Field-Tested Engineering Notes

Engineering Papers from Real Deployments

Writeups from production work rather than lab benchmarks: air-gapped inference, latency targets that have to actually be met, and the less glamorous parts that take most of the effort. Each entry says where the numbers come from and what we got wrong.

Browse Papers
Research & Field Literature

Published Papers

Sovereign & Air-Gapped Inference 4 entries
Air-Gapped Systems • 2026 Published

Sovereign Weights: Quantized 405B-Class Models in Zero-Egress Air-Gapped Environments

Field Notes, Defense & Aerospace

Field notes from standing up Llama 3.1 405B and DeepSeek-R1 on in-house GPU nodes with zero cloud egress of any kind. The parts that cost us the first month were the FP8 weight layouts and the vLLM scheduler defaults, not the models themselves. Also covered: why the parts-retrieval system that predates everyone on the project still caps us below the latency we wanted, and what we would have done differently if we were starting today.

FP8 vLLM SCIF TPO Retrieval 2026
25 pp • Tech Report, rev 2 Request Access →
Air-Gapped Systems • 2026 In Review

Zero-Trust Cross-Boundary Model Gateways: HSM-Anchored Inference Topologies

Critical Infrastructure Practice

A gateway that proxies requests to hosted frontier models while keeping prompts and completions out of vendor logs, with an HSM issuing a per-request audit receipt. The red-team results are the honest part: two bypasses we fixed, one that remains open to an attacker holding an AWS root role. We discuss all three.

HSM Red Team VPC
31 pp • Working draft Request Access →
Sovereign Systems • 2025 Published

Cross-Border Enterprise Compute: Sovereign Data Mesh for Multi-National Financial Institutions

Financial Infrastructure Practice

Most sovereignty work is plumbing before it is models. For nine months we threaded inference across banking cores in North America and Europe without letting a request cross the wrong border. Covers the routing rules, the egress controls added after a compliance review, and the two latency budgets that still do not meet spec.

Data Residency GDPR PIPEDA Latency Budgets
28 pp • Architecture Note Request Access →
Sovereign Systems • Internal Working Note

National Sovereign LLM Clusters: Capacity Planning from a Rollout That Ran Long

Public Sector Practice

A capacity memo from a national-cluster rollout scheduled for six weeks that took four months. Sizing, isolation domains, and the continuity-of-operations work nobody had budgeted for. Some numbers in here are estimates; we mark which ones.

Capacity Planning
Internally circulated, not for citation Request Access →
Latency & Throughput Engineering 4 entries
Latency Engineering • 2026 Published

Sub-50ms TTFT Under Extreme Concurrent Load: Event-Driven Inference Gateways

Power & Utilities Practice

What it took to hold a 50 ms first-token target on smart-grid state estimation over five million meter streams while respecting a zero-data-retention constraint that made caching almost useless. The load-testing methodology and the three bottlenecks we had to remove to get from the soft target to a hard one.

TTFT AMI Streams Zero Retention Load Tests
33 pp • Tech Report Request Access →
Latency Engineering • 2026 Published

FP8 / FP4 Serving Without Accuracy Degradation: A Benchmark Audit

Model Optimization

A direct comparison of FP16 against FP8 and FP4 kernels across vLLM, TensorRT-LLM, and SGLang on reasoning workloads we actually run. We tracked GPU memory pressure, cache eviction, and per-task accuracy deltas rather than headline averages. Includes the degradation we saw on one kernel family that the vendor marketing did not mention.

FP8 FP4 vLLM / TensorRT-LLM / SGLang Benchmarks
41 pp • Benchmark Study Request Access →
High-Throughput • 2026 Published

Hyper-Scale Transaction Gateways: High-Volume Real-Time Requests

BFSI & Telecom

Design notes on an inference layer that sits in front of high-volume UPI and telecom telemetry streams. The hard parts were stream ordering, an idempotent cache that survived leader elections, and keeping data localized per jurisdiction. We share the outage numbers too, including the one we caused.

