling is unlimited; single-process ceiling was reached at ~2.85 billion decisions (memory-bound).

Test Parameters

Agent: Single unified AGI agent (the AGI is One)
Symbols: Real TSX: RY.TO, TD.TO, SHOP.TO, CNQ.TO, BNS.TO
Record: Full PPP triplet + reasoning trace on every decision
Oversight: Binary human delta + FOI escalation on 2% of trades
Capture: Full AKI atomic capture on every inference
Total decisions: 100,000,000
Peak latency (FOI escalation): 11.2 ms

Python-only baseline. These results were produced using the Python SDK (agdr-aki via PyPI). The Rust implementation delivers significantly lower latency. See the Rust note above.

What This Proves

AgDR can handle real Canadian high-velocity trading environments without slowing the desk down. The Atomic Kernel Inference (AKI) primitives held perfectly: no partial records, no timing windows.

Every one of the 100 million records remains fully verifiable. The Merkle root is intact and forward-secret: tamper one record and the chain breaks.

This is the proof that AgDR works at market scale.

Part of the AgDR v0.2 foundational standard · Canonical source on GitHub