§ Performance Validation results

DRIFT, measured.

How DRIFT performs against human review, against itself, and across domains.

80%+

NASA Review Concordance

100%

Mesh Trust Failure Catch

Time to First Catch

Mission Review Validation

DRIFT vs. a NASA review board

During the Landolt photometric calibration mission's SRR/MDR review, DRIFT was deployed against the full document corpus alongside the human review board. Outputs were compared against the board's eventual findings.

Concordance

DRIFT matched the human review board's findings at over 80% concordance and surfaced additional issues the team had not yet flagged.

Coverage

The Review Board Ingestion Module achieved an asymmetric coverage match rate of 81.2% across the full review corpus.

Mesh Validation

Distributed peer reasoning, edge hardware

Three Raspberry Pi nodes running independent DRIFT engines. Tested against trust failure scenarios where one node's data qualification status degrades and peer nodes must catch the failure before downstream consequences.

Catch Rate

100% catch rate on trust failure detection across the three-node mesh. Peer nodes consistently caught failures invisible to the originating node.

Detection Latency

2 second time-to-first-catch across the three-node mesh. Trust failures are surfaced before they propagate to dependent decisions.

Hardware Validation

Edge hardware performance, characterized

DRIFT engine and mesh performance measured on Raspberry Pi 5 edge hardware. Three-node configuration with deterministic measurement methodology, controlled thermal and power state, and per-node scenario profiles exercising distinct engine paths.

Test Configuration

Three Raspberry Pi 5 nodes (4GB RAM each) over gigabit Ethernet. Pi OS Lite 64-bit, Python 3.13. CPU governor locked to performance, swap disabled, active cooling. 600-second runs at 10 Hz publish cadence per node, with three distinct scenario profiles.

Categories Measured

— Engine cycle latency (full landscape construction, p50 / p95 / p99 / max)

— Input qualification time per sensor reading

— Engine memory footprint per node (working set and process RSS)

— Inter-node landscape state packet size (msgpack serialized)

— Mesh arrival jitter at sustained 10 Hz cadence

— Packet delivery, deduplication, and loss rates over 10-minute runs

— Thermal stability and throttle behavior under sustained load

Detailed Measurements

Specific p50, p99, and max values for each category, full methodology, and the preliminary measurement report (v1.1, April 2026) are available under NDA. Formal Stage 1 campaign with extended-duration runs and multi-day replication is scheduled for later in 2026.

Methodology Note

Detailed methodology, source documents, and validation procedures are available under NDA. DRIFT is a sealed engine; results presented here are output-side observable. The engine's internals are not disclosed.

Want the methodology under NDA?

Request a Walkthrough