Stream Ordering Idempotency Localization
36 pp • Tech Report Request Access →
High-Throughput • 2025 Published

Petabyte-Scale Telemetry Fusion at Industrial Rotating-Asset Fleets

Energy & Extraction

Streaming vibrational and thermal waveforms into OSIsoft PI and SCADA historians for around a thousand rotating assets. The predictive target was a multi-day advance warning on mechanical failures; in practice lead time ranged from about two days to zero depending on the fault type. We document the spread rather than the best case.

OSIsoft PI SCADA Predictive Alerts
38 pp • Field Report Request Access →
Hardening & Deterministic Output 4 entries
Hardening • 2026 Published

Reducing Non-Deterministic Hallucinations in Safety-Critical Operations

Safety-Critical Systems

Constrained decoding, grammar sampling, and formal verification each solve a slice of the hallucination problem and none of them solve it entirely. We discuss what each technique actually prevented in depot logistics, grid balancing, and adjudication use, and where we still fall back on human review.

Constrained Decoding Formal Verification
34 pp • Tech Report Request Access →
Hardening • 2026 Published

ERP Line-Item Auditing with Formal Guardrails: Catching Pricing Discrepancies at Scale

Industrial Sourcing

Document reasoning over SAP NetWeaver and Oracle ERP exports to flag pricing discrepancies that human reviewers were missing. The guardrails narrow what the model is allowed to assert; review velocity improved, but the interesting number is the false-positive rate, which we report honestly.

SAP NetWeaver Oracle ERP Audit Guardrails
29 pp • Case Study Request Access →
Hardening • 2025 Published

Zero-Knowledge Clinical Reasoning: PHI-Safe Pipelines in Confidential Enclaves

Regulated Healthcare

Clinical reasoning that never retains or exposes protected health information: data is stripped before model ingestion, run inside a confidential enclave, and cross-checked against current formularies. Operationally it is more about enclave build and key management than about the model. Compliance evidence and design choices are here.

Confidential Enclave HIPAA PHI
26 pp • Tech Report Request Access →
Security Audit • 2026 Published

Crawford Security Findings: BFSIboard.org-Aligned Security Research Report

stage-x.space Research Team

An open security research walkthrough of Crawford & Company's public data surface: project inventory, entity and relationship mapping across the BFSIboard-aligned corpus, and the threat-modeling notes that came out of it. Scope and method are documented so the findings can be independently reproduced.

Security Research Threat Modeling
18 pp • Security Report Read Paper →
Case Study • 2026 Published

Isolating and De-Identifying a Multi-Platform Insurance Codebase: A Field Study of the Renovo Platform

Crawford Apps & Claims Automation Team

A defensive engineering field study documenting the forensic migration of 5 claims repositories (9,590 files / 355 MB) from iCloud Drive, scrubbing 15K PII matches with zero residual leaks and isolating 554 confidential claims workbooks and 410 SQL database dumps.

Data Isolation PII De-Identification Migration Operations
12 pp • Field Study Read Paper →
Research Program • 2026 In Review

Optimization of Fuel-Burn Control Loops in 1,400°C Kiln Combustion

Heavy Industrial Edge

A kiln control loop that adjusts fuel feeder rates and draft speed over OPC-UA on plant-edge compute. Early results show fuel-burn reduced by single-digit percentages and unplanned stops down roughly a third, but clinker quality variance is still under study. This entry is a preview while the paper is finalized.

OPC-UA Kiln Control
Forthcoming Request Access →
Context, before you read the numbers
47 days
Median time from signed engagement to first production inference, across our last six deployments. The range is wide; the slowest took more than twice that.
$100M+ sourcing spend
Client-reported savings from our pricing-discrepancy audit work. Not independently audited, and concentrated in a few large engagements rather than spread evenly.
99.2% telemetry uptime
Aggregate across managed fleets over the trailing twelve months. Individual customers saw 98% to 99.9% depending on maintenance windows we did not control.
<50ms first token
The target we design to for grid and trading workloads, and roughly what we measure at p50. Tail latency sits higher; the TTFT papers show the full distribution